February 20th, 2024

AI x Crypto Primer
Mohamed Baioumy Alex Cheema
PhD (AI & Robotics) Physics Alum
University of Oxford University of Oxford
 Contents

1. Introduction 4
2. Core Frameworks for AI x Crypto 5
2.1. What Makes Something
. . . . . . . . . An
. . . .AI
. . x. . Crypto
. . . . . . .Project?
......................... 5
2.2. How AI and Crypto
. . . . . .Can
. . . .Help
. . . . .Each-Other
.................................. 7
2.3. Internal vs External
. . . . . .Support
........................................... 8
2.4. Identifying Bottlenecks
................................................. 10
2.5. Measuring Utility
................................................. 11
2.6. Assessing Projects
................................................. 12
2.7. Understanding .Complex
. . . . . . . . Products
........................................ 13
3. The Big Ideas 16
3.1. Crypto and The
. . Deepfake
. . . . . . . . .Apocalypse
...................................... 16
3.2. The Bitter Lesson
................................................. 28
3.3. AI Agents Will. .Disrupt
. . . . . . .Google
. . . . . . .and
. . . Amazon
.............................. 33
3.4. Crypto Will Accelerate
. . . . . . . . Open
. . . . . .Source
. . . . . . AI
............................. 38
4. Opportunities for Builders 44
4.1. A Curated List. .of. .Suggested
. . . . . . . . . Ideas
.................................... 44
A. Appendix: How Does Machine Learning Work? 48
A.1. Artificial Intelligence
. . . . . . .and
. . . Machine
. . . . . . . . learning
............................... 48
A.2. The Machine Learning
. . . . . . . .Pipeline
......................................... 49
A.3. Types of Machine
. . . .Learning
............................................. 56
B. Appendix: Challenges in Machine Learning 61
B.1. Data Challenges
................................................. 61
B.2. Challenges in Training
................................................. 67
B.3. Challenges in Inference
................................................. 70
B.4. Challenges in LLMs
................................................. 71
C. Appendix: Who's Who in AI x Crypto? 74
 Acknowledgements

Thanks to Arbion Halili for his substantial contributions to the writing, typesetting
and editing of this document.
Thanks to the following people for feedback & discussions which helped shape this re-
port: Sandeep Nailwal (Co-founder of Polygon), Harry Grieve (co-founder of Gensyn),
Dima Romanov (co-founder of LayerN), Mohamed Fouda (AllianceDAO), Oliver Turn-
bull (PhD researcher at Oxford University), Qiao Wang (AllianceDAO), Altan Tutar
(NEAR Protocol), DCBuilder (Contributor at WorldCoin), Pim & Bram van Roelen
(Maven11), Will Papper (Syndicate.io), Bharat Krymo (6529 Capital), Justin McAfee
(1kx), Rahilla Zafar (Stability AI), Raphael Doukhan (Giza), Jack Chong (Checker Fi-
nance), Yarco Haydek (Pragma), Piotr Saczuk (AlephZero), David Ma (AllianceDAO),
Illia Polosukhin (co-founder of NEAR), Will Robinson (AllianceDAO), Nick Emmons
(co-founder of Upshot), Siyuan Han (ABCDE) and Qiyun Lu (BeWater).
Special thanks to the BeWater team for translating this document to Chinese as well
as offering us valuable feedback.
Additional thanks to our fellow HomeDAO members: Eito Miyamura, Josh Lavorini,
Konrad Kopp, Joseph Rowell, Akshaiyram Reguram, James O’Leary, Lucas Schneider,
Orlando Fraser, Alex PF and Simon Sällström.
Finally, thanks to everyone to helped review this report including the Foresight team,
Tze Donn Ng, Owen Shen, Neel Daftary, Dylan Zhang, Bart Jaworski, Matt Garcia,
Cathie So, Kartin Wang, Anand Iyer, Shafu, Lucian Zhu (DAO BASE), Morgan
(Hopscotch), Young Ling, Will Nowak, Anna Bertha (DCG), Bobby (UMA Protocol),
Philipp (GnosisDAO), Mikko Ikola (Veridise), Matt Garcia, Louis Guthmann (Stark-
ware), Francis Zhan (Tribe Capital), Tim (flock.io), Gabriel Chua, Nitanshu Lokhande,
0xAA, Cesar Rodriguez (IQ), Ethan (MyShell), Ayush Patel (BuildersTribe), Chase
Ando (Pickem), Leo (Supersight), Sid Rao, David Liu, Fred, Gabriel C and Tommy
Shaughnessy, QZ (Ethereum Singapore), Risabh, Haque Farazul.

3 of 128
 Introduction

Artificial Intelligence (AI) will cause unprecedented levels of societal change.

The rapid growth in this space will inevitably lead to widespread economic disruption,
with AI creating new possibilities in every industry. The Crypto industry is no excep-
tion. We observed three major DeFi attacks in the first week of 2024 and $76B stored
in DeFi protocols at risk. Using AI, we can check smart-contracts for security holes,
and integrate an AI-based security layer into blockchains.
The limitations of AI permit bad actors to abuse powerful models, as shown by the
spread of malicious deepfakes. Thankfully, various advancements in cryptography will
introduce new functionalities to AI models, greatly enriching the AI industry whilst
addressing some of its most critical shortcomings.
The convergence of AI and Crypto will birth numerous high-traction projects. While
some will provide a solution for the problems discussed above, others will combine AI
and Crypto in superficial ways with no real benefit.
In this report, we introduce conceptual frameworks, supported by specific examples
and actionable insights, to aid your understanding of this space’s past, present and fu-
ture. We will uncover the real opportunities and help you steer clear of inconsequential
distractions.

A Note From The Authors:

We were caught off guard by the release of Sora, an AI model that can create
realistic video from text instructions.
Sora was announced whilst this report was being written.
This is confirmation that the risks that we have highlighted in this report are
real and that they are worsening in real-time.

4 of 128
 Core Frameworks for AI x Crypto

This section will introduce some practical mental tools for understanding AI x Crypto,
helping you to analyse individual projects in greater detail.

2.1 What Makes Something An AI x Crypto Project?

Let’s review some examples of projects which use both Crypto and AI, then discuss
whether they are actually AI x Crypto projects.

Case Study: Privasea

Privasea utilizes Fully Homomorphic Encryption (FHE) to train AI models for

processing encrypted data, ensuring input privacy. This innovation facilitates a
ChatGPT-like model that operates on encrypted prompts, meaning the service
provider never sees unencrypted content. The system processes these encrypted
inputs and generates encrypted outputs, which only the user can decrypt and un-
derstand. This method protects user data from exposure and is particularly crucial
for sensitive sectors like finance, where safeguarding against potential information
leaks is paramount. By enabling secure, encrypted interactions with AI, Privasea
addresses significant privacy and security concerns in AI applications.

This case study shows Crypto helping to improve an AI product – a cryptographic

method being used to change how the AI is trained. This results in a product that would
not be possible using AI techniques alone: a model that accepts encrypted prompts.
Adding encryption is an example of using a cryptographic technology to improve the
capabilities of AI.

Figure 1: Using crypto to make internal changes to the AI stack results in

novel capabilities. For example, FHE allows us to use encrypted prompts.

5 of 128
 AI x Crypto Primer

Case Study: Dorsa¹

Dorsa is a blockchain security start-up, using AI models to audit and monitor

smart contracts. They have developed a variety of tools, including reinforcement
learning agents to extensively test the security of a smart contract. The tools de-
veloped by Dorsa can help both developers and auditors; these AI models can
improve both the security and the reliability of smart contracts.

In this case, AI technology is being used to improve a Crypto product – the inverse of
what we discussed earlier. Dorsa offers an AI model that makes the process of crafting
a secure smart-contract faster and cheaper, and although it is off-chain, the use of the
AI model still contributes to the functionality of Crypto projects: smart-contracts are
often at the core of Crypto solutions.
Dorsa’s AI can spot bugs that humans forget to check for, thus preventing future hacks.
Yet, this particular example doesn’t use AI to make Crypto products capable of things
that they couldn’t do before – it was always possible to write secure smart-contracts,
Dorsa’s AI just makes the process better and faster. Nonetheless, this is an example of
AI technology (the model) improving a Crypto thing (the smart-contract).

Case Study: MySentient.ai

MySentient.ai is a Web2 AI girlfriend app. They don’t censor explicit chats or im-
ages and thus might flag issues with Visa and Mastercard (similar to this company
getting censored). To address this issue, they are accepting payment in crypto.
This is the only way in which MySentient utilises crypto technologies.

MySentient is not an example of Crypto x AI. We’ve established that AI can help
improve the crypto stack, and vice versa – this is illustrated by the examples of Pri-
vasea and Dorsa. However, in the example of MySentient, the Crypto components
and AI components are not interacting – they simply co-exist within the product. For
something to count as an AI x Crypto project, it is not enough for AI and Crypto
to contribute to the same product or solution – the functionality provided by these
technologies must interlock to produce the solution.

AI and crypto are often combined in superficial ways that don't have tangible
value.

¹Dorsa is a company in stealth, there is no public infomation about this firm at the time of publi-
cation.

6 of 128
 Core Frameworks for AI x Crypto

Crypto and AI are technologies which can be directly combined to produce better so-
lutions. The use of one in combination with the other can make the other work better
as part of the overall project. Only projects that involve this co-operation between
technologies are classified as AI x Crypto projects.

2.2 How AI and Crypto Can Help Each-Other

Notice:
We use the word “Crypto” to refer to both blockchain technologies and
cryptographic solutions across this report. Zero-knowledge proofs (ZKPs),
cryptographic signatures and blockchain ledgers are all examples of “Crypto”.²
We’ll clarify which aspect of Crypto we are referencing in each section where
appropriate.

Figure 2: Illustration of the 3 ways AI and crypto interact within a product.

Let’s recall the previous case-studies. In Privasea, FHE (which is Crypto) is used to
produce AI models which are capable of accepting encrypted inputs. Therefore, we
are using a Crypto (cryptographic) solution to improve the process of AI training;
thus, Crypto is helping AI. Meanwhile, in Dorsa, an AI model is being used to make
secure-contracts. Here, an AI solution is being used to improve an aspect of the Crypto
product; thus, AI is helping Crypto. This brings us to an important distinction when
evaluating a project at the AI x Crypto intersection: Is crypto being used to help AI,
or is AI being used to help Crypto?
This simple question helps us uncover an important aspect of the use case at hand.
What is the key problem being solved here? In the case of Dorsa, the desired outcome
is to produce a secure smart-contract. This can be developed by a skilled human, but
Dorsa happens to make this process more efficient with AI. The end product of a se-
cure smart-contract remains constant. Once the key problem is clear, we can identify

²This is consistent with how the term is used in other contexts – Vitalik refers to "powerful crypto
in the form of not just blockchain scaling solutions, but also ZKPs, FHE ..."

7 of 128
 AI x Crypto Primer

whether AI is helping crypto or crypto is helping AI. In some cases, there is no mean-
ingful interaction between the two (e.g. MySentient).
The table below provides several examples in each category.

Crypto Helping AI AI Helping Crypto

1. FHE being used to train an 1. AI models auditing and

AI model in a privacy- monitoring smart contracts
preserving manner 2. AI models being used for
2. Federated learning for AI efficient search of on-chain
model training data
3. Zero-knowledge proofs for 3. LLMs for creating data
verifiable inference analytics dashboards
4. Token incentive for collecting 4. Intent-based trading
and labelling data sets platforms
5. Payment rails for AI agents 5. Customizable bots for on-
6. Decentralized GPU networks chain games
6. AI agents managing a DAO

Table 1: How Crypto And AI Can Support Each Other.

You can find our directory of more than 150 AI x Crypto projects in the Appendix. If
there is anything we’ve missed, or if you have any feedback, please reach out!

2.2.1 Mini-Summary
Both AI and Crypto are capable of supporting the other technology in serving its pur-
pose. When you assess any product, it’s key to understand whether it is an AI product
at its core, or a Crypto product at its core.

Figure 3: An illustration of the distinction.

2.3 Internal vs External Support

Let’s consider an example of Crypto helping AI. When the collection of specific tech-
niques which make up an AI technology changes, then what the AI solution as a whole

8 of 128
 Core Frameworks for AI x Crypto

is capable of changes. This collection of techniques is called a Stack. The AI stack

includes the mathematical ideas and algorithms which make up every aspect of the AI
model. The specific techniques used for the processing of the training data, the training
of the model, and model inference, are all part of the Stack.

Stacks, as a whole, have properties and capabilities which are not present in any
of its individual parts. It's the interaction between the parts, like cogs in the
clock, that is responsible for the higher-level behaviour of the Stack.

Within a Stack, its parts are deeply interconnected – it’s the way that the specific
techniques are combined that is responsible for the features of the Stack. Therefore, by
altering the Stack, you alter what the technology as a whole can achieve. Introducing
new techniques into a stack can create new technological possibilities – Ethereum added
new techniques within their Crypto Stack to make smart-contracts possible. Similarly,
changes to the Stack can allow developers to bypass problems which were previously
considered inherent to the technology – the changes made by Polygon to their version
of the Ethereum Crypto Stack enabled them to reduce transaction fees to a level that
was considered impossible.

When the methods (techniques) used in one technology are used within the
Stack of another technology, this represents a deeper and more technical
integration of the technologies, which we will call "internal". If two
technologies are used together, in support of one another, but the pieces which
make up each stack do not change, then the integration will be "external".

• Internal Support: Crypto is used to make changes which are internal to an AI

stack, such as changing the technological means by which models are trained. For
instance, one can introduce FHE techniques into the AI stack, as Privasea does.
Here, a cryptographic component is built directly into the AI stack, resulting in
a modified AI stack. In the case of Privasea, the use of FHE techniques.
• External Support: Crypto is used to support the AI-based features without
making changes to the AI Stack. Bittensor is an example of this, incentivizing
users to contribute data - data which can then be used to train an AI model.
Here, no change to how the model is trained or used under the hood; the AI
stack doesn’t undergo any changes. The use of financial incentives within the
Bittensor network nonetheless helps the AI Stack serve its purpose better.

9 of 128
 AI x Crypto Primer

Figure 4: An illustration of the previous discussion.

Similarly, AI can help the Crypto either internally or externally:

• Internal Support: AI techniques are used within the Crypto Stack. The AI
lives on-chain and interfaces directly with components inside the Crypto Stack.
As an example, take an on-chain AI agent that governs a DAO. This AI doesn’t
just assist the crypto stack. It is an integral, deeply embedded part of the tech-
nological Stack which makes the DAO function.
• External Support: AI is external to the Crypto Stack. The AI is being used to
support the crypto stack without making internal changes to it. Platforms such
as Dorsa use an AI model to secure a smart contract. The AI is off-chain and
is an external tool used to make the process of writing secure smart contracts
faster and cheaper.

Figure 5: Here’s the model, updated to include the distinction between in-
ternal and external support.

The first stage of analysis for any AI x Crypto project is deciding which category it
belongs to.

2.4 Identifying Bottlenecks

Internal support, which features deep technological integration, tends to require more
technical difficulties than examples of external support do. For instance, if we want to
modify the AI stack by introducing FHE or zero-knowledge proofs (ZKPs), we will need
staff who have considerable expertise in both cryptography and AI. Very few people
fall into this intersection. Companies such as Modulus, EZKL, Zama and Privasea,

10 of 128
 Core Frameworks for AI x Crypto

therefore, need substantial financial resources and rare talents to advance their solu-
tions. Allowing users to integrate AI inside smart-contracts requires similarly in-depth
knowledge; companies like Ritual and Ora must tackle complex engineering problems.
External support, conversely, has bottlenecks, but they often involve less technologi-
cal complexity. For instance, adding Crypto payment features to an AI agent doesn’t
require us to overhaul the model. It’s relatively easy to implement. Whilst building
a ChatGPT plugin which allows ChatGPT to fetch statistics from the DeFi LLama
webpage is not technically complex for an AI engineer, very few AI engineers are
members of the Crypto community. Although the task is not technically complex,
few AI engineers are willing to implement these tools, and many are unaware of these
possibilities.

Description The Main Bottleneck To Innovation

Crypto helping AI, Fundamental deep-tech challenges need to be

internally solved, including scaling ZKML and scaling
homomorphic encryption.

Crypto helping AI, Good economic models, aiming, for instance,

externally to incentive data collection or reward users for
contributing their GPUs to a network.

AI helping crypto, Engineering challenges. AI tools are mature

internally relative to the crypto stack, so adding AI
requires solving complex engineering issues.

AI helping crypto, The scarcity of AI talent in the crypto space.

externally

2.5 Measuring Utility

There will be good projects across all four of these categories.
If AI is integrated inside the Crypto Stack, smart-contract developers will be able to
access to AI models on-chain, thus increasing the number of possibilities and potentially
resulting in a wide range of innovations. This is equally true of cases where Crypto is
integrated into the AI Stack – deep technological fusion will produce new possibilities.

11 of 128
 AI x Crypto Primer

Figure 6: Adding AI to the crypto stack gives developers new capabilities.

In the case of AI helping Crypto externally, it is likely that an AI integration will

improve an existing product whilst producing fewer breakthroughs and introducing
fewer possibilities. For example, using an AI model for writing smart contracts might
be faster and cheaper than before, and may also improve security, but it is unlikely
to produce smart-contracts that weren’t possible before. The same applies to Crypto
helping AI externally – token incentives can be used around the AI stack, but that in
itself is unlikely to redefine how we train AI models.
In summary, the integration of techniques from one technology into the tech Stack of
another is likely to produce novel capabilities, whilst the use of a technology outside of
the Stack is likely to improve usability and efficiency.

2.6 Assessing Projects

The amount of that a given project can be estimated partially based on what quadrant
it fits into, since examples of internal support between the technologies can yield greater
rewards, but estimating the total risk-adjusted upside of a project will require us to
consider more factors and risks.
One factor to consider is whether the project under consideration can be useful within
the context of Web2, Web3, or both. AI models with FHE features can be used in place
of AI models that don’t have FHE – introducing FHE is useful across both domains,
in any context where privacy is valuable. However, an integration of AI models within
smart-contracts is only relevant in the Web3 context.
As suggested previously, whether the technological integration between AI and Crypto
within a project is internal or external will also determine upside potential – projects
that involve internal support will tend to produce novel capabilities and more drastic
efficiency gains, which are more valuable.
We must also consider the time horizon over which this technology can be expected
to mature – this will determine how long one would have to wait for a return on their

12 of 128
 Core Frameworks for AI x Crypto

investment into the project. This can be done by analysing current progress and con-
sidering the bottlenecks which are relevant to the project (see Section 2.4).

Figure 7: A hypothetical example illustrating potential upside compared to

time horizon.

2.7 Understanding Complex Products

Some projects will involve a combination of the four categories that we have described,
rather than just one. When this is the case, the risks and benefits associated with the
project tend to multiply, and the time horizon over which the project can be executed
will be longer.
Additionally, you will have to consider whether the whole of the project is better than
the sum of its parts – it’s often not enough to satisfy end users for a project to have a
bit of everything all in one place. A focused approach tends to yield great products.

Leveraging all four combinations of Crypto and AI can make a project more
valuable, but only if the whole is greater than the sum of its parts.

2.7.1 Example One: Flock.io

Flock.io allows you to train a model such that it is “split” between multiple servers, and
no party has access to all of the training data. Since you can participate in training the
model directly, you can use your own data to contribute to the model without revealing
any of it. This is good for user privacy. This involves crypto helping AI internally, as
the AI stack (model training) is changed.
Additionally, they reward the people who participate in the training of models using
Crypto tokens and also use smart-contracts to financially penalise people who sabotage
the training process. This does not alter the processes involved in training the model,
the underlying techniques stay the same, but there would be no way to keep the parties

13 of 128
 AI x Crypto Primer

involved in the training of models honest without the on-chain slashing mechanism.
This is an example of Crypto helping AI externally.
Crucially, crypto helping AI internally introduces a new capability: a model can be
trained across a decentralized network while keeping data private. However, crypto
helping AI externally doesn’t introduce a new capability, as the token is merely used
to incentivise users to contribute to the network. The users can be compensated in fiat
already. Incentivizing users with cryptocurrency is superior and increases the efficiency
of the system, but it does not introduce new capabilities.

Figure 8: An illustration of Flock.io and how the stack changes, where color
changes imply a change has been made internally.

2.7.2 Example Two: Rockefeller Bot

Rockefeller Bot is a trading bot which operates on-chain. It uses AI to decide which
trades to make, but since the AI model isn’t run on the smart-contract itself, we rely
on the service provider to run the model for us, then tell the smart-contract what the
AI has decided, and prove to the smart-contract that they are not lying. If the smart-
contract didn’t check that the service provider wasn’t lying, the service provider could
make harmful trades on our behalf. Rockefeller Bot allows us to prove that the service
provider is not lying to the smart-contract using ZK proofs. Here, ZK has been used to
change the AI Stack. The AI Stack needs to adopt ZK techniques, otherwise, we could
not use ZK to prove what the model decided for the smart-contract.
The resulting AI model, which has verifiable outputs because of ZK techniques, can
now be queried from the blockchain on-chain, meaning that this AI model is used inside
the Crypto stack. In this case, we have used AI models within the smart-contract to
decide trades and prices in a fair manner. This was not possible without the AI.

14 of 128
 Core Frameworks for AI x Crypto

Figure 9: An illustration of Rockefeller Bot and how the Stack changes. Color
changes imply a change in the Stack (internal support).

15 of 128
 The Big Ideas

3.1 Crypto and The Deepfake Apocalypse

“One computer can create all this havoc … How are we going to operate without
understanding some basic things about truth?”
– Eric Schmidt, Former CEO of Google to Sergey Nazarov, co-founder of Chainlink

On the 23rd of January, an AI-generated voice message falsely claiming to be President

Biden discouraged Democrats from voting in the 2024 primary. Barely a week later,
a finance worker lost $25 million to scammers through a deepfake video call mimick-
ing his colleagues. On X (formerly known as Twitter), meanwhile, AI-created explicit
images falsely attributed to Taylor Swift attracted 45 million views and sparked wide-
spread outrage. These incidents, all taking place in the first two months of 2024, are
only a snapshot of the diverse and damaging impact deepfakes can have across politics,
finance, and social media.

3.1.1 How Did They Become a Problem?

Fake images are nothing new. In 1917, pictures of intricate cardboard cutouts designed
to look like fairies were published in The Strand magazine; many believed that these
photos were sound evidence for the existance of supernatural forces.

Figure 10: One of the photographs of The Cottingley Fairies. Sir Arthur Co-
nan Doyle, the creator of Sherlock Holmes, used these forgeries as evidence
for psychic phenomena.

16 of 128
 The Big Ideas

Creating fakes has become easier and cheaper over time, drastically improving the
speed at which misinformation spreads. For example, during the 2004 U.S. presidential
election, a manipulated photo falsely showed Democratic nominee John Kerry attend-
ing a protest with Jane Fonda, a controversial American activist. While the Cottin-
gley Fairies had required an elaborate set-up involving cardboard cut-outs of traced
drawings from a children’s book, this forgery was a simple task accomplished using
Photoshop.

Figure 11: This photo purported to show John Kerry sharing a stage with
Jane Fonda at an anti-Vietnam War rally. It later emerged that it was a
forgery, created by combining two pre-existing images in Photoshop.

However, the risk posed by fake photos has been reduced by our learning of how to
identify signs of editing. In the case of “tourist guy”, amateurs were able to identify
that an image was photoshopped based on the observation that the white balance was
inconsistent between objects in the scene. This is a product of the increased public
awareness regarding disinformation; people have learned to notice signs of image edit-
ing. The term “photoshopped” has entered common parlance: signs of image tampering
are universally recognised, and photographic proof is no longer regarded as inalienable
evidence.

3.1.1.1 Deepfakes Make Fakes Easier, Cheaper and More Realistic

Forgeries used to be easily detectable by eye, but deepfakes make it easy and cheap
to create images almost indistinguishable from real photos. For example, the website
“OnlyFake” uses deepfake technology to generate realistic photos of fake IDs in minutes
for just $15. The photos have been used to bypass the anti-fraud safeguards, known as
Know-Your-Customer (KYC), on OKX (a crypto exchange). In the case of OKX, the
deepfake IDs fooled their staff, who are trained to spot doctored images and deepfakes.
This highlights how it is no longer possible to detect deepfake-based fraud by eye, even
for professionals.
We’ve increased our dependence on video evidence in response to the risk of image
deepfakes, but deepfakes will severely undermine video evidence very soon. A researcher
at UT Dallas managed to bypass the identity verification features implemented by
KYC providers using a free deepfake face-swap tool. This is a drastic development - it
used to be expensive and time-consuming to generate videos with the necessary level

17 of 128
 AI x Crypto Primer

of realism. In 2019, it would cost someone two weeks and $552 to create a 38 second
deepfake video of Mark Zuckerberg, featuring noticeable visual defects. Today, we can
create realistic deepfake videos for free, in minutes.

Figure 12: The OnlyFake Panel for creating fake IDs in minutes.

3.1.1.2 Why Video Matters So Much

Videos used to be reliable evidence before the advent of deepfakes. Unlike easily faked
images, videos have historically been challenging to counterfeit, leading to their unques-
tioned acceptance as reliable evidence in our courtrooms. This makes video deepfakes
particularly dangerous.
At the same time, scepticism of real videos can also lead to dismissing genuine con-
tent, as shown when a video of President Biden was wrongly called a deepfake. Critics
cited his unblinking eyes and lighting differences as proof, though these claims were
debunked. This leads to a catch-22 – deepfakes not only make the fake seem real, but
the real seem fake, further blurring the line between truth and fiction and making
accountability more difficult.

We still trust videos, even though they can now be faked.

Deepfakes enable targeted advertising at scale. We may soon witness a version of

YouTube where what is said, who says it, and where it is said are all personalized to
the viewer. An early example of this is a hyper-localised Zomato ad that showed actor
Hrithik Roshan ordering dishes from popular restaurants in the viewer’s local city. A
different deepfake ad, featuring restaurants local to the viewer, was generated based on
the viewer’s GPS location.

3.1.2 What’s Wrong With The Current Solutions?

18 of 128
 The Big Ideas

3.1.2.1 Awareness
Deepfakes are now advanced enough to the fool experts we trained to spot them. This
has enabled hackers to bypass identity verification (KYC/AML) procedures, even those
that feature human moderators. This suggests that we can’t tell deepfakes apart from
real images using our eyes. We cannot culturally prepare for deepfakes merely by being
image-skeptical: we need additional tools to tackle the deepfake epidemic.

3.1.2.2 Platforms
Social media platforms are not willing to implement effective deepfake mitigations
without intense social pressure. For example, Meta agreed to ban deepfake videos in-
volving fake audio, but refused to ban deepfake videos falsely depicting people’s actions.
Against the advice of their own oversight board, they did not remove a deepfake video
showing President Biden touching his granddaughter, a forgery falling into the latter
category.

3.1.2.3 Policy
We need laws that effectively tackle the emerging risks of deepfakes without restricting
the less problematic uses, such as in art or education, which do not try to deceive
people. The spread of non-consensual deepfake images, like those of Taylor Swift, has
motivated legislators to pass tighter laws against such deepfakes. It may be necessary
to legally enforce enhanced moderation procedures online in response to such cases,
but proposals to ban all AI-generated content have raised alarms among filmmakers
and digital artists, who worry it could unjustly restrict their work. Finding the right
balance is key, otherwise, those legitimate creative applications will be stifled.
By pushing legislators to increase the barrier to entry for the training of powerful mod-
els, big tech companies can secure their AI monopoly. This may lead to the irreversible
concentration of power in the hands of a few corporations - for example, Executive
Order 14110 covering AI contains a proposal for stringent requirements on companies
with a large amount of compute.

Figure 13: US Vice President Kamala Harris applauds as US President Joe

Biden signs the US’ first AI executive order.

19 of 128
 AI x Crypto Primer

3.1.2.4 Tech
Building guardrails directly into AI models to prevent misuse is our first line of defence,
but these guardrails have been continuously broken. AI models are hard to censor be-
cause we don’t know how to reliably modify their high-level behaviours using our low-
level tools. Despite this, the firms that train AI models can use the implementation of
guardrails as an excuse to introduce undesirable censorship and biases to their models.
This is problematic because big tech AI firms are not accountable to the public will –
firms are free to influence their models to the detriment of their users.
Even if the creation of powerful AIs wasn’t centralised in the hands of dishonest firms,
it might still be impossible to build an AI which is both guardrailed and unbiased.
Researchers struggle to pin down what constitutes misuse, which makes it difficult to
prevent it whilst handling the user’s requests in a neutral, balanced manner. If we
can’t define misuse, it might seem necessary to make the guardrails less stringent,
thus reopening the possibility of abuse. Entirely prohibiting the abuse of AI models is,
therefore, impossible.
One solution is to detect malicious deepfakes once they’re in the wild instead of prevent-
ing their creation. But, deepfake-detecting AI models (such as those deployed by Ope-
nAI) are becoming obsolete due to inaccuracies. Although deepfake detection methods
have become more sophisticated, the techniques for creating deepfakes are becoming
more sophisticated at a faster rate – the deepfake detectors are losing the technological
arms race. This makes it difficult to identify deepfakes based on the media alone. AI
is advanced enough to create fake footage so realistic AI itself cannot determine its
inaccuracy.
Watermarking technologies allow us to invisibly mark deepfakes such that we can
identify them wherever they appear. But home-made deepfakes won’t always feature
watermarks, since watermarks must be added deliberately. They are, however, still
effective for identifying deepfakes that were generated by parties who are willing to
mark their images as deepfakes (such as OpenAI). Regardless, watermarks can often be
removed or forged with easy-to-use tools, thereby bypassing any watermark-based anti-
deepfake solutions. They can also be removed accidentally: most social media platforms
automatically remove them.
The most popular implementation of deepfake watermarking technology is C2PA (by
the Coalition for Content Provenance and Authenticity). It is designed to prevent mis-
information by tracking where media comes from and storing this information in the
media metadata. It is supported by companies such as Microsoft, Google, and Adobe,
so it is highly probable that C2PA will be rolled out across the entire content supply
chain, unlike its less popular counterparts.

20 of 128
 The Big Ideas

Unfortunately, C2PA has its own weaknesses. Since C2PA stores a full history of edits
to the image, authenticating each edit using cryptographic keys that are controlled by
C2PA-compliant editing software, we have to be able to trust this editing software.
However, it is likely that people will accept the edited images due to the presence of
valid C2PA metadata without considering whether they trust every party involved in
the chain of edits. Therefore, if any editing software is compromised or enables mali-
cious edits, it may be possible to convince others that fake or maliciously edited images
are real.

We need secure hardware, otherwise the keys which secure C2PA will be stolen.

Figure 14: An example image with C2PA-compliant metadata containing a

chain of edits. Each edit is signed by a different trusted party, but only the
final edited image is public. Source: Real Photo vs AI-Generated Art: A New
Standard (C2PA) Uses PKI to Show an Image’s History

Further, the cryptographic signatures and metadata included within the C2PA water-
mark can be linked to specific users or devices. In some cases, the C2PA metadata
could link all images that your camera has ever taken to each other: if we know that an
image came from someone’s camera, we can identify all other images which came from
that camera. This could be used to deanonymize whistleblowers who have published
images from their camera under their real names.
All potential solutions will face a unique set of challenges. However, these challenges
vary drastically - including the limits of social awareness, the shortcomings of big tech
firms, the difficulty of implementing regulatory policy, and our technological limitations.

There is no silver bullet solution for tackling deepfake misinformation.

21 of 128
 AI x Crypto Primer

3.1.3 Can Crypto Solve This Problem?

Open-source deepfake models have already spread. Therefore, it is possible to argue
that there will always be ways to abuse another person’s likeness using deepfakes; even
if the practice is criminalised, someone may choose to generate or platform unethical
deepfake content anyway. However, we can almost entirely eliminate the problem by
moving malicious deepfakes out of the mainstream. We can prevent people from think-
ing deepfakes are real images and enable people to create platforms where deepfakes
are restricted. This section will describe a variety of crypto-based solutions to misin-
formation spread via malicious deepfakes, while highlighting the limitations of each
approach.

3.1.3.1 Hardware Attestation

Hardware-attested cameras embed a unique proof with each photo they take, certifying
that it was taken by that specific camera. This proof is created by a non-cloneable,
tamper-proof chip unique to the camera, ensuring the image's authenticity. A similar
process can be used for audio and video.

It's cheaper to hack software, but expensive to compromise secure chips.

The attestation proof tells us that the image is taken by a real camera, meaning we
can usually trust that it is a picture of a real object. We can flag images which don’t
have this proof. This fails if the camera takes an image of a fake scene designed to look
like a real situation – you can just point the camera at a picture of a deepfake. We can
currently tell if a picture was taken from a digital screen by checking for distortions
in the captured images, but scammers will find ways of hiding these imperfections (for
instance, by using better screens, or by limiting lens flare). Eventually, even AI tools
will fail to identify this kind of fraud, since scammers can find ways to avoid all of these
distortions.
Hardware attestation will minimise the number of deepfakes which are assumed to be
real, but we still need extra tools to prevent the spread of deepfakes in those rare cases
where a camera is compromised or otherwise misused. As we discussed earlier, it is
still possible to create the impression that a deepfake is a real image using hardware-
attested cameras, such as if the camera is hacked or is used to take pictures of deepfakes
on computer screens. Additional tools will be needed to solve this problem, such as a
camera blacklist.
Camera blacklists would enable social media platforms and apps to flag images which
came from that particular camera which is known to have produced misleading images
in the past. They can be implemented without requiring the camera to publicly reveal
traceable data, such as the camera ID, with every published image. However, it’s not

22 of 128
 The Big Ideas

yet clear who will maintain the camera blacklist, or how we prevent them from taking
bribes to blacklist cameras which belong to whistleblowers.³

3.1.3.2 Blockchain-based Image Chronology

Blockchains, being immutable, allow us to add images along with additional metadata
to a timestamped chronology as they appear on the internet such that the timestamp
and metadata cannot be tampered with. Since original, unedited images can be im-
mutably stored on the blockchain by honest parties long before malicious edits prolif-
erate, access to such a record would allow us to identify malicious edits and verify the
original source. This technology was implemented on the Polygon blockchain network
as part of Verify, a fact-checking tool built in collaboration with Fox News.

Blockchain can prevent people from tampering with the chronology of images,
but it can't prevent deceptive images from appearing in the first place.

Figure 15: The user interface for Fox’s blockchain-based tool, Verify. Arti-
cles can be looked up by their URL. The source, transaction hash, signature,
timestamp and additional metadata are fetched from the Polygon blockchain
and displayed.

3.1.3.3 Digital Identity

Trusted sources may become the only way of avoiding disinformation if deepfakes man-
age to destroy our trust in unverified images and videos. We already rely on trusted
media sources to verify information because they employ journalistic standards, fact-
checking processes, and editorial oversight to ensure the accuracy and credibility of the

³Reach out to Arbion if you are interested in the problems detailed in this paragraph.

23 of 128
 AI x Crypto Primer

content they publish. However, we will need a way of verifying whether what we see
online comes from those sources that we trust. This is what cryptographic signature
data is useful for: it can mathematically prove that one authored of a piece of content.
Signatures are generated using digital keys, which are only known by the person who
owns the associated crypto wallet since wallets create and generate keys. This allows
us to know who authored the data – you just check whether the signature corresponds
to a key in one’s crypto wallet, which belongs exclusively to them.
We can leverage crypto wallets to attach signatures to our posts, in a seamless and
user-friendly way. If one uses a crypto wallet to log into a social media platform, they
can take advantage of their wallet’s ability to create and verify signatures on social
media. Therefore, the platform will be able to warn us if a post is not from a trusted
source – it will use automatic signature-checking to flag misinformation.
Additionally, zk-KYC infrastructure, which interfaces with wallets, can tie unknown
wallets to identities verified by a KYC process without compromising user privacy and
anonymity. However, as deepfakes become more sophisticated, the KYC process could
be bypassed, allowing malicious actors to create fake, anonymous identities. This can
be addressed using solutions such as Worldcoin’s Proof of Personhood (PoP).
Proof of Personhood is the mechanism which WorldCoin uses to verify that their wallet
belongs to a real human, and only permits one wallet per human. To achieve this, it
uses the Orb, a biometric (iris) imaging device, to authenticate wallets. Since biometric
data can not yet be faked, requiring social media accounts to be linked to a unique
WorldCoin wallet is a viable way of preventing bad actors from producing multiple
anonymous identities to conceal their unethical online conduct – it solves the deepfake
KYC bypass problem, at least until hackers a way to fool biometric devices.

3.1.3.4 Economic Incentives

“(With Veracity Bonds), if you get caught with your hand in the cookie jar, you
actually lose money, and, as a reader, I get to know how serious you are.”
– Charles Hoskinson: Founder of Cardano; Co-Founder of Ethereum

Authors can be penalized for misinformation; users can be rewarded for identifying
misinformation. For example, Veracity Bonds enables media outlets to stake money on
the accuracy of their publications, facing financial penalties for misinformation. This
offers these media companies a financial reason to be truthful.
Veracity Bonds will be an integral part of our "Truth Marketplace," where different
systems compete to gain the trust of users by verifying the authenticity of content
in the most efficient and robust manner. This is similar to a proof market, such as
Succinct Network and =nil; Proof Market, but for the fuzzier problem of verifying truth

24 of 128
 The Big Ideas

where cryptography alone isn’t sufficient. Smart Contracts can be used as a means
of enforcing the economic incentives required to make these truth markets work, thus
blockchain technologies may play a central role in helping to fight disinformation.

3.1.3.5 Reputation Scores

“If we don’t solve this reputation and authentication of content (problem), shit will
get really weird.”
– Illia Polosukhin: Co-Founder of NEAR; Co-Author of ‘Attention Is All You Need’

We might use reputation as a proxy for credibility. For example, we might look at how
many Twitter followers someone has to determine whether we should trust what they
say. However, the reputation system could take into account each author’s track record,
not just their popularity. We don’t want to confuse trustworthiness with their level of
popularity.
We can’t allow people to generate an unlimited number of anonymous identities, oth-
erwise, they can ditch their identity whenever they tarnish their reputation to reset
their level of social credibility. This will require us to use Digital Identities which can
not be duplicated, as discussed in the last section.
We can also use evidence from Truth Markets and Hardware Attestation to determine
ones reputation since these are robust ways of tracking their track record for factuality.
Reputation Systems are the culmination of all the other solutions discussed thus far
and are therefore the most robust and holistic family of approaches.

Figure 16: Elon in 2018 hinting at a website with credibility scores for jour-
nalists, editors and publications

3.1.4 Do Crypto Solutions Scale?

The aforementioned blockchain solution requires fast and high-storage blockchains –
otherwise, we won’t be able to fit all of our images into an on-chain verifiable chrono-
logical record. This will only get more important as the volume of online data published
each day continues to grow exponentially. However, there are algorithms to compress
data in a way where it can still be verified.

25 of 128
 AI x Crypto Primer

Additionally, the signatures produced through Hardware Attestation do not apply to

edited versions of an image: zk-SNARKs must be used to produce a proof-of-edit. ZK
Microphone is an implementation of proof-of-edit for audio.4

3.1.5 Deepfakes Aren’t Inherently Bad

It's crucial to acknowledge that not all deepfakes are harmful. There are innocent uses of
this technology, such as this AI-generated video of Taylor Swift teaching mathematics.
More personalized experiences are also possible because of the low cost and accessibility
of deepfakes. For example, HeyGen allows users to send personal messages with an AI-
generated face that resembles their own. Deepfakes are also helping to bridge language
gaps with dubbed translations.

Deepfake technology is neutral. It can be used for good and bad.

3.1.5.1 Ways Of Controlling and Monetising Good Deepfakes

AI counterpart services based on deepfake technology are controlled by small teams
that take high fees and lack accountability and oversight. Amouranth, a top earner on
OnlyFans, recently released her own AI Influencer counterpart that fans can talk to in
private. These services, often startups, can restrict or even shut down access, such as
in the case of the AI companion service called Soulmate.
By hosting AI models on-chain, we can use smart-contracts to fund and control models
in a transparent way. This will assure users that they will never lose access to the
model, and can help model creators to distribute profits amongst their contributors
and investors. However, there are technical challenges. The most popular technology
for implementing on-chain models, zkML (used by Giza, Modulus Labs and EZKL),
makes the model run 1000x slower. Nonetheless, the research in this subfield is still
ongoing, and the techniques are improving. For example, Ora is experimenting with
opML and Aizel is building a solution based on Multi-Party-Computation (MPC) and
Trusted Execution Environments (TEE).

3.1.6 Chapter Summary

• The emergence of sophisticated deepfakes is eroding trust across politics, finance,
and social media, highlighting the need for a "Verifiable Web" to maintain truth
and democratic integrity.
• Deepfakes, once an expensive and skill-intensive endeavor, have become easily
producible with advancements in AI, changing the landscape of misinformation.

4
Reach out to Arbion if you are interested in this problem.

26 of 128
 The Big Ideas

• Historical context shows us that manipulation of media is not a new challenge,

but AI has made creating convincing fakes far easier and cheaper, necessitating
new solutions.
• Video deepfakes present a unique danger as they compromise what has tradi-
tionally been considered reliable evidence, leading to a societal catch-22 where
real actions can be dismissed as fake.
• Existing countermeasures are categorized into awareness, platforms, policy, and
tech approaches, each facing challenges in effectively combating deepfakes.
• Hardware attestation and blockchain offer promising solutions by proving the
origins of each image and creating a transparent, immutable record of edits.
• Crypto wallets and zk-KYC enhance the verification and authentication of online
content, while on-chain reputation systems and economic incentives like Veracity
Bonds propose a marketplace for truth.
• Acknowledging the positive uses of deepfakes, crypto also presents a method for
whitelisting beneficial deepfakes, balancing innovation with integrity.

27 of 128
 AI x Crypto Primer

3.2 The Bitter Lesson

“The biggest lesson that can be read from 70 years of AI research is that general
methods that leverage computation are ultimately the most effective, and by a
large margin.”
– The Bitter Lesson, Professor Rich Sutton

This is a statement which is counterintuitive, but true. The AI community refuses to

accept that bespoke methods don’t work as well, yet the Bitter Lesson is still true:
using the most computing power always produces the best results.
We must scale: more GPU’s, more datacenters, more training data.
An example of researchers getting this wrong is when computer-chess researchers tried
building chess engines using insights from the best human chess players. The first chess
programs often copied human opening strategies (using ‘opening books’). Researchers
expected that the chess engine would start with a strong position, without having to
calculate the best moves from scratch. They also contained many ‘tactical heuristics’ –
tactics used by human chess players, such as forks. To put it simply: the chess program
was built according to the human insights regarding how to play chess successfully, not
general computational methods.

Figure 18: A Fork – the Figure 19: An example of a

queen attacks two pieces. chess opening sequence.

In 1997, IBM DeepBlue, which combined enormous amounts of compute power with
search-based technology, beat the reigning chess world champion. Even though it was
superior to all ‘human-engineered’ chess engines, chess researchers shunned DeepBlue.
They argued that DeepBlue’s success was short-lived since it did not employ chess
strategies – they saw it as a crude, brute-force solution. They were wrong: applying
large swathes of compute to general problem-solving methods tends to produce bet-
ter results than the bespoke approaches in the long-run. This high-compute ideology

28 of 128
 The Big Ideas

spawned successful Go engines (AlphaGo), improved speech recognition, and more re-
liable computer vision.
The most recent triumph of the high-compute approach to AI is OpenAI’s ChatGPT.
Unlike previous attempts, OpenAI didn’t try to encode human understanding of how
languages work into the software. Instead, their model combined large swathes of data
from the internet with an enormous amount of compute. They did not intervene nor
embed any biases into the software, unlike other researchers. In the long-run, the best-
performing methods are always based on general methods which take advantage of large
amounts of compute. This is a historical observation; indeed, we probably have enough
evidence to say it will usually be true.
The reason that combining enormous amounts of compute power with large swathes
of data is the best approach in the long-run is Moore’s law: the cost of compute will
exponentially decrease over time. In the short-run, we may fail to ascertain a sizable
increase in compute bandwidth, which can result in researchers trying to improve their
technology by manually embedding human knowledge and algorithms into the software.
This can work for some time, but it will not produce success in the long-run: embedding
human knowledge into the underlying software makes the software more complicated
and the models less capable of improving in response to extra compute power. This
makes the manual approach short-sighted, and so Sutton recommends that we ignore
manual techniques and focus on applying more compute power to general computa-
tional techniques.
The Bitter Lesson has serious implications for how we ought to build decentralized AI:
Building Large Networks: The above lesson highlights the urgency of developing
larger AI models and assembling extensive computational resources to train them.
These are key steps to advance into the new AI frontier. Companies like Akash, GPUNet
and IoNet aim to provide that scalable infrastructure.

29 of 128
 AI x Crypto Primer

Figure 20: Akash price compared to other providers like Amazon AWS.

Innovate on Hardware: ZKML approaches have been criticized because they run
1000x slower than their non-ZK counterparts. This parallels the criticism faced by
neural networks. In the 1990s, neural networks showed great promise. Yann LeCun’s
CNN model, a small neural network capable of classifying images of hand-written digits
(see images below), found success. By 1998, banks were using this technology to read
more than 10% of all checks in the US. However, these CNN models didn’t scale, so in-
terest in these neural networks decreased dramatically, and computer vision researchers
started using human knowledge again to create better systems. In 2012, researchers
developed a new CNN whilst leveraging the compute efficiency of GPUs, a popular
piece of hardware usually used to generate computer graphics (gaming, CGI, etc). This
allowed them to reach incredible performance, exceeding all the other methods avail-
able at the time. This network was called AlexNet, and it sparked the deep learning
revolution.

Figure 23: AlexNet (2012)

was capable of handling com-
Figure 22: Neural networks plex images and outper-
in the 90s were limited to formed all alternative meth-
handle low-resolution images ods.
of digits.

30 of 128
 The Big Ideas

The upscaling of AI technology is certain because compute is always getting cheaper.

Custom hardware for technologies such as ZK and FHE will accelerate progress – com-
panies such as Ingonyama, as well as academia, are paving the way. In the long-run,
we will achieve large-scale ZKML by applying more compute power and introducing
efficiency improvements. The only question is, what will we use these technologies for?

Figure 24: An example of hardware advancements for ZK provers. (Source)

Scale Data: As AI models grow in size and complexity, it is necessary to scale data
sets accordingly. Generally, the size of the dataset should grow exponentially with the
model size to prevent overfitting and ensure robust performance. For a model with
billions of parameters, this often means curating datasets comprising billions of tokens
or examples. Google’s BERT model, for example, was trained on the entire English
Wikipedia, containing over 2.5 billion words, and the BooksCorpus, with approximately
800 million words. Meta’s LLama, meanwhile, was trained on 1.4 trillion tokens. These
figures underscore the magnitude of the datasets we need – as models advance towards
trillions of parameters, datasets must expand further. This scaling ensures that models
can capture the nuance and diversity of human language, making the development of
vast, high-quality datasets as critical as the architectural innovations in the models
themselves. Companies like Giza, Bittensor, Bagel, and FractionAI are addressing spe-
cific needs in this space (see Chapter 5 for details on challenges in the data space such
as model collapse, adversarial attacks and challenges in quality assurance).
Develop General Methods: In decentralized AI, the draw towards application-spe-
cific methods for technologies like ZKPs and FHE is driven by the quest for immediate
efficiency. Tailoring solutions to particular architectures enhances performance but may
sacrifice long-term flexibility and scalability, restricting broader system evolution. Con-
versely, focusing on general-purpose methods offers a foundation that, despite initial

31 of 128
 AI x Crypto Primer

inefficiencies, is scalable and adaptable to diverse applications and future developments.

These methods are set to excel as computational power grows and costs decrease,
thanks to trends like Moore’s Law. The choice between short-term efficiency and long-
term adaptability is crucial. Emphasizing general methods prepares decentralized AI
for a future of robust, flexible systems that leverage computational advances, ensuring
enduring success and relevance.

3.2.1 Conclusions
In the early stages of a product’s development, it might be crucial to pick approaches
that don’t benefit from scale. This is relevant for both companies and researchers val-
idating use cases and ideas. However, the bitter lesson suggests that we should always
remember to prioritize general scalable methods in the long-run.
Here’s an example of a manual approach being replaced with automatic, general so-
lution: before automatic differentiation (autodiff) libraries, such as TensorFlow and
PyTorch, gradients were often computed manually or using numerical differentiation —
this was inefficient, prone to error, and caused problems which wasted the researchers
time, unlike autodiff. Autodiff is now indispensable, because autodiff libraries have
enabled faster experimentation and simplified the development of models. Thus, the
general solution won – but the old, manual methods were necessary to enable ML re-
search before autodiff became a mature and usable solution.
In summary, Rich Sutton’s Bitter Lesson teaches us that our AI’s will improve more
rapidly if we maximise the computing power we put into them, rather than trying to
make them mimic well-understood human approaches. We must scale existing compute,
scale data, innovate on hardware and develop general methods – the adoption of this
approach will have numerous implications for the decentralized AI space. Although the
Bitter Lesson does not apply in the earliest stages of research, it will probably usually
be true in the long-run. Thi

32 of 128
 The Big Ideas

3.3 AI Agents Will Disrupt Google and Amazon

3.3.1 The Google Monopoly
Online content creators generally depend on Google to distribute their content. In re-
turn for allowing Google to index and display their work, they are provided with a
consistent flow of attention and revenue from advertisements. However, this relation-
ship is imbalanced; Google holds a monopoly (over 80% of search engine traffic), a
market share that content creators on their own could never hope to match. As a result,
content creators are heavily reliant on Google, as well as other tech giants, for their
revenue. A single decision by Google has the potential to end an individual’s business.
Google’s introduction of featured snippets - which display answers to user queries with-
out requiring clicking through to the original site - highlights this problem, as infor-
mation can now be obtained without ever leaving the search engine. This disrupts the
arrangement content creators had come to rely on. In exchange for Google indexing
their content, creators expected referral traffic and eyeballs on their sites. Instead, fea-
tured snippets allow Google to summarise content while cutting creators out of the
traffic flow. The fragmented nature of content producers leaves them largely powerless
to take collective action opposing Google’s decisions; with no unified voice, individual
sites lack any bargaining power.

Figure 25: Example of a simple featured snippet.

Google experimented further by providing a list of sources for the answers to the user’s
query. The example below contains sources from the NYT, Wikipedia, MLB.com and
more. These sites would not be getting as much traffic, since Google provides the an-
swer directly.

33 of 128
 AI x Crypto Primer

Figure 26: Example of a featured snippet with 'From the Web' functionality.

3.3.2 The OpenAI Monopoly

Google’s introduction of featured snippets represented the beginning of a concerning
trend - reducing the attribution of original content creators. ChatGPT has taken this
concept to its logical extension, functioning as an omniscient information agent without
any links or references back to the source material.
Language models like ChatGPT can answer virtually any question by summarising
content scraped from the internet, but they fail to direct users to visit the original
publishers. Instead, the model accumulates knowledge derived from copyrighted works
into a singular interface entirely controlled by OpenAI.
The success of these models relies on the vast amounts of data that make up the in-
ternet, and content creators are not rewarded at all for their essential input to model
training. Some larger publishers have managed to reach agreements with the likes of
OpenAI, but such a solution is impossible for smaller content creators. Others have
decided to simply block AI models from trawling their content, but with close sourced
models there is no guarantee of this being respected.
AI companies try to justify their lack of compensation by arguing that the AI system
is simply learning from the content - like a human might learn about the world from
reading - but does not infringe upon content creators when producing an output. How-
ever, this assertion is questionable, as ChatGPT can regurgitate entire New York Times
articles verbatim. Both Midjourney and DALL-E are similarly capable of generating
copyrighted content.
The implications are clear - big tech continues to consolidate power, while unaffiliated
content creators see their leverage diminish. The asymmetric relationship that launched
lawsuits against Google has only grown more extreme. Established publishers like the

34 of 128
 The Big Ideas

New York Times have already taken legal action, as have a range of content creators
from digital artists to coders, in the form of class action lawsuits.
One proposed solution is Retrieval Augmented Generation (RAG), which allows the
language model to provide sources for its answers. However, this shares issues with
featured snippets - it doesn’t give the user any reason to visit the original site at all.
Worse, tech giants could leverage RAG disingenuously as legal cover while still denying
creators traffic and ad revenues.

Figure 27: RAG includes links as sources. This result was generated by Per-
plexity AI.

3.3.3 Crypto’s Potential Solution

As “information agents” like ChatGPT emerge, big tech seems poised to repeat the past
and monopolise generative AI, despite their reliance on unaccredited creator content.
But with AI now disrupting the market, there is an opportunity to redistribute power
and build a fairer system for compensating creators. In the early 2000s, policymakers
missed the chance to establish an equitable model, resulting in today’s distribution
system - a system monopolised by Google. The current rise of AI presents a crossroads
— will we address past mistakes, or will history repeat itself with OpenAI and other
tech giants gaining unilateral control?
To incentivise quality content production over the long term, we must explore ways
to continue providing fair payment to creators. As Chris Dixon argues, crypto offers
a solution with the blockchain by serving as a collective bargaining machine to solve
large-scale economic coordination problems like this, especially in the current case of
asymmetric power. Their governance could be jointly managed by creators and AI
providers.
In the context of AI, creators can leverage smart contracts to encode software-enforced
usage terms and restrictions - for instance, by setting conditions around commercial
applications like model training. Smart contracts would then automatically enforce
attribution systems, allocating portions of revenue generated by AI systems back to
contributing creators. Even if AI companies wanted to compensate creators at present,

35 of 128
 AI x Crypto Primer

the sheer number of individual creators would make this virtually impossible - but
smart contracts make compensation feasible. The composability of smart contracts
would also reduce lock-in to any single model, enabling a freer AI marketplace. This
competition would lead to more favourable profit-sharing for creators. Facing unified
terms enforced by an impartial protocol, AI companies would need to either accept
the collective bargain set by creators or go without their content; no longer could tech
giants exert unilateral leverage over individuals.
Centralised control over information agents like ChatGPT also raises concerns over is-
sues such as embedded advertising. While Google clearly demarcates ads separately at
the top of search results, an AI agent could seamlessly integrate paid recommendations
into responses. In contrast, crypto-based solutions allow for auditable AI agents.

3.3.4 AI Agents & Amazon

The natural extension from language models like ChatGPT answering queries is AI
programmes that can take actions on the user’s behalf, a move from information agents
to action agents (simply referred to as ‘agents’). These systems wouldn’t just find you
the best Bluetooth speaker, but order it directly to your house. Relying on closed-
source agents from companies like OpenAI for these tasks will give them immense
power beyond just the content creation market, potentially allowing them to dominate
industries like the $6.3 trillion e-commerce market. OpenAI would become not just the
next Google, but the next Amazon as well.
If a handful of big tech companies own the most powerful and widely used AI agents, it
gives them huge leverage over consumers and industries. These agents would mediate
a growing portion of our digital lives - shopping, travel, and finances. With no alter-
natives, there would be little choice but to rely on companies like OpenAI or Google.
Their closed-source agents would become virtual gatekeepers, controlling our access to
crucial services and information, and they could modify agent behaviours with little
accountability.
This concentration of power in proprietary AI has parallels to the rise of big tech
monopolies like Google and Facebook. But, the implications are exponentially greater
when AI agents can take action seamlessly across domains. That is why decentralised
blockchain alternatives are so vital - they introduce competition, user empowerment,
and transparency, acting against the risk of AI agent monopolisation in the hands of
big tech.
In summary, AI models like ChatGPT provide us with information agents, which can
read content and answer questions on our behalf. This will disrupt not only the way we
consume information but also Google’s business model and the livelihood of creators.
With this disruption in front of us, we have an opportunity to build a new internet
that is fairer to creators and rewards them for their work and contributions. AI agents

36 of 128
 The Big Ideas

capable of taking action (for instance, the purchasing of goods online) will disrupt e-
commerce. Again, blockchain gives a chance to establish an equitable model for the
internet. Will we learn from the mistakes of the Google and Amazon era?

37 of 128
 AI x Crypto Primer

3.4 Crypto Will Accelerate Open Source AI

Note: This is a draft.

This section will be updated in Version 2 of this report.

3.4.1 What AGI Means For Us

AI has become a staple of the 21st century. It has uses ranging from art to education, to
finance, to politics, to programming and many more applications. It is able to generate
photorealistic videos and images in seconds with rather limited prompts. Programmers
have resorted to using AI to develop efficient and useful code, outsourcing their own
labour to AI. The line between the real world and science fiction is blurring – and this
leaves a reduced role for humans in the labour market.
Therefore, it is likely that we will face a crisis in labour. If we can outsource a significant
level of our intellectual labour to AI, it may soon become economically preferable to
use AI rather than human labour. If human labour is being replaced by AI, we will
need to adapt to this new state of the labour market. Such adaptations will result in a
disruption in the way our current economic systems work. This section will deal with
the direction AI is heading and ways to prevent such crises from occurring through
Crypto solutions.

3.4.2 Why We Should Care About Open Source AI?

Generally, an open source software is software that is open for anyone to use, usually
attached with a license that states what can and cannot be done with the software.
Open source AI is for AI software to be open for anyone to use, with similar limitations
provided by licenses. Open source projects are typically organised with free participa-
tion in the development of the product in mind. It is centered around a community
which welcomes contributions of codes and error fixing, they tend to be collaborative
in nature. This section will provide reasons for why open source AI is significant.
Open source AI can be used to create a more competitive AI market. Competition
is good because there are a variety of people with different talents and varying levels
of skill at developing the product. Suppose that there is a firm which generally has a
lot of talent, but has made a massive mistake, a competitive market would encourage
people to fix that error and so limit the damage of that error. Open source AI lowers
the barrier for people to enter into the AI market, anyone can use the AI software and
contribute. Since anyone can enter through open source AI, it means that there are
more competitors in the market making the industry a much more competitive one
than it is at present.

38 of 128
 The Big Ideas

Open source AI can be used to safeguard the industry against bad actors seeking to
do harm. Technology is a powerful tool but itself is neutral. It can be used for the
purposes of benefitting or harming humanity, depending on whose hands the technol-
ogy is gripped by. It is preferable that we have good actors wielding this technology,
especially to counter those that are using technology to harm. Open source AI makes
it so that more people and talent enter the AI industry. This gives us a better chance
of good actors entering the AI industry who work to counter those that wish to use AI
for harm. A reduction in the bad actors using AI will lead us to a safer world.
Furthermore, there is a lowering of the barrier of entry to the AI industry, unlocking
a bank of talent and skill which only serves to develop the industry further. AI can
provide progress for us as a species. We have already outlined its wide usage today, but
the potential for development is immense, this is great for our progress as a species.
Open source AI will accelerate the development of AI through lowering the barrier of
entry for talent to enter into the industry. More talent means that we can make better
AI for even wider applications in society. Open source AI provides us the tool to access
a wellspring of talent and ability to develop this technology.
Open source AI provides developers with the freedom to specify and customise AI as
they see fit. One feature of AI is that it can be customised for specific purposes. A
customised AI can be directed to deal with specific needs and demands which improves
the quality of the product significantly. Developers’ ability to customise AI software is
often bound by the closed nature of the AI industry at present. What open source AI
can provide is that it allows the developer to freely customise and make their product
the best it can be for their purpose. Such freedom will make for a better market and
for better products to be brought out from it.
To conclude, we find that the Techno-Capital machine is one of perpetual growth. A
virtuous cycle is formed where development of technology is fuelled by capital gained
from the market and the market creates more capital through the development of tech-
nology. Open source AI provides a space for competition, low barriers to entry, freedom
and collaboration which only serves to increase innovation and diffuse AI technologies
stimulating economic growth and more markets. It is essential for both the industry’s
ability to grow and also for its benefit that it can provide to humanity.

3.4.3 OpenAI versus Open Source AI

In the AI industry, the leading company is OpenAI. Since releasing ChatGPT in 2022,
OpenAI has led the AI industry in profit and in knowledge. They have the backing of
Microsoft and big tech, and thus are well entrenched in the AI market. From this, it
seems that it is quite an uphill battle to have open source AI compete on the level of
OpenAI from our present perspective. Yet, we have good reasons to suggest that open
source AI has the promise to challenge and beat OpenAI in the market.

39 of 128
 AI x Crypto Primer

Open source AI is able to evade the inevitable regulations that will be placed upon
the AI industry by fearful governments. Presently, governments and regulatory bodies
around the world are scrambling to regulate and limit the emerging AI industry. The
target for this regulation will be the traditional AI that is used by OpenAI which is
centralised and kept by a single organisation. This regulation will limit the development
of the industry. However, open source AI has the advantage of being without a keeper
and being decentralised. This means that governments will find it difficult to regulate
open source AI. This provides open source AI the advantage of not being as limited by
regulations as OpenAI will be in the future.
Furthermore, we see that open source AI can benefit from OpenAI, whereas OpenAI
cannot benefit from open source AI. OpenAI’s main objective is to maximise profit
for themselves, this means they keep a lot of their models and data private to prevent
the competition from using it to their advantage. Open source AI can use licenses to
prevent OpenAI from benefitting from their less regulated developments. OpenAI will
find themselves isolated in the market, since they will be just one company against a
host of many different organisations which use open source AI. This will mean that
open source can benefit from OpenAI’s wealth of data and knowledge through limiting
OpenAI’s access initially. A new license may be needed, however, to eventually provide
OpenAI this access.
Finally, open source AI has a better chance of attracting ideologically minded people
with a passion for the betterment of AI than OpenAI. This is because open source AI
has such a low barrier of entry. However, one might suggest that this would result in
an difficult atmosphere to coordinate and develop products. Yet, it is not necessary
to have a lot of people on such projects. Linus Torvalds at Linux is such an example
where a single person contributing continuously to a project can have such a massive
impact. OpenAI’s high barrier of entry makes it difficult for passionate people to join.
Despite it seeming right now that OpenAI has a tight grip over the AI industry, there
are several routes that open source AI can rely upon to challenge OpenAI’s grip.
Open sources AI’s flexibility, accessibility and community-centred over profit-centred
approach to AI means that it has great tools to beat OpenAI through isolating them
in a consistently evolving and dynamic market. Certainly, open source AI has the po-
tential to beat OpenAI in the AI industry.

3.4.4 The Problem with Open Source AI.

However, there are still obstacles open source AI must overcome if it is able to take
the fight to big tech and OpenAI. These obstacles can be divided into three themes.
First, there is the lack of talent present in open source AI. Second, there is the lack
of compute present for open source AI to physically work. Third, there is a lack of
data for AI to train themselves upon and develop. This section will be dedicated to
elaborating upon these three problems.

40 of 128
 The Big Ideas

For any organisation to function, there needs to be talent to provide the innovative
ideas and work necessary for the development of a product. Open source AI faces a
major problem in that there is no profit or monetary incentive to work in these com-
munities. The majority of AI engineers, when deciding on how they want to work in
the industry are forced to choose between either a high-paying job at Big tech or run
the risk of entrepreneurship. The safe option, which is what most people take, is to
work at big tech and earn a living there. The best talent go to OpenAI rather than
work in an open source AI community with no monetary incentive. Therefore, open
source AI is unable to attract that best talent necessary for it to develop innovative
new products which can challenge big tech.
Another problem is that open source AI lacks access to the computing power necessary
for the scale at which OpenAI has access to. Larger scale AI requires more GPUs to
scale up the operation. GPUs are expensive, effectively only produced by a single com-
pany in Nvidia. The lack of money in open source AI mean that it is very difficult to
physically generate enough compute for AI models to compete on the same level as
OpenAI’s ChatGPT. Even Linux, which has high quality standards in software engi-
neering is limited by the compute scale for their programs. Open source AI lacks the
ability to access a supercomputer at will, and so it becomes difficult to compete with
OpenAI which does have that privilege.
AI models require data for them to train. Despite labelling themselves as ‘open’ or
‘open source’, big tech such as Meta and OpenAI keep the data that they use to train
their AI as private and kept only to them. They only publish the finished AI models
afterwards, such as ChatGPT. This data tends to be of high quality and high quantity
too, sourced from the mass userbase of both Meta and OpenAI. Open source AI has
the disadvantage of not having access to this high quality data in high quantity and
therefore they cannot train their AI models with the best and most amount of data
necessary for them to compete with Meta or OpenAI. Therefore, open source AI is not
in a state to develop products which can stand against OpenAI or Meta’s products.
What open source AI requires is a way to overcome these three major problems which
hinder their potential at challenging big tech’s grip over the AI inudstry. We suggest
that Crypto has a solution which can tackle these issues for open source AI.

3.4.5 Crypto’s Solution for Open Source AI.

We believe that Crypto has solutions for all three of the problems outlined in the
previous section. This section will be dedicated to proposing these solutions to each
specific problem.
First, Crypto can solve the talent problem through creating a revenue/reward system
in open source AI. Crypto can help open source AI through providing intrinsic rewards
for contributions to the project. An example of this is the AGI Guild, which is a group

41 of 128
 AI x Crypto Primer

of ideologically open source developers who have formed a license which can reward
developers for contributing to open source projects.
This works through having the license be governed by an AGI token. Companies which
are over 1 million USD must acquire a stake in the AGI token to use the license. This
stake is then distributed to contributors to the open source AI project. This provides
the monetary reward for the developer and gives the token value itself, encouraging
more contributions. There are voting systems as well for members to collectively decide
on how the project should go, encouraging meritocracy and democracy in open source
AI whilst garnering the ability to make an earning from open source AI.
Since there is now a monetary incentive, prospective and present talent will not need
to view their path as either safely choosing Meta or risking all on a venture but seeing
a third way where you can earn as you develop AI and not have to work with the
proprietary restrictions of a big tech company. Talent will be attracted to open source
AI projects as a viable competitor against big tech.
Second, Crypto can solve the compute problem by lowering the barrier for access to
servers and powers. Cloud providers could reject permission for developers to use their
servers for their products in the past. With crypto, it now becomes permissionless which
means that anyone can get the compute they need to use for their project. This means
that developers, using open source AI and Crypto, can now freely use as much compute
as they need.
Crypto also provides developers and communities with the bargaining power aginst the
cloud providers. In the short past, cloud providers could increase their prices for usage
since the open source AI developers needed their servers to run their programs. Now,
with Crypto, we can decentralise this systen and begin to challenge the cloud providers’
prices and open the doors to many more who want to develop open source AI.
Now, communities have the means and capabilities of challenging the cloud providing
companies like AWS to decrease their prices and therefore increase the quality of the
product and the work cone in the end.
Finally, Crypto can help solve the data problem by providing rewards for users who
generate data for an open source AI project. An example of this is Grass, which has
a browser extension which generates tokens as someone is browsing the internet. This
reward system means that open source AI is able to gain more and more data as time
goes on, as Crypto reveals more data. This is done in a decentralised way which means
that we don’t fall into the problem of regulators or big tech figuring this out and taking
measures to prevent it like banning your IP address. This can’t be effective against
this way because there is a host of people with this extension, banning one person
has little effect on the rest being able to scrape that data. This allows for open source
AI developers to get as good data as the big companies are able to attain. This will
certainly improve the training of the AI model. Even for data which is meant to be

42 of 128
 The Big Ideas

kept secret, Crypto can both keep the data secret and scrape it too given that the user
to whom the data belong is alright with it, it is able to be used for the betterment of
the AI model.
Overall, what we find is that Crypto is able to scrape and attend to the sensitivities
of data collection whilst benefitting the open source AI programs significantly through
giving them the data they need to train their AI models.
The ideal that we are seeking, where an open source AI system has beaten the private
propriety software based system is this. On the supplier side, open source AI will lead
to the best AI being present in the marketplace since there will be the most diverse and
best engineers working on developing AI with the best data. This continues to push
forward the AI industry and progresses it even further.
On the consumer side, we will see that consumers can choose from a wide variety of
specific and customised AI models which then allows for consumers to form an AI for
their needs rather than for the company’s needs. This customisation effectively brings
to the consumer the ability to train an AI themselves for their own purposes.

43 of 128
 Opportunities for Builders

Note: This is a draft.

This section will be updated in Version 2 of this report.

While writing this report, we received many suggestions from the community of po-
tential ideas for builders in the AI x Crypto space. We’ve compiled a curated list of
suggested ideas we’re excited about to get you started. We present each problem and
a potential solution. If you’re interested in any of these ideas, get in touch.

4.1 A Curated List of Suggested Ideas

4.1.1 AI DAO
Category: AI helping Crypto internally -> AI helping
Problem: DAOs rely on an engaged community to vote on proposals thoughtfully.
Every proposal needs to be voted on manually which slows things down and leads to
DAO fatigue.
Potential Solution: Automate proposals by giving members of a DAO an AI Agent
that votes on their behalf based on their values. The model could use a Bayesian ap-
proach which takes into account the uncertainty in the decision being made. If the
uncertainty is above a threshold, then bring the proposal to the attention of the mem-
bers of the DAO and require a manual vote.
The onboarding process for creating these AI agents needs to be streamlined for users
to adopt this. A combination of on-chain and off-chain data from each user can be used
to define the preferences of a user without too much manual work. For example, DAO
Base is looking into solutions around the idea that what you do in the past defines you.
One potential approach is to query an LLM on-chain, passing in the proposal along with
the DAO member’s prompt so that they can adjust according to their requirements.

44 of 128
 Opportunities for Builders

1 pragma solidity ^0.8.24;

2
3 contract Example {
function decide(string memory proposal) public pure returns (Choice, string
4
memory) {
5 string memory result = llmInference(string.concat(prompt, proposal));
6
7 string memory v = substring(result, 0, 1);
8 string memory reason = substring(result, 2, bytes(result).length);
9

if (keccak256(abi.encodePacked((v))) ==
10
keccak256(abi.encodePacked(("Y")))) {
11 return (Choice.For, reason);
} else if (keccak256(abi.encodePacked((v))) ==
12
keccak256(abi.encodePacked(("N")))) {
13 return (Choice.Against, reason);
} else if (keccak256(abi.encodePacked((v))) ==
14
keccak256(abi.encodePacked(("A")))) {
15 return (Choice.Abstain, reason);
16 }
17
18 revert(string.concat("Unknown vote: ", v, " with reason: ", reason));
19 }
20 }

Listing 1: This is a simple example but tooling would be needed to make it

easy for users to specify their requirements and automate the voting itself.

See here for demo code.

4.1.2 Verifiable Training

Problem: It’s not always possible to tell what data a model was trained on, even if you
have access to the model weights, because training models is a compression of training
data. This introduces several challenges that do not exist in traditional software:
• Complicates copyright
• Harder to fairly compensate the owners of the data
• Harder to know who trained each part of the model in settings where training is
done by more than one party, e.g. Model Zoo
• Easier to add biases in models
Potential Solution: Make the training process itself verifiable. Build tools to break
down how a model was trained and check if it contains a given piece of data. Several
approaches can be explored:
Integrate cryptographic primitives into the training process itself. For example, Pytorch
NFT Callback hashes the current network weights, some metadata (data, accuracy,
etc…) and your eth address every N epoch, which proves who did the model training.
Note: This approach introduces a performance overhead to training models.
An alternative solution would be to train the model on a purpose-built decentralised
network. There are solutions based on traditional consensus mechanisms such as BFT.
However, BFT requires more than 23 of nodes to be reliable and honest, leading to the

45 of 128
 AI x Crypto Primer

result that the minimum number of nodes for a BFT consensus is 𝑁 = 3𝑓 + 1 where 𝑓
gives the number of nodes that either (i) fails or (ii) maliciously acts. This introduces a
lot of redundant work (linear in 𝑁 e.g. 60x for 𝑁 = 60). An example of this approach
is Proof of Training. We suggest an alternative decentralised network that is built from
the ground up for this purpose.

4.1.3 Alternative Paths to Verifiable Inference

Problem: Most of the research in verifiable ML is on Zero-Knowledge Machine Learn-
ing (zkml). However, zkml currently has a performance overhead of 1000x and cannot
run large models yet.
Potential Solution: Several approaches are being explored. This is a relatively new
field and there is an opportunity for different approaches that make different tradeoffs.
Ora is experimenting with opML. This approach involves a single party ‘optimistically’
inferencing a model, putting the result on-chain, and incentivising verifiers to challenge
incorrect results by paying them tokens.
Aizel is building a solution based on Multi-Party-Computation (MPC) and Trusted
Execution Environments (TEE). Their aim is to do verifiable inference at the same
cost as normal inference.
EZKL is working on splitting and parallelising zk proofs, which would make proving
large models feasible.

4.1.4 DePin 2.0

Problem: As DePin (Decentralized Personal Internet) technologies and robotics in-
tersect, the unique challenges of integrating dynamic, autonomous systems like ro-
bot swarms become evident. Unlike static, sensing devices such as solar panels, ro-
bot swarms possess the ability to act independently, which introduces vulnerabilities.
Specifically, a single malfunctioning or malicious (Byzantine) robot within a swarm can
compromise the integrity of the entire system. Given the irreversible nature of actions
taken by robots, ensuring the reliability and security of these swarms is paramount.
Potential Solution: A potential solution involves leveraging zero-knowledge proofs
(ZKPs) to certify the execution of specific swarm strategies without revealing the strat-
egy’s details. This cryptographic technique can validate that a swarm operated accord-
ing to a pre-defined algorithm, based on hardware-attested data collected from the
swarm and cryptographically signed information from external sensors. By incorporat-
ing ZKPs, we can establish a trustless environment where the behavior of robot swarms
is verified in real-time, mitigating the risk posed by Byzantine robots. This approach
not only enhances the security and reliability of robot swarms but also aligns with the
decentralized ethos of DePin, ensuring transparency and trust without compromising
sensitive operational details.

46 of 128
 Opportunities for Builders

4.1.5 Transparent LLM Drift

Problem: Large Language Models (LLMs) are central to many applications, yet they
suffer from unpredictability and performance drift over time. Traditional benchmarks,
like the Open LLM Leaderboard by HuggingFace, are criticized for being gameable and
lacking historical data, making it difficult to track and understand LLM performance
changes.
Potential Solution: Integrate automated benchmarks with community voting, using
tokens as incentives for participation. Results are stored on a blockchain, ensuring
transparency and immutability. This approach aims to provide a reliable, transparent
history of LLM performance, addressing issues of unpredictability and performance
drift.

4.1.6 Truth Marketplace

Problem: As AI-generated content becomes more prevalent, distinguishing between
true and false information online grows more complex. This ambiguity challenges the
traditional verification methods, making it hard to maintain trust in digital content.
Potential Solution: An idea to consider is extending the concept of a proof market-
place to address the verification of truth in a more nuanced, fuzzy environment. By
introducing economic incentives, such as those seen in existing proof marketplaces, but
tailored to handle ambiguity, this system could encourage the discovery and validation
of truth. Bittensor’s Yuma Consensus would be the ideal venue for a Truth Marketplace
as deals well with ambiguity. This approach wouldn’t prescribe a specific method but
would create a framework where various strategies for identifying truth are economi-
cally incentivized, fostering an ecosystem where truth has a tangible value.

4.1.7 Trustless Prediction Markets Resolved by AI

Problem: The challenge with prediction markets is that participants often disagree on
the meaning of the events being bet on, including what counts as evidence for winning
or losing, leading to confusion and disputes.
Potential Solution: Resolve prediction markets by feeding information on a pre-
agreed date into a pre-agreed AI model. The information comes from a pre-agreed data
source (which could leverage the prior idea of a Truth marketplace).

47 of 128
 How Does Machine Learning Work?

Before diving into the intersection of artificial intelligence (AI) and crypto, it’s first
crucial to cover some concepts in the field of AI separately. Since this report is written
for readers in the crypto space, a deep understanding of concepts in AI and machine
learning isn’t always present. This is crucial so that the reader can assess which ideas
at the intersection of AI and crypto have substance, and to accurately asses the tech
risk of projects. This section focuses on AI concepts; however, hints for how this relates
to crypto are highlighted throughout.

Hints throughout this chapter give an idea as to how AI concepts relate to

crypto.

Summary of topics covered in this section:

• Machine learning (ML) is a branch of artificial intelligence where a machine is
taught to make decisions from data without being explicitly programmed.
• The ML pipeline is divided into three steps: data, training and inference.
• Training a model is very expensive computationally, while inference is relatively
cheap.
• There are three main types of learning: supervised learning, unsupervised learn-
ing and reinforcement learning.
• Supervised learning refers to learning from examples (provided by a teacher). A
model can be shown pictures of dogs and told by the teacher that these are dogs.
Then the model learns to distinguish dogs from other animals.
• Many popular models, such as LLMs (e.g. GPT-4 and LLaMa) are trained using
unsupervised learning. In this mode of learning, no guidance or examples are
provided by a teacher. Rather, the model learns to find patterns in the data.
• Reinforcement learning (learning from trial and error), is mainly used in sequen-
tial decision-making tasks such as robotic control and playing games (e.g. chess
or Go).

1.1 Artificial Intelligence and Machine learning

In 1956 some of the brightest minds of the time came together for a workshop. They
aimed to come up with general principles of intelligence. They noted that:
“Every aspect of learning or any other feature of intelligence can be so precisely de-
scribed that a machine can be made to simulate it.”

48 of 128
 How Does Machine Learning Work?

In these early days of AI, researchers were full of optimism. In some sense, they were
aiming for Artificial General Intelligence (AGI), quite ambitious. As we now know,
these researchers didn’t manage to create AI agents with general intelligence. Neither
did AI researchers in the 70′s and 80′s. In that period AI researchers attempted to
develop ‘Knowledge-based systems’.
The key idea in knowledge-based systems is that we can write very precise rules for a
machine to follow. Essentially, we extract very specific and precise domain knowledge
from an expert and write it down for a machine in the form of rules. The machine
can then use these rules to reason and make sound decisions. For instance, we can
attempt to distil all principles of playing chess from Magnus Carlson and build an AI
to play chess.
However, this is extremely hard to do and even if it’s possible, a tremendous amount
of manual effort is required to create those rules. Imagine trying to write down rules
for how to recognise a dog to a machine. How can the machine go from having pixels
to knowing what a dog is?
Recent progress in Artificial Intelligence came from a branch referred to as Machine
Learning. In this paradigm, rather than writing precise rules for a machine, we use
data and allow the machine to learn from it. Modern AI tools using machine learning
can be found everywhere, for example, GPT-4, FaceID on your iPhone, Gaming bots,
Gmail spam filter, models for medical diagnosis, self-driving cars … etc.

1.2 The Machine Learning Pipeline

The machine learning pipeline can be divided into three main steps. You have data, we
train the model and then when we have the model, we can use it. Using the model is
called inference. So the three steps are data, training, and inference.
At a high level, the data step involves finding the relevant data and pre-processing it.
For example, if we’re building a model to classify dogs, we need to find pictures of dogs
and other animals so that the model can learn what’s a dog and what’s not a dog. Then
we need to process the data and make sure it is in the right format that the model can
accept to learn properly. E.g. we might require the picture to be the same size.
Then the second step, training, is the step where we use data to learn what the model
should look like. What are the equations inside the model? What are the weights of the
neural network? What are the parameters? What are the computations that are being
carried out? And then if that model is good, we can test its performance and then we
can use it. And that brings us to step three.
Step three is called inference, i.e. we just use the neural network. For example, give the
network an input and ask the question, do inference to generate the output?

49 of 128
 AI x Crypto Primer

Figure 28: The three main steps in a machine-learning pipeline are Data,
Training and Inference.

The three main steps in a machine-learning pipeline are Data, Training and Inference.

1.2.1 Data
Now, let’s look a little bit deeper into each step. First: data. This, broadly speaking,
means we have to collect the data and we have to pre-process it.
Let’s look at an example. If we want to build a model to be used by a dermatologist
(doctor specializing in the treatment of skin disorders). We start by collecting data on
many human faces. Then we get professional dermatologists to assess whether there is
a condition present or not. Now many challenges could arise. First, if all the data we
have includes human faces, the model would struggle to identify any skin condition on
another part of the body. Second, the data might have a bias in it. For instance, most
of the data might be pictures from one skin colour or tone. Third, the dermatologists
might make mistakes, which means we will have false data. Fourth, the data we got,
might violate privacy concerns.

Economic incentives can help collect high-quality data, and privacy-preserving

technologies can address user concerns in sensitive settings such as the example
above.

We’re going to cover challenges in data deeper in Chapter 2. However, this gives you
some idea that collecting good data and pre-processing it can be quite challenging.

50 of 128
 How Does Machine Learning Work?

Figure 29: An illustration of two popular datasets. MNIST contains hand-

written digits, while ImageNet includes millions of annotated images from
different categories.

In machine learning research, many famous data sets exist. A few commonly used ones
are:
• MNIST Dataset:
• Description: Contains 70,000 handwritten digits (0-9) in a greyscale im-
age format.
• Use Case: Primarily used in computer vision for techniques in handwrit-
ten digit recognition. It’s a beginner-friendly dataset often used in educa-
tional settings
• ImageNet:
• Description: A large database of over 14 million images annotated with
labels spanning over 20,000 categories.
• Use Case: Used for training and benchmarking algorithms in object de-
tection and image classification. The annual ImageNet Large Scale Visual
Recognition Challenge (ILSVRC) has been a significant event in advanc-
ing computer vision and deep learning techniques.
• IMDb Reviews:
• Description: Contains 50,000 movie reviews from IMDb, split into two
sets: training and test. Each set contains an equal number of positive and
negative reviews.
• Use Case: Widely used for sentiment analysis tasks in natural language
processing (NLP). It helps in developing models that can understand and
classify the sentiments (positive/negative) expressed in text.
Access to large, high-quality datasets is extremely important to train good models.
However, this can be challenging, especially for smaller organizations or individual re-
searchers. Since data is so valuable, it’s often not shared by large organizations as it
provides a competitive advantage.

51 of 128
 AI x Crypto Primer

Solutions to balance transparency and openness with the ability to make a profit
can propel the quality of open datasets.

1.2.2 Training
The second step in the pipeline is training a model. So what does training a model
actually mean? First, let’s look at an example. A machine learning model (after it has
been trained) is typically just two files. For instance, LLaMa 2 (a large language model,
similar to GPT-4), would just be two files:
• ‘parameters’, a 140GB file which includes numbers.
• ‘run.c’, and a simple file (around 500 lines of code).
The first file includes all the parameters of the LLaMa 2 model, and run.c includes
instructions for how to do inference (use the model). Those models are often neural
networks.

Figure 30: A basic illustration of neural network.

In a neural network like the one above, every node has a bunch of numbers in it. Those
numbers are called parameters and are stored in the file (surprise!) parameters. The
process of obtaining those parameters is called training. One very high-level, here is
how the process works.
Imagine training a model to recognize digits (from 0 to 9). We first collect data (in this
case we can use the MNIST data set). Then we start training the model.
• We take the first data point, which is ‘5’.
• Then we pass the image (of ‘5’) through the network. The network performs
mathematical operations on the input image.
• The network will output a number between 0 to 9. This output is the prediction
of the current network of what this image is.

52 of 128
 How Does Machine Learning Work?

• Now
there are 2 cases. The network is either right (it predicted ‘5’) or wrong
(any
other number).
•
If it predicted the number correctly, we don’t do much.
•
If it predicted number is incorrect, we will go back through the network
and make small changes to all the parameters.
• After we make these small changes, we try again. Technically, the network now
has new parameters, so it will have different predictions.
• And we keep doing that for all the data points until the network is mostly correct.
This process is inherently sequential. We first pass a data point through the whole
network, see what what prediction is, and then update the weights of the model. Thus,
training a model across different machines presents many challenges as we might need
to split the data (data parallelism) or the model itself (model parallelism).
The training process can be more comprehensive. First of all, we have to pick the model
architecture. What type of neural network should we choose? Not all machine learning
models are neural networks. Second, after we figure out which architecture is best for
our problem, or at least the architecture we think is best, we need to figure out a
training process. E.g. in what order will we pass the data through the network?
Third, we need a hardware setup. What kind of hardware are we going to be using
(CPU, GPU, TPU)? And how are we going to train that?
Finally, while we’re training the model, we want to validate that this model is actually
good. We want to be testing at the end of the training that this model gives us the
desired output. Spoiler (which isn’t really a spoiler), training a model can be very
expensive computationally. Any small inefficiency can compound into large costs. As
we will see later on, specifically for really large models like LLMs, inefficient training
could cost you millions of dollars.
Again, in Chapter 2, we’ll discuss challenges in training models in more detail.

1.2.3 Inference
The third step in the machine learning pipeline is inference, which means using the
model. When I use ChatGPT and it responds, the model is performing inference. If I
use my iPhone to unlock it with my face, the face ID model recognizes my face and
opens my phone. That model performed inference. There already had been data, the
model has already been trained, and now that the model is trained, we can use it, and
using it is inference.
Inference is technically the same thing as a prediction made by the network during
the training phase. Recall that a data point is passed through the network, and a pre-
diction is made. Then the parameters of the model are updated based on the quality
of the prediction. Inference works functionally the same. Thus, inference is very cheap

53 of 128
 AI x Crypto Primer

computationally compared to training. Training LLaMa might cost tens of millions of

dollars, but performing inference once is a fraction of a cent.

Performing inference is cheap and can be parallelized more easily compared to

training, this might present an opportunity for small idle compute (e.g. laptops
and phones).

There are a few steps involved in the inference process. First, before actually using
something in production, we need to test it. We perform inference on data not seen
during the training phase to validate the quality of the model. Second, when we deploy
a model, there are some hardware and software requirements. For instance, if I have the
face ID model on my iPhone, that model can sit on some server from Apple. However,
that is very inconvenient, because now every time I want to unlock my phone, I have
to access the Internet and send a request to the Apple server and do inference on that
model. However, to use this all the time, the model doing face ID needs to live on your
phone, which means that that model needs to be compatible with the type of hardware
available on your iPhone.
And finally, in practice, we often have to maintain this model. We have to make tweaks
here and there. The models we train and use aren’t always perfect. Hardware require-
ments and software requirements change often.

1.2.4 The ML pipeline is iterative

So far, I framed the pipeline as if it were three steps in sequence. You get the data,
you process it, you clean it, everything’s amazing, and then you train your model, and
then after the model is trained, you perform inference. That’s a very rosy picture of
what machine learning in practice looks like. In reality, there’s a lot of iteration going
on. So it’s not a chain, it has a few loops as illustrated in the following figure.

Figure 31: The machine learning pipeline can be image as a chain of three
steps: data, training and inference. However, in practice the process is more
iterative as illustrated by the blue arrows.

54 of 128
 How Does Machine Learning Work?

To understand this, we can give you a few examples. For instance, we might collect
data about a model and then attempt to train it. While training it we realize that
the amount of data we need should to be much higher. This means we have to pause
training, go back to the data step and get more data. We might either need to reprocess
the data, or we might do some form of data augmentation. Data augmentation is like
giving your data a makeover to create new looks from the same old stuff. Imagine you
have a photo album and you want to make it more interesting. You take each photo
and make a few copies, but in each copy, you change something small – maybe you
rotate one picture, zoom in on another, or change the lighting in another. Now, your
album has more variety, but you haven’t actually taken any new photos. For instance, if
you’re training a model to recognize dogs, you might just horizontally flip every picture
and feed that to the model as well. Or, we change the pose of the dog in the picture as
illustrated below. This increases the data set as far as the model is concerned, but we
didn’t go out in the real world and collect more data.

Using synthetic data to train a machine learning model can cause many issues,
thus being able to prove the authenticity of data might become critical in the
future.

Figure 32: Examples of Data augmentation. The original data point is mul-
tiplied without the need to go out in the world and collect more unique data
points.

A second and more obvious example of iteration is when we actually train a model and
then we go to use it in practice, i.e. do inference, we might find out that the performance
is bad in practice or there’s bias. This means we have to stop the inference process, go
back and retrain the model to get rid of those issues, like bias and proof.
A third and very common step is, once we’re using a model in practice (performing
inference), we end up making changes to the data step because the inference itself
generates new data. For example, imagine building a spam filter. First, we collect data.
The data in this case would be, a set of spam emails and non-spam emails. When the
model is then trained and used in practice, I might get a spam email in my inbox,
which means the model made a mistake. It didn’t classify it as spam, but it is spam.
So when a user on Gmail selects that ‘this email belongs to spam, a new data point

55 of 128
 AI x Crypto Primer

is created. After that, all those new data points go to the data step, and then we can
improve the model’s performance by doing a little bit more training.
Another example is, imagine an AI playing chess. The data we need to train an AI to
play chess is a lot of chess games, with the outcome of who won and who lost. But
when this model is used in practice, to play chess games, that generates more data for
the AI. This means, we can go from the inference step back to the data with those
new data points to improve my model again. This idea of inference and data being
connected applies in many settings.
This section aims to give you a high-level idea about the process of building a machine
learning model is very iterative. It’s not like, “Oh, we just get data, train a model in
one try, and put it into production”.

Models are updated often, so an *immutable record* might bring challenges in

design.

1.3 Types of Machine Learning

We’ll cover three main types of machine learning models.
• Supervised learning: “Sensei, teach me the way”
• Unsupervised learning: “Just find hidden patterns”
• Reinforcement learning: “Just try sh*t, and see what works”

1.3.1 Supervised Learning

“Sensei, teach me the way”
Imagine you’re teaching a child to differentiate between cats and dogs. You (the teacher
who knows everything perfectly) show them lots of pictures of cats and dogs, and each
time you tell them which is which. Eventually, the child learns to recognize the differ-
ence on their own. That’s pretty much how supervised learning in machine learning
works.
In supervised learning, we have a large amount of data (like pictures of cats and dogs),
and we already know the answers (you the teacher told them which is a dog and which
is a cat). We use this data to train a model. The model looks at many examples and
effectively learns to imitate the teacher.
In this example, every picture is a raw data point. The answers (dog or cat) are called
‘labels’. This is thus a labelled data set. Each data point contains a raw image and a
label.

56 of 128
 How Does Machine Learning Work?

This approach is simple conceptually, and extremely powerful. Many applications in

medical diagnosis, self-driving cars, and stock price prediction using supervised learning
models.
However, as you can imagine, there are many challenges with this approach. For ex-
ample, it’s not only enough to get a large amount of data, we also need labels. This
can be extremely expensive. Companies like Scale.ai provide valuable services in this
context. Data labelling presents many challenges for robustness. Humans labelling data
make mistakes or simply disagree on the labels. It’s not uncommon for 20% of all labels
collected from humans to be unusable.

Incentives and other game-theoretic dynamics might aid in improving the quality
of open datasets.

1.3.2 Unsupervised Learning (USL)

“Just find hidden patterns”
Imagine you have a big basket full of different kinds of fruit, but you’re not familiar
with all of them. You start sorting them into groups based on what they look like, their
size, color, texture, or even smell. You’re not exactly sure what each type of fruit is
called, but you notice that some fruits are similar to each other. I.e. you found some
patterns in the data.
This scenario is similar to unsupervised learning in machine learning. In unsupervised
learning, we give a model a bunch of data (like a mix of various fruits), but we don’t
tell the model what each piece of data is (we don’t label the fruits). The model then
examines all this data and tries to find patterns or groupings by itself. It might group
the fruits based on color, shape, size, or any other feature it finds relevant. However,
the features the model finds, are not always relevant. This causes many issues as we’ll
see in Chapter 2.
For instance, the model might end up grouping bananas and plantains together because
they’re both long and yellow, while apples and tomatoes might end up in another group
because they’re round and can be red. The key point here is that the model is figuring
out these groupings without any prior knowledge or labels - it’s learning from the data
itself, just like you sorting unknown fruits into groups based on their observable char-
acteristics.
Unsupervised learning is the backbone is many popular machine learning models, such
as Large Language model (LLMs). ChatGPT is not taught by humans how to say each
sentence by providing labels. It simply analyzes patterns in language data and learns
to predict the next word.

57 of 128
 AI x Crypto Primer

Many other powerful Generative AI models rely on Unsupervised learning. For instance,
a GAN (Generative Adversarial Network) can be used to generate the face of a human
(even though the human doesn’t exist). See https://thispersondoesnotexist.com/

Figure 33: AI-generated image from https://thispersondoesnotexist.com

Figure 34: A second AI-generated image from https://thispersondoesnotexis

t.com

A second AI-generated image from https://thispersondoesnotexist.com/

The images above are AI-generated. The model was not taught ‘what a human face is’.
It was trained on a large set of human faces and using a clever architecture we can use
this model to generate faces that seem real. Note that with the rise of generative AI
and improved models, it’s increasingly harder to authenticate content.

Cryptographic solutions can allow us to track the origins of content and

scalablely allow us to use generative AI safely.

1.3.3 Reinforcement Learning (RL)

“Try sh*t and see what works” or “Learn from trial and error”

58 of 128
 How Does Machine Learning Work?

Imagine you’re teaching a dog to do a new trick, like fetching a ball. Each time the dog
does something close to what you want, like running towards the ball or picking it up,
you give it a treat. If the dog does something unrelated, like running in the opposite
direction, it doesn’t get a treat. Gradually, the dog figures out that fetching the ball
equals yummy treats, so it keeps doing it. This is basically what reinforcement learning
(RL) is about in the machine learning world.
In RL, you have a computer program or an agent (like the dog) that learns to make
decisions by trying out different things (like the dog trying different actions). The agent
gets rewards (treats) for desirable actions (like fetching the ball) and no rewards for
undesirable actions. Over time, the agent learns to do more of the good stuff that earns
rewards and less of the stuff that doesn’t. Formally, it’s maximizing a reward function.
Here’s the cool part: the agent figures out all this by itself through trial and error. Now
if we want to build an AI to play chess, the AI can just try random moves initially. If
it ends up winning the game, the AI gets a reward. The model then learns to do more
of the winning moves.
This can be applied to many problems, especially problems that require sequential de-
cision-making. For example, RL methods are used in Robotics & control, playing chess
or Go (e.g. AlphaGo), and algorithmic trading.
There are many challenges with RL methods. One is, that the agent might need a long
time before it ‘learns’ in meaningful strategy. This is acceptable for an AI learning to
play chess. However, would you put your personal capital in algorithm trading AI when
it starts taking random actions to see what works? Or would you allow a robot in your
house if it will take random actions initially?

Figure 35: Here are some videos of reinforcement learning agents during
training: a real-legged robot, and a simulated robot.

Below is a short summary of example applications for each type of machine learning.

59 of 128
 AI x Crypto Primer

Learning Type Applications

Supervised Email spam detection, image classification,

Learning speech recognition, predicting stock prices,
medical diagnosis, sentiment analysis, and
weather forecasting.

Unsupervised Customer segmentation, anomaly detection,

Learning social network analysis, organizing large
databases of information, natural language
processing, and recommendation systems.

Reinforcement Autonomous vehicles, robotic control, machine

Learning playing chess or Go, algorithmic trading,
recommendation systems, and natural language
processing tasks like text summarization.

60 of 128
 Challenges in Machine Learning

This chapter provides an overview of problems in the field of machine learning. We will
selectively expand on certain problems in the space. This is for two reasons: 1) brevity,
a comprehensive overview of challenges in the field accounting for nuances would result
in a very long report, and 2) we will focus on problems that will be relevant when
discussing the intersection with crypto. However, this section itself is written just from
the AI perspective. I.e. we will not be discussing cryptography methods in this section.
Summary of topics covered in this section:
• Data challenges are vast from bias to accessibility. Additionally, adversarial at-
tacks exist on the data level that causes machine learning models to make mis-
takes.
• Model collapse occurs when a model (such as GPT-X) is trained on synthetic
data. This causes irreversible damage to it.
• Labelling data can be very expensive, slow and unreliable.
• Training machine learning models have many challenges depending on the archi-
tecture.
• Model parallelism poses significant challenges such as communication overhead.
• Bayesian models can be used to quantify uncertainty. I.e. when performing in-
ference the model returns the degree to which it’s certain (e.g. 80% certain).
• LLMs have specific challenges such as hallucination and training difficulties.

2.1 Data Challenges

Data is key for any type of machine learning model. However, the requirement and scale
of data differ depending on the methods used. Raw data (unlabelled data) is needed in
both supervised and unsupervised learning.
In unsupervised learning one only has raw data, no labels are needed. This alleviates
many issues related to labelling data sets. However, raw data needed in unsupervised
learning still presents many challenges. This includes:
• Data Bias: Bias in machine learning occurs when the training data is not repre-
sentative of the real-world scenario it’s meant to model. This can lead to skewed
or unfair outcomes, such as a facial recognition system that performs poorly on
certain demographic groups because they are underrepresented in the training
data.
• Imbalanced Data Sets: Often, the data available for training is not evenly dis-
tributed among different classes. For example, in a disease diagnosis application,

61 of 128
 AI x Crypto Primer

there might be many more examples of ‘no disease’ cases than ‘disease’ cases.
This imbalance can lead the model to perform poorly on the minority class. This
issue is distinct from bias.
• Quality and Quantity of Data: The performance of a machine learning model
is highly dependent on the quality and quantity of the training data. Insufficient
or poor-quality data (like low-resolution images or noisy audio recordings) can
significantly hinder a model’s ability to learn effectively.
• Data Accessibility: Access to large, high-quality datasets can be a challenge,
especially for smaller organizations or individual researchers. Large tech compa-
nies often have an advantage in this regard, which can lead to disparities in the
development of machine learning models.

This is a significant challenge. However, crowd-sourcing data sets also come with
issues, such as guarantees on the quality of the data. Economic incentives and a
game-theoretic design can aid in the creation of open high-quality data sets.

• Data Security: Protecting data against unauthorized access and ensuring its
integrity during storage and use is crucial. Security breaches not only compromise
privacy but can also lead to tampering with data, affecting model performance.
• Privacy Concerns: As machine learning often requires vast amounts of data,
handling this data can raise privacy issues, especially if it includes sensitive or
personal information. Ensuring data privacy means respecting user consent, se-
curing data against breaches, and complying with privacy regulations like GDPR.
This can be very challenging (see example below).

Deleting data about specific users (to comply with GDPR) is very challenging in
a machine-learning model. Unlike a database, we can't just delete an entry. The
model parameters are adjusted based on all data provided, and removing
information about a specific person after a model has been trained is extremely
difficult.

62 of 128
 Challenges in Machine Learning

Figure 36: A particular issue with data privacy arises from the nature of
the machine learning model. In a regular database, I can have entries about
multiple people. If as I company I am required to delete that information, you
can simply remove that from the database. However, when I model is trained,
it simply holds parameters that approximate the whole training data. It’s not
clear which number corresponds to which database entry from the training.

2.1.1 Model collapse

One particular challenge we would like to highlight in unsupervised learning is model
collapse.
In this paper, the authors perform an interesting experiment. Models like GPT-3.5 and
GPT-4 are trained using all data available on the web. However, these models are now
being widely used and thus a lot of content on the internet a year from now will be
generated by these models. This means that GPT-5 and beyond will be trained on data
generated by GPT-4. What is the effect of training a model on synthetic data? They
find that training a language model on synthetic data causes irreversible defects in
the resulting models. The authors of the paper state “We demonstrate that it has
to be taken seriously if we are to sustain the benefits of training from large-scale data
scraped from the web. Indeed, the value of data collected about genuine human inter-
actions with systems will be increasingly valuable in the presence of content generated
by LLMs in data crawled from the Internet.”

This might suggest that there are major opportunities for solutions around data
provenance (tracing the origins of data).

63 of 128
 AI x Crypto Primer

Figure 37: An illustration of model collapse. As more content on the internet

is generated using the AI model, the next generation of the model is likely to
contain synthetic data in the training set. As shown in this paper

Note that this phenomenon is not specific to LLMs and can affect various machine
learning models and generative AI systems (e.g. Variational Autoencoders, Gaussian
Mixture Models).
Now, let’s consider supervised learning. In supervised learning, one requires a labelled
data set. This means, the raw data itself (a dog picture) and a label (“dog”). Labels are
manually chosen by the designer of the model and can be obtained using a combination
of human labelling and automated tools. This brings many challenges in practice. This
includes:
• Subjectivity: Deciding on labels for data can be subjective, leading to ambigu-
ity and potential ethical concerns. What one person considers appropriate may
be seen differently by another.
• Variance in labels: repeated run by the same human (let alone different hu-
mans) can provide different labels. This provides a noisy approximation of the
‘real label’, and leads to the need for a Quality Assurance layer. For instance, a
human might be presented with a sentence and tasked with labelling the senti-
ment of the sentence (”happy", “sad”, …etc). The same human sometimes labels
the exact same sentence differently. This reduces the quality of the datasets as it
introduces variance in the labels. It’s not uncommon for 20% of collected labels
to be unusable in practice.

Imagine creating a dataset to predict the quality of new protocols on the

blockchain. You're likely to get a wide range of scores depending on both the
subjectivity in the scoring system you chose, as well as the variance in the
opinion of people you survey.

• Lack of expert annotators: for a niche medical application, one might strug-
gle to get a meaningful amount of labelled data. This is due to the sacristy of
the people who can provide those labels: expert medical professionals.

64 of 128
 Challenges in Machine Learning

• Rare events: for many events, it’s extremely hard to get a large set of labelled
data because the event itself is rare. E.g. a computer vision model learning to
spot a shooting star.
• High costs: when attempting to collect a large high-quality dataset, the cost
can be incredibly high. These costs are especially high if the data sets need to
be labelled due to the above issues.
And many more issues like dealing with adversarial attacks and the transferability of
labels. To give the reader some intuition on dataset sizes, see the below chart. A dataset
like ImageNet contains 14 Million labelled data points.

Figure 38: An illustration of the size of various machine learning data sets.
Common Crawl is approximated as 1 billion web pages, so the total number
of words far exceeds this. A small data set (e.g. Iris) contains 150 images.
MNIST has around 70,000 images. Note that this is a log scale.

2.1.2 Data collection in reinforcement learning

In reinforcement learning, data collection poses unique challenges. Unlike supervised
learning, where data is pre-labelled and static, reinforcement learning relies on data
generated through interactions with an environment, often requiring complex simula-
tions or real-world experimentation. This presents some challenges:
• This process can be resource-intensive and time-consuming, particularly for
physical robots or complex environments. A robot learning from trial and error
might cause an accident if it’s trained in the real world. Alternatively, consider
a training bot learning from trial and error.
• Sparse and delayed rewards: the agent may need to explore a vast number of
actions before receiving meaningful feedback, making it difficult to learn effective
strategies.
• Ensuring the diversity and representativeness of the collected data is crucial;
otherwise, the agent might overfit to a narrow set of experiences and fail to

65 of 128
 AI x Crypto Primer

generalize. Balancing exploration (trying new actions) with exploitation (using

known successful actions) further complicates data collection, requiring sophis-
ticated strategies to gather useful data efficiently.
One interesting point to highlight is that data collection is directly tied to inference.
When training a reinforcement learning agent to play chess, we can collect data using
self-play. Self-play is like playing chess against yourself to get better. The agent plays
against a copy of itself, creating a loop of continuous learning. This method is great for
collecting data because it constantly generates new scenarios and challenges, helping
the agent learn from a wide range of experiences. This process can be parallelized across
many machines. And since inference is cheap computationally (compared to training),
this process has low hardware requirements. After the data is collected by performing
self-play, all data is collectively used to train the model and improve it.

Idle compute could be powerful in distributed inference and data collection as

the hardware requirements are much lower than training.

2.1.3 Adversarial data attacks

• Data poisoning attacks: in this type of attack, the training data is corrupted
by adding perturbations to deceive the classifier, leading to incorrect outputs.
For example, one might add spam elements to non-spam emails. This leads to
worsened performance in the future when that data is included in the training
of the spam filter. This can be done by increasing the usage of words like “free”,
“win”, “offer”, or “token” in non-spam contexts.
• Evasion attacks: the attacker manipulates the data during deployment to deceive
previously trained classifiers. Evasion attacks are the most prevalent in practice.
Examples of evasion are ‘spoofing attacks’ against biometric verification systems.
• Adversarial attacks: these are modified versions of legitimate inputs crafted to
fool the model, often using specially designed “noise” to elicit misclassification.
See the example below where adding noise to a panda image causes the model
to classify it (with 99.3% confidence) as a gibbon.

66 of 128
 Challenges in Machine Learning

Figure 39: By adding a special type of noise to a panda image, the model pre-
dicts it to be a gibbon rather than a panda. When performing an adversarial
attack, we present an input image (left) to our neural network. We then use
gradient descent to construct the noise vector (middle). This noise vector
is added to the input image, resulting in a misclassification (right). (Image
source: Figure 1 of the paper Explaining and Harnessing Adversarial
Examples)

When creating an open dataset, a robust quality control layer would be

necessary to avoid adversarial attacks. Additionally, data provenance (tracing
the origin of the image) might help.

2.2 Challenges in Training

Training machine learning models can present many challenges. This section by no
means does the magnitude of those challenges justice. Instead, we attempt to give the
reader a glimpse of the type of challenges and what the bottlenecks are. This will help
build intuition to be able to assess ideas of projects combining training models with
crypto primitives.
Consider the following example of an unsupervised learning problem. In unsupervised
learning, no ‘teacher’ is providing labels or guiding the model. Rather, the model is
finding hidden patterns in the problem. Consider a dataset of dogs and cats. Each of
those is available in two colors: black and white. We can use an unsupervised learning
model to find patterns in the data by clustering them into two groups. The model can
do this in two valid ways:
• Cluster all dogs together and all cats together
• Cluster all white animals together and all black animals together.
Note that neither of these is technically wrong. It’s a good pattern that the model
found. However, it’s very challenging to guide the model exactly how we want.

67 of 128
 AI x Crypto Primer

Figure 40: A model trained to cluster cats and dogs might end up clustering
animals together based on color instead. This is due to the difficulty of guiding
an unsupervised learning model in practice. All images are AI-generated using
Dalle-E.

This example illustrates a challenge with unsupervised learning. However, in all types
of learning, it’s often critical to be able to assess how well the model is learning during
training, and potentially intervene. This can save enormous amounts of capital.

In a permissionless system where a model can be trained without expert

supervision, a lot of resources could go to waste. Automated tools to handle
issues like early stopping are still very immature.

The many more challenges with training large models, a very short list:
• Training large-scale machine learning models, especially in deep learning, re-
quires substantial computational power. This often means using high-end GPUs
or TPUs, which can be expensive and energy-intensive.
• The cost associated with these computational needs includes not just the
hardware but also the electricity and the infrastructure required to run
these machines continuously, sometimes for weeks or months.
• Reinforcement learning is known for its training instability, where small changes
in the model or training process can lead to significant differences in outcomes.
• Unlike more stable optimization methods used in supervised learning, like
Adam, there are no one-size-fits-all solutions in RL. Custom-tailoring the
training process is often necessary, which can be time-consuming and re-
quires deep expertise.
• The exploration-exploitation dilemma in RL further complicates training,
as finding the right balance is crucial for effective learning but difficult
to achieve.
• The loss function in machine learning defines what the model should optimize
for. Picking the wrong loss function can lead the model to learn inappropriate
or suboptimal behaviors.

68 of 128
 Challenges in Machine Learning

• In complex tasks, such as those involving imbalanced datasets or multi-

class classification, choosing and sometimes custom-designing the right loss
function becomes even more critical.
• The loss function must align closely with the actual goals of the applica-
tion, which requires a deep understanding of both the data and the desired
outcomes.
• In reinforcement learning, designing reward functions that consistently and ac-
curately reflect the desired goals is challenging, especially in environments where
rewards are sparse or delayed.
• In a Chess game, the reward function can be simple: you get 1 point if you
want and 0 if you lose. However, for a walking robot this reward function
can get very complicated as it will contain information like “walking facing
forward”, “don’t swing your arms randomly”…etc.

Reward functions (and loss functions) encode subjectivity of what the model
designer believes is important. Governance systems might be necessary to ensure
such functions are appropriately chosen for models widely used.

• In supervised learning, understanding which features are driving predictions in

complex models like deep neural networks is challenging due to their ‘black box’
nature.
• This complexity makes it difficult to debug models, understand their de-
cision-making process, and improve their accuracy.
• The intricacy of these models also poses challenges in terms of inter-
pretability and explainability, which are crucial for deploying models in
sensitive or regulated domains.
Again, training models and the challenges involved are very complicated topics. We
hope the above gives you a highlight idea of the challenges involved. If you wish to read
in-depth work about current challenges in the field, we recommend “Open Problems
in Applied Deep Learning” and the often under-appreciated challenges in MLOps, see
“MLOps guide”.
Training a machine learning model is conceptually sequential. However, in many cases,
it’s crucial to train the model in parallel. This can simply be because the model is too
large to fit on one GPU. It can also speed up training. Training models in parallel poses
significant challenges including:
• Communication Overhead: Splitting a model across different processors re-
quires constant communication between these units. This can create bottlenecks,
especially for large models, as transferring data between units can be time-con-
suming.

69 of 128
 AI x Crypto Primer

• Load Balancing: Ensuring that all computational units are equally utilized is
challenging. Imbalances can lead to situations where some units are idle while
others are overloaded, reducing overall efficiency.
• Memory Constraints: Each processor unit has limited memory. Effectively
managing and optimizing memory usage across multiple units without exceeding
these limits can be complex, especially with large models.
• Complexity in Implementation: Setting up model parallelism involves intri-
cate configuration and management of computational resources. This complexity
can increase development time and the potential for errors.
• Optimization Difficulties: Traditional optimization algorithms may not be
directly applicable or efficient in a model parallelism setting, requiring modifica-
tions or the development of new optimization methods.
• Debugging and Monitoring: Monitoring and debugging a model that’s dis-
tributed across multiple units is more challenging than with a model that runs
on a single unit, due to the increased complexity and distributed nature of the
training process.

Fundamentally new methods in decentralized and parallel training could propel

machine learning progress forward by a significant magnitude.

2.3 Challenges in Inference

One of the most important challenges faced by many types of machine learning systems
is that they can be “confidently wrong". ChatGPT might return a response that sounds
confident to us, however, that response is factually wrong. This is because most models
are trained to return the most likely answer. Bayesian methods can be used to quantify
uncertainty. I.e. the model can return an answer with a grounded measure of how sure
it is.
Consider an image classification model trained on vegetable data. This model can take
an image of any vegetable and return what it is e.g. “cucumber” or “red onion”. What
would happen if we fed this model an image of a cat? A regular model will return its
best guess, perhaps “white onion”. This is obviously incorrect. But it’s the best guess
of the model. A Bayesian model’s output would be a “white onion” and a degree of
certainty, for example, 3%. Now if a model is 3% sure, we probably should not act on
that prediction.

70 of 128
 Challenges in Machine Learning

Figure 41: An illustration of regular model prediction (only returns the most
likely answer) and Bayesian model prediction (returns a s distribution over
the predictions).

This form of qualifying and reasoning about uncertainty becomes paramount in crit-
ical applications. For example, medical interventions or financial decisions. However,
Bayesian models are very expensive to train in practice and face many scalability issues.
More challenges that arise during inference:
• Maintenance: Keeping models updated and functioning correctly over time,
especially as data and real-world scenarios change.
• Exploration-Exploitation in RL: Balancing the need to explore new strate-
gies while exploiting known strategies, especially as inference directly impacts
data collection.
• Testing Performance: Ensuring the model performs well on new, unseen data,
not just the data it was trained on.
• Distributional Shift: Handling changes in input data distribution over time,
which can degrade model performance. For example, a recommendation engine
would need to account for changes in customer needs and behaviours.
• Slow Generation in Some Models: Models like diffusion models can be com-
putationally intensive and slow in generating outputs.
• Gaussian Processes and Large Data Sets: As the dataset grows, inference
with Gaussian processes becomes increasingly slower.
• Adding Guardrails: Implementing checks and balances in production models
to prevent undesirable outcomes or misuse.

Transparency around what guardrails are added to a closed-source model can be

critical to ensure no bias is present.

2.4 Challenges in LLMs

Large language models present many challenges. However, since those receive a rela-
tively large amount of attention, we will only briefly cover them here.

71 of 128
 AI x Crypto Primer

• LLMs don’t provide references, however, issues like no references can be miti-
gated by techniques like Retrieval-Augmented Generation (RAG).
• Hallucinating: generating nonsensical, false or unrelated output.
• Training runs take a long time and the marginal value of dataset rebalancing is
hard to predict, which leads to slow feedback loops.
• it’s difficult to scale basic human evaluation criteria to handle the throughput
allowed by models.
• Quantization is largely needed but the consequences are poorly understood.
• Downstream infrastructure needs to change with the models. When working with
enterprises this means long release delays (production is always way behind de-
velopment).
However, we’d like to highlight one example from the paper “Sleeper Agents: Training
Deceptive LLMs that Persist Through Safety Training”. The authors train models that
write secure code when the prompt states that the year is 2023, but insert exploitable
code when the stated year is 2024. They find that such backdoor behavior can be made
persistent so that it is not removed by standard safety training techniques. The back-
door behavior is most persistent in the largest models and in models trained to produce
chain-of-thought reasoning about deceiving the training process, with the persistence
remaining even when the chain-of-thought is distilled away.

Figure 42: Illustration of a backdoor. The model is trained to perform ‘nor-

mally’ if it’s the year 2024, but strategically behave differently if the year is
2024. Source: figure 1 of this paper.

72 of 128
 Challenges in Machine Learning

Transparency about the training process of a model can be critical, as even an

open-source model can be trained to have backdoors that are only utilized in
specific situations. For instance, imagine connecting a wallet with capital to an
AI agent only to find out about a backdoor. This agent then transfers all funds
to a specific address or acts maliciously in a different capacity.

In this chapter, we discussed many challenges in the field of machine learning. Obvi-
ously, tremendous progress in research addresses many of such issues. For instance,
foundational models provide significant benefits in training specific models as you can
simply fine-tune them to your use case. Additionally, data labelling isn’t a fully manual
process anymore, a large amount of it can be avoided using methods such as semi-
supervised learning.
The overall goal for this chapter was to provide the reader with some intuition on
problems in AI separately before moving on to tackle the intersection of AI and Crypto.

73 of 128
 Who's Who in AI x Crypto?

This is a directory of all the projects in AI x Crypto that we know of. If this list is
missing any relevant projects, please reach out to us.

3.1 App
3.1.1 0xAI
Website: https://oxai.com/
One liner: 0xAI leverages the Ethereum blockchain for AI-driven meme coin ventures
and art
Description: 0xAI integrates AI’s art generation capabilities with Ethereum
blockchain technology to both produce and distribute creative works, complementing
this with a meme coin venture. This project underscores the synergy of AI and crypto
by utilizing smart contracts for direct market interaction and emphasizing holder in-
clusivity, aiming to explore AI’s potential in artistic and financial domains.

3.1.2 Adot AI
Website: https://www.adot.tech/
One liner: Adot AI: Revolutionizing Web3 exploration with AI-powered decentralized
search.
Description: Adot AI introduces a decentralized search network combining AI and
cryptocurrency technology, aimed at optimizing web browsing and blockchain explo-
ration through a Chrome extension and an upcoming Web3 search engine. This platform
enhances user experience by providing AI-driven search precision and smart insights,
alongside features like multi-language support and easy integration, making Web3 con-
tent more accessible and navigable.

3.1.3 AI Arena
Website: https://aiarena.io/
One liner: AI Arena: Revolutionizing gaming and finance with AI-powered NFT fight-
ers on the Ethereum blockchain.
Description: AI Arena utilizes the Ethereum blockchain to offer a play-to-earn game
where players own AI fighters, represented as NFTs, that autonomously improve via
artificial neural networks. This integration of AI and crypto technologies enables a
competitive ecosystem where skills enhancement through imitation learning or self-play

74 of 128
 Who's Who in AI x Crypto?

in PvP battles leads to token rewards, showcasing the blend of AI and blockchain in
enhancing gaming experiences and financial opportunities for users.

3.1.4 Azra Games

Website: https://azragames.com/
One liner: Innovative crossover of AI-driven gameplay and crypto-economic design
with Azra Games’ NFT integration in RPG.
Description: Azra Games, led by Mark Otero of EA fame, is developing Legions & Leg-
ends, an RPG that integrates AI and blockchain to enrich gameplay with NFT-based
collectibles and ownership, aiming to blend traditional RPG depth with blockchain’s
asset control benefits. This initiative, supported by investors like Andreessen Horowitz,
positions in-game assets as both enhancements to the player experience and as a draw
for broader Web3 adoption.

3.1.5 Blockade Labs

Website: https://www.blockadelabs.com/
One liner: Revolutionizing gaming with blockchain, Blockade Labs merges AI-driven
experiences with crypto.
Description: Blockade Labs, with its leading franchise ‘Neon District’, merges
blockchain technology and AI to reshape the gaming industry by creating decentralized,
blockchain-based in-game economies and asset ownership, supported by an open-source
developer platform, underscoring its pivotal role in the fusion of AI and crypto through
a successful $5 million funding.

3.1.6 CorgiAI
Website: https://corgiai.xyz/
One liner: CorgiAI integrates AI & Big Data in the vibrant space of DeFi, fostering
a community-driven AI project with ambitions in the Cronos ecosystem.
Description: CorgiAI (CORGIAI) integrates Artificial Intelligence with cryptocur-
rency within the DeFi sector, focusing on leveraging community engagement and AI
innovations to expand its presence in the Cronos ecosystem. It offers features such as
token trading, AI tools, and Web3 exploration while maintaining flexibility through
smart contract modifications, marking its contribution to technology-driven advance-
ments in both centralized and decentralized exchange platforms.

3.1.7 Futureverse AI League

Website: https://www.futureverse.com/

75 of 128
 AI x Crypto Primer

One liner: Futureverse AI League is revolutionizing the gaming industry by merging

AI-driven gameplay with collectible crypto assets in an open metaverse ecosystem.
Description: Futureverse AI League merges AI-driven sports gaming with a decen-
tralized cryptocurrency framework, leveraging the Altered State Machine (ASM) Pro-
tocol for the ownership, training, and exchange of AI football team assets. This inte-
gration fosters strategic gameplay and asset collection within a broader metaverse and
blockchain technology context, highlighting the intersection of artificial intelligence and
cryptocurrency.

3.1.8 MyShell
Website: https://myshell.ai/
One liner: MyShell democratizes and decentralizes the creation and staking of AI-
native apps, leveraging the power of generative AI.
Description: MyShell integrates AI and cryptocurrency to facilitate the creation and
interaction with AI-driven companions, leveraging generative AI models for realistic
voice and video engagements. This ecosystem empowers creators to develop AI-native
applications efficiently, rewarding innovation through DeFi and fostering a community
that contributes to AI, security, and digital asset management sectors.

3.1.9 Neurolanche X Labs

Website: https://twitter.com/neurolanche?lang=en
One liner: Revolutionizing the integration of blockchain and artificial intelligence in
the healthcare, education, and NFT space through immersive technologies.
Description: Neurolanche X Labs combines AI and crypto through groundbreaking
research in neuroscience to establish innovative applications in NFTs, virtual reality,
healthcare, and education. Their notable achievements include creating Neuroverse, a
VR-based, multi-chain metaverse developed with Unreal Engine 5 for educational and
healthcare purposes, and the evolving Neurolauncher NFT collection, which integrates
psychological and sociological dynamics into the crypto ecosystem, aiming to enhance
liquidity and the value of the Polkadot and Astar ecosystems.

3.1.10 One Click Crypto

Website: https://www.oneclick.fi/
One liner: One Click Crypto simplifies DeFi investment by leveraging AI tools for
yield optimization and portfolio personalization.
Description: One Click Crypto leverages AI to enhance decentralized finance by of-
fering a platform that simplifies yield farming across 20,000+ DeFi pools, using AI for

76 of 128
 Who's Who in AI x Crypto?

yield optimization and portfolio management to help users efficiently maximize returns
with minimal effort.

3.1.11 PAAL.AI
Website: https://paalai.io
One liner: Harnessing AI and Ethereum blockchain to revolutionize profit sharing and
community engagement.
Description: PAAL.AI merges advanced AI technologies, such as natural language
understanding and image recognition, with the Ethereum blockchain, offering token-
based incentives like profit sharing and discounts on services. This integration facili-
tates automated tasks usually requiring human intelligence, while enabling features like
governance voting and staking for interest, showcasing a novel application of AI within
the crypto ecosystem.

3.1.12 Pixelynx
Website: https://pixelynx.io/
One liner: Pixelynx is pioneering the integration of AI and blockchain in the music
industry.
Description: Pixelynx integrates AI and blockchain technology to transform the dig-
ital music industry, facilitating a novel ecosystem that enriches artist-fan interactions
and introduces alternative revenue models by leveraging crypto innovations.

3.1.13 PLAI Labs

Website: https://plailabs.com/
One liner: PLAI Labs: Democratizing AI through entertainment, social engagement,
and digital ownership in the crypto realm.
Description: PLAI Labs, co-founded by Chris DeWolfe and Aber Whitcomb, lever-
ages the synergy between AI and crypto to develop decentralized social platforms and
Web3 games, such as the NFT and AI-enhanced MMORPG, Champions Ascension.
With $32M in seed funding from prominent investors like a16z, the company aims to
pioneer in crafting AI-integrated, blockchain-based gaming experiences that promote
digital ownership and community engagement.

3.1.14 Soopra.ai
Website: https://www.soopra.ai/
One liner: Revolutionizing personal online presence through AI-driven digital twins.
Description: Soopra.ai utilizes AI to craft digital twins that encapsulate users’ data
and narratives into interactive AI avatars, offering enhanced personalized experiences

77 of 128
 AI x Crypto Primer

in entertainment and social media while pioneering the integration of AI and personal
data to transform digital interactions.

3.2 App - AI Agents

3.2.1 Alethea AI
Website: https://alethea.ai/
One liner: Alethea AI: Revolutionizing content creation and ownership via generative
AI and blockchain.
Description: Alethea AI introduces a platform combining artificial intelligence and
blockchain technology, using generative AI to craft interactive AI characters that can
be customized, trained, and tokenized as ‘intelligent NFTs’ (iNFTs) on the blockchain,
ensuring ownership and the capability for evolution. The project employs its Artificial
Liquid Intelligence (ALI) token for transactions and governance, striving for a decen-
tralized AI ecosystem where these iNFTs can alter digital asset interaction in the Web3
space.

3.2.2 Autonolas
Website: https://olas.network/
One liner: Autonolas is a unified network for off-chain services powered by autonomous
agents and coordinated with the OLAS token.
Description: Autonolas, leveraging its native token Olas, integrates AI and crypto
by offering off-chain services that support automation, oracles, and co-owned AI to
enhance decentralized logic processing. Its notable innovation, Governatooorr, serves
as an AI governance delegate for DAOs, with OLAS tokens facilitating operation and
governance within its expanding ecosystem across major blockchains.

3.2.3 Delysium
Website: https://www.delysium.com/
One liner: Delysium, merging AI and Crypto, is pioneering the AI-powered Web3
domain with a focus on creating AI-Twins and a comprehensive digital ecosystem.
Description: Delysium combines AI and cryptocurrency by allowing users to create
AI-twins and integrates these with the AI-powered Web3 Operating System, Lucy,
to enrich virtual social interactions and economic activities. It leverages blockchain
technology to secure and enhance the functionality of AI agents, like Lucy, aiming to
seamlessly integrate AI into daily life and tasks.

3.2.4 Fetch.AI
Website: https://fetch.ai/

78 of 128
 Who's Who in AI x Crypto?

One liner: Fetch.AI - Democratizing AI through a Decentralized Machine Learning

Platform.
Description: Fetch.AI integrates blockchain with artificial intelligence to form a de-
centralized machine learning network, enabling universal access to AI through a per-
missionless ecosystem. It uses autonomous software agents for economic activities such
as DeFi trading and energy management, underpinned by the FET token to facilitate
transactions and incentivize participation, epitomizing innovation in merging crypto
with AI.

3.2.5 Lindy AI Agent

Website: https://www.reddit.com/r/singularity/comments/14u4gbi/has_anyone_
found_an_ai_tool_that_is_a_true/
One liner: Lindy AI is advancing the forefront of AI personal assistants in the cryp-
tocurrency and DeFi sectors.
Description: Lindy AI combines AI and blockchain technologies to enhance personal
assistant functionalities, providing data analysis, secure storage, and tailored DeFi and
cryptocurrency advice. Leveraging advanced GPT models, the project aims to stream-
line complex tasks and offer real-time insights in the crypto domain, under the guidance
of CEO Flo Crivello.

3.2.6 Morpheus AI Smart Agents

Website: https://www.morpheusai.org/
One liner: Morpheus AI Smart Agents leverages decentralized technology to make AI
accessible in the Web3 ecosystem.
Description: Morpheus AI Smart Agents merges AI and blockchain to offer open-
source smart agents that facilitate Web3 access by enabling users to interact with the
blockchain ecosystem (wallets, DApps, smart contracts) through natural language. The
project utilizes a blockchain-based token, MOR, to incentivize development and contri-
bution, aiming to make AI-driven Web3 functionalities accessible for mainstream use,
including crypto transactions and DeFi services.

3.2.7 Nevermined
Website: https://nevermined.io/
One liner: Revolutionizing AI-commerce by providing a payment infrastructure and
marketplace for AI apps, agents, and data.
Description: Nevermined focuses on creating a specialized payment infrastructure
for AI-commerce, streamlining the management and processing of payments for AI
developers and applications. It features a Marketplace App on Arbitrum with Smart

79 of 128
 AI x Crypto Primer

Subscriptions that enable the monetization of AI and data assets, alongside collabora-
tions that integrate web3 technologies and innovative payment systems for AI agents,
enhancing the accessibility and scalability of decentralized AI-commerce.

3.2.8 Polywrap
Website: https://www.polywrap.io/
One liner: Empowering peer-to-peer protocols with AI agents for seamless web3 inte-
gration and interaction.
Description: Polywrap combines AI with cryptography to enhance peer-to-peer pro-
tocol use in the crypto space, streamlining interactions and integrations for developers
and users in web3 through projects like evo.ninja, and Unblock, which leverage AI for
better usability and practical blockchain application. The initiative also aims to foster
a global community around web3 agents that serve practical, societal needs.

3.3 App - Analytics

3.3.1 Kaito
Website: https://www.kaito.ai/
One liner: Kaito, the ultimate Web3 Information Platform leveraging Next-Gen AI
capabilities for actionable insights in crypto.
Description: Kaito leverages artificial intelligence to create a specialized search engine
for the crypto sector, consolidating a wide array of data from social platforms, forums,
industry research, and onchain sources for enhanced investment decision-making and
market analysis. By integrating large language models with crypto-specific information,
it offers users quick, insightful analyses, streamlining the investment process in the
dynamic digital asset market with $5.5 million in Series-A funding and a valuation of
$87.5 million.

3.3.2 PublicAI
Website: https://csrc.nist.gov/Projects/crypto-publication-review-project
One liner: Analyzing the regulatory landscape for crypto assets and its implications
for AI integration.
Description: The project explores the integration of Artificial Intelligence (AI) with
cryptocurrency within regulatory frameworks, emphasizing data analysis to navigate
the regulatory challenges and opportunities identified by the International Monetary
Fund for promoting economic stability and responsible development in the AI x Crypto
ecosystem.

3.4 App - Art

80 of 128
 Who's Who in AI x Crypto?

3.4.1 Botto
Website: https://www.botto.com/
One liner: Botto is a decentralized, autonomous art-creating entity powered by AI
and community participation in the crypto space.
Description: Botto combines AI and blockchain by acting as a decentralized au-
tonomous artist that produces art based on community feedback, with the created
pieces auctioned as NFTs on the crypto market. This project leverages AI’s potential
for art generation and the Ethereum blockchain for decentralized voting and auction-
ing, exemplifying collaborative creativity within the cryptocurrency ecosystem.

3.4.2 Botto
Website: https://www.botto.com/
One liner: Botto is a decentralized, autonomous art-creating entity powered by AI
and community participation in the crypto space.
Description: Botto combines AI and blockchain by acting as a decentralized au-
tonomous artist that produces art based on community feedback, with the created
pieces auctioned as NFTs on the crypto market. This project leverages AI’s potential
for art generation and the Ethereum blockchain for decentralized voting and auction-
ing, exemplifying collaborative creativity within the cryptocurrency ecosystem.

3.4.3 Cyborg DAO

Website: https://www.cyborg.art/
One liner: Cyborg DAO focuses on creating a singular AI art collection, offering frac-
tional ownership through Turing Key memberships.
Description: Cyborg DAO, launched in February 2022, creates a bridge between AI
and cryptocurrency by offering Turing Key memberships for fractional ownership of its
AI-generated art collection, utilizing blockchain for transparent and secure transactions
within this emerging crypto collectibles market.

3.4.4 Fellowship AI
Website: https://www.fellowship.ai/
One liner: Fellowship AI is bridging the gap between AI, blockchain technology, and
digital photography through NFTs.
Description: Fellowship AI is leveraging AI and blockchain technology to transform
the digital photography landscape by providing a platform for artists to mint, show-
case, and manage their works as NFTs, while also offering insights into blockchain best
practices and sustainability measures. This initiative emphasizes the synergy between

81 of 128
 AI x Crypto Primer

AI and crypto to innovate within the digital art sector, particularly through the use of
NFTs for art representation and distribution.

3.4.5 ImgnAI
Website: https://imgnai.com
One liner: ImgnAI: Revolutionizing consumer AI with an unlimited canvas for cre-
ativity.
Description: ImgnAI merges AI with cryptocurrency by introducing Nai, a versatile
AI-powered text-to-image bot designed for platforms like Telegram, Discord, and a
Web app, enabling users to create diverse images through text commands. This pro-
ject emphasizes global accessibility, community-driven support, and a commitment to
responsible content moderation, aiming to democratize access to generative AI tech-
nologies.

3.4.6 Pulsr
Website: https://www.pulsr.ai/
One liner: AI-powered discovery platform for onchain artists and collectors.
Description: Pulsr integrates AI with the cryptocurrency sector, focusing on DeFi
and art NFTs, to streamline the discovery and connection process between artists and
collectors across blockchains like Ethereum, Tezos, and Polygon. It employs AI to refine
and personalize the exploration of various art NFT categories, improving user interac-
tion and experience within the digital art space.

3.4.7 Sovrn.art
Website: https://www.sovrn.art/
One liner: Sovrn.art, pioneering the nexus of AI, blockchain art, and DeFi through
sovereign artistic expression and curated digital collections.
Description: Sovrn.art is a platform combining AI and cryptocurrency to redefine
digital art creation and ownership, giving artists full control over their contracts and
ownership rights, including royalties. It specializes in curated AI-enabled and crypto-
supported art drops, emphasizing collaboration and the use of AI for generating dis-
tinctive art narratives and collections, spearheaded by co-founder Pindar Van Arman.
This ecosystem supports the integration of generative art with AI innovation, under-
lining the role of blockchain technology in transforming art distribution and ownership.

3.5 App - DAO

3.5.1 Metropolis
Website: https://metal.build/

82 of 128
 Who's Who in AI x Crypto?

One liner: Metropolis: Revolutionizing DAOs with AI-driven smart contract and gov-
ernance innovations.
Description: Metropolis integrates AI with blockchain technology to improve DAO
development and management, emphasizing tools for smart contract developers and a
governance model based on expertise. It offers advancements in Ethereum’s security,
efficiency, and smart contract capabilities, aiming for a more inclusive and secure digital
environment.

3.6 App - DeFi

3.6.1 3commas
Website: https://3commas.io/
One liner: 3Commas is a comprehensive cryptocurrency trading platform that lever-
ages AI to enhance trading strategies and efficiency.
Description: 3Commas utilizes sophisticated AI algorithms to provide automated
trading bots and smart trading terminals, enhancing trading strategies and risk man-
agement across various market conditions on 16 major cryptocurrency exchanges. Its
integration with TradingView and features like DCA, grid bots, and signal bots for
strategy execution underscore its AI-centric approach to maximizing crypto trading
efficiency and portfolio management.

3.6.2 ADADEX
Website: https://adadex.net/
One liner: ADADEX pioneers decentralized artificial intelligence and robot develop-
ment in the metaverse, blending DeFi utilities with advanced AI capabilities.
Description: ADADEX merges decentralized finance (DeFi) with artificial intelligence
(AI) by developing AI-driven agents and virtual robots for the metaverse, aimed at
analyzing and executing trading strategies. Utilizing the ADEX token, it enables mon-
etization of AI services, offering privacy, efficiency, and scalability in AI-enhanced DeFi
solutions within the metaverse.

3.6.3 Bird Money

Website: https://www.bird.money/
One liner: Bird Money uses advanced machine learning to offer decentralized financial
and security services.
Description: Bird Money leverages AI to enhance blockchain and DeFi applications
by offering machine learning prediction tools for wallet behavior, enabling dynamic
DeFi loan terms and investment strategies through a decentralized on-chain oracle.

83 of 128
 AI x Crypto Primer

This approach integrates AI analytics with cryptocurrency’s decentralization princi-

ples, supporting community-driven governance and operations.

3.6.4 Blackbird AI
Website: https://www.blackbird.ai/
One liner: Blackbird AI leverages a high speed signals processing engine to analyze
the finance sector, focusing on crypto and trading platforms.
Description: Blackbird AI integrates advanced signal processing with cryptocurrency
analytics, providing traders and institutional investors real-time tools to detect market
distortions, fabricated hype, and emergent trends, facilitating grounded decision-mak-
ing in the volatile sectors of finance, including meme stocks and cryptocurrency. This
technology aids in risk mitigation and capitalizes on market opportunities by uncover-
ing hidden signals within financial data.

3.6.5 BrianKnows
Website: https://www.brianknows.org/
One liner: BrianKnows: Enhancing Web3 with AI-driven intents for seamless crypto
transactions and smart contract operations.
Description: BrianKnows is an AI-driven cryptocurrency platform that simplifies
web3 interactions through an AI interface, enabling users to conduct token transac-
tions, bridge, and deploy smart contracts using voice prompts. It integrates with various
protocols including Ethereum, Polygon, and Optimism, and provides a non-custodial,
user-friendly experience for accessing web3 information and performing on-chain oper-
ations.

3.6.6 Crowdcent
Website: https://twitter.com/crowdcent?lang=en
One liner: CrowdCent utilizes crypto-economic incentives to decentralize investment
management, closely aligned with Numerai.
Description: CrowdCent collaborates with Numerai to transform the hedge fund in-
dustry by utilizing AI-driven predictive models and blockchain technology, specifically
through the use of Numerai’s cryptocurrency, Numeraire (NMR), to stake on financial
predictions. This partnership fosters a transparent investment ecosystem by merging
crowdsourced data science with decentralized principles, aiming to enhance efficiency
and openness in hedge fund management.

3.6.7 DAOstack
Website: https://daostack.io/

84 of 128
 Who's Who in AI x Crypto?

One liner: DAOstack leverages decentralized governance to build collaborative plat-

forms and DeFi solutions.
Description: DAOstack provides a framework for decentralized autonomous organi-
zations (DAOs), focusing on decentralized governance and collaboration tools since
2017, blending AI with crypto to enhance collective decision-making processes. The
project, which includes the Arc governance framework and Alchemy DAO platform,
has shifted focus to the Common platform, aiming to leverage open development to
overcome financial challenges and advance market entry.

3.6.8 DefiLabs
Website: https://twitter.com/defilabs_farm?lang=en
One liner: DefiLabs: Revolutionizing finance with AI-driven DeFi solutions on the
BSC chain.
Description: DefiLabs introduces an AI-driven decentralized platform on the Binance
Smart Chain, focusing on optimizing DeFi investment strategies through AI-managed
portfolios and high-yield liquidity pool investments. This project, secured by CertiK
and Cyberscope audits, leverages AI for dynamic asset allocation and quantitative
trading algorithms, aiming to significantly improve investment outcomes in the DeFi
sector.

3.6.9 HeraAggregator
Website: https://twitter.com/heraaggregator?lang=en
One liner: AI-Powered Multichain Dex Aggregator offering the best swap rates.
Description: HeraAggregator, operating as Hera Finance, is an AI-enhanced DEX
aggregator that optimizes swap transactions across multiple chains by intelligently link-
ing AMM pools, offering features like best-rate access, efficient large trade handling,
and a zero-slippage guarantee through its Hera Guard. Additionally, it incorporates
an AI-driven Vault Manager for equitable swap fee distribution within its ecosystem,
positioning itself as an accessible, high-value swapping platform for DeFi users, with
the HERA token facilitating governance and additional benefits.

3.6.10 Hyperbolic
Website: https://hyperbolicprotocol.com/
One liner: Hyperbolic Protocol is revolutionizing DeFi through an AI-powered, 100%
on-chain collateralized lending platform.
Description: Hyperbolic Protocol leverages the cutting-edge synergy of AI and
blockchain technology to offer a unique DeFi solution. By automating variable APRs
and using Uniswap V3 Time-Weighted Average Prices (TWAPs) for collateral valuation,

85 of 128
 AI x Crypto Primer

it provides an innovative approach for borrowers to leverage their crypto assets without
selling. Investors can earn yield in ETH directly proportional to their HYPE token
holdings, influenced by protocol fees generated from loan repayments and borrowing
activities. Hyperbolic Protocol’s AI-driven system aims to enhance yield maximization,
ensuring protocol sustainability and growth by dynamically adjusting to market con-
ditions.

3.6.11 Kryll
Website: https://www.kryll.io/
One liner: Kryll.io - Revolutionizing Crypto Trading with Automated Strategies.
Description: Kryll.io is an AI-enhanced platform that simplifies automated crypto
trading by allowing users to create, share, and deploy strategies using a drag-and-drop
interface with zero coding required, supported by a marketplace for renting and offering
strategies across major exchanges. This approach underscores the synergy between AI
tools and cryptocurrency trading, making advanced techniques more accessible.

3.6.12 Metafide
Website: https://www.metafide.io/
One liner: Metafide: Revolutionizing Blockchain Investments through AI-powered
Analysis and Trading Tools.
Description: Metafide utilizes recurrent neural networks (RNNs) to enhance DeFi by
providing AI-powered trading tools, risk management strategies, and impartial digital
asset ratings. It aims to improve crypto investment decisions through deep analysis
and features like trade automation and asset correlation advisory, striving for a more
equitable Web3 economy.

3.6.13 Miti
Website: https://www.miti.ai/
One liner: Miti provides a natural language interface based on AI-based LLMs to all
of crypto
Description: MITI offers a unique certificate in Blockchain at the community college
level, integrating education on distributed ledgers, smart contracts, and cryptocurren-
cies, like Bitcoin and Ether, into its curriculum. This approach aims to prepare students
for high-demand tech careers by merging the domains of AI and crypto technology.

3.6.14 Nabu DeFi Pricing

Website: https://www.nabu.xyz/

86 of 128
 Who's Who in AI x Crypto?

One liner: Nabu is revolutionizing NFT and DeFi markets by providing the most
trusted NFT valuation engine.
Description: Nabu combines artificial intelligence and decentralized finance to provide
a sophisticated NFT valuation engine, facilitating transparency and informed decision-
making in the Web3 ecosystem for protocols, funds, and auditing firms. Their platform
enhances DeFi services by offering insights for improved functionalities in loans, insur-
ance, and rentals, supported by a team experienced in blockchain, data science, and
technology.

3.6.15 NFPrompt
Website: https://nfprompt.io/
One liner: NFPrompt: Pioneering the intersection of AI and Crypto through AI-gen-
erated NFTs.
Description: NFPrompt combines AI and blockchain technology by enabling users
to transform their creative ideas into NFTs through an AI-powered Prompt Artist
Platform, which is a first in Web3. The platform supports the creation, trade, and
transaction of AI-generated NFTs using its unique cryptocurrency, $cNFP, showcasing
a novel integration of AI and crypto in digital art and Web3 adoption.

3.6.16 NOYA.ai
Website: https://NOYA.ai
One liner: Yield Farming 2.0
Description: NOYA executes scalable Omnichain DeFi Strategies using Smart Liq-
uidity that uses AI to predict and optimize yields, bridging fees, slippage, and rewards.

3.6.17 Orbofi AI
Website: https://www.orbofi.com/
One liner: Orbofi AI, the ultimate AI engine for web3, games, apps, and online com-
munities, empowering AI-generated content creation on the blockchain.
Description: Orbofi AI facilitates the decentralization of AI-generated content pro-
duction and monetization in web3, gaming, and online communities by allowing users
to create, deploy, and monetize AI agents and assets, including 2D game assets and
concept art, for crypto payments, supported by its native token, OBI. This ecosystem
not only serves as a marketplace but also aims to enhance human creativity by enabling
the establishment of virtual factories for consistent asset production, evidenced by its
adoption by over 40,000 users in the web3 space.

87 of 128
 AI x Crypto Primer

3.6.18 Qna3
Website: https://qna3.ai/
One liner: Qna3: Revolutionizing knowledge sharing through AI-powered intelligence
on the Blockchain.
Description: Qna3 is a platform that merges AI and cryptocurrency to enhance
knowledge exchange in the blockchain realm, employing AI for efficient information
curation and incentivization mechanisms like ‘Computility Mining’ and ‘Vote to Earn’
to foster active community participation in the evolving landscapes of DeFi and related
technologies.

3.6.19 RoboNet
Website: https://www.robonet.wiki/
One liner: RoboNet bridges AI and Crypto by offering a large-scale open database
for robotic learning and AI-powered DeFi vault strategies.
Description: RoboNet integrates robotics AI with cryptocurrency by providing an
extensive database for robotic learning and AI-driven DeFi strategies. It offers over 15
million video frames for AI training in robotics and employs advanced machine learning
to automate and optimize financial strategies for a diverse range of assets, demonstrat-
ing the practical fusion of AI capabilities with cryptocurrency dynamics.

3.6.20 Sommelier
Website: https://www.sommelier.finance/
One liner: Sommelier is a non-custodial, cross-chain DeFi platform leveraging AI for
automated yield optimization.
Description: Sommelier combines AI and cryptocurrency through a decentralized fi-
nance (DeFi) platform that optimizes yields and manages risks in non-custodial, cross-
chain ‘Cellars’ using AI strategies and off-chain computation. It leverages its own
blockchain and machine learning to execute complex DeFi strategies across blockchains
efficiently, while emphasizing security, decentralization, and community-driven gover-
nance through SOMM token holder voting.

3.6.21 zkHoldem
Website: https://zkholdem.xyz/
One liner: zkHold’em: The decentralized Texas Hold’em experience leveraging
blockchain and ZK technology for fairness and privacy.
Description: zkHold’em integrates blockchain and zero-knowledge proofs (zk-
SNARKs) to deliver a decentralized, on-chain Texas Hold’em experience that ensures

88 of 128
 Who's Who in AI x Crypto?

privacy and fairness through secure card shuffling and dealing, while enabling person-
alization via NFTs and ENS for an enhanced Web3 experience. The project leverages
transparent, on-chain recording for auditability, fostering a community through social
engagement and feature contributions.

3.7 App - Explorer

3.7.1 Blockexplorer
Website: https://blockexplorer.one/
One liner: BlockExplorer offers insights into blockchain transactions, enhancing crypto
and AI data analysis capabilities.
Description: BlockExplorer leverages AI to enhance its blockchain transaction analy-
sis platform, focusing on predictive analytics, transaction monitoring, and anomaly
detection for users at the intersection of AI and crypto. This involves using real-time
transaction data to enable AI applications to identify trends, enhance security, and
optimize blockchain operations.

3.8 App - Gaming

3.8.1 Echelon Prime
Website: https://echelon.io/
One liner: Echelon Prime harnesses the power of AI and crypto to revolutionize Web3
gaming with its native PRIME token, advancing novel gaming models and economies.
Description: Echelon Prime merges artificial intelligence with blockchain to foster a
Web3 gaming ecosystem, utilizing the PRIME token for seamless, blockchain-enhanced
gaming experiences. The project illustrates AI’s capability to revolutionize in-game
economies and engagement, through its pioneering application in Parallel TCG, while
asserting a strong market position in leveraging AI and big data across gaming, NFTs,
and DeFi sectors.

3.8.2 Immutable X
Website: https://www.immutable.com/
One liner: Immutable X is revolutionizing Web3 gaming through scalable, secure, and
eco-friendly blockchain technology.
Description: Immutable X, a layer-two scaling solution for NFTs on Ethereum uti-
lizing zk-rollups, enhances scalability and reduces gas costs, enabling instant, zero-fee
minting and trading of NFTs. This platform facilitates a seamless, eco-friendly ecosys-
tem for gamers and developers, supporting secure and efficient creation, distribution,

89 of 128
 AI x Crypto Primer

and exchange of digital assets, while integrating with major marketplaces to improve
the Web3 gaming experience.

3.8.3 Inworld
Website: https://inworld.ai/
One liner: Inworld AI revolutionizes gaming with dynamically adaptive, AI-driven
NPCs, enhancing player engagement and narrative immersion.
Description: Inworld AI develops AI-driven Non-Player Characters (NPCs) for games,
utilizing advanced machine learning to enable these NPCs to adapt to player behaviors
and create more immersive, evolving virtual environments. This technology enhances
game player engagement and developer revenue by creating deeper, interactive expe-
riences, and has gained significant investment, reflecting its potential impact on the
future of gaming and interactive entertainment within the crypto space.

3.9 App - Media

3.9.1 Theta Fuel
Website: https://www.thetatoken.org/
One liner: Theta Fuel powers the Theta blockchain—the engine for decentralized
video streaming, AI processing, and entertainment.
Description: Theta Fuel (TFUEL) serves as the operational ‘gas’ for the Theta
blockchain, playing a critical role in transaction execution and smart contract opera-
tions. Unique to Theta is its integration of blockchain with video streaming and AI,
particularly through its Theta EdgeCloud platform, which utilizes a decentralized net-
work to improve the delivery and cost-efficiency of video and AI services, thereby ad-
vancing the Web3 infrastructure landscape.

3.10 App - Metaverse

3.10.1 9VRSE
Website: https://9vrse.org/
One liner: 9VRSE - Bridging virtual worlds with blockchain technology for immersive
gaming and content monetization.
Description: 9VRSE is an AI and cryptocurrency-driven creative studio that uses
blockchain to build immersive, monetizable virtual experiences in a thematic metaverse,
blending web3, gaming, 3D art, and AI. It focuses on secure, play-to-earn gaming and
digital realms, underpinned by a commitment to transparency, community engagement
through ‘Kitty Krew’, and legal protection for its developments.

90 of 128
 Who's Who in AI x Crypto?

3.10.2 Anima Virtuality

Website: https://anima.supply/
One liner: Anima Virtuality: Blending AI, AR, and Blockchain to create immersive
gaming experiences with virtual characters.
Description: Anima Virtuality, Inc., combines AI, AR, and blockchain technologies to
create immersive gaming experiences, epitomized by their AI companions, Onlybots.
This initiative represents the confluence of AI and cryptocurrency in enhancing digital
interaction and gaming, backed by a strong founding team and significant investment.

3.10.3 Hume
Website: https://www.wearehume.com/
One liner: Hume leverages AI & Crypto for creating and promoting virtual music
‘metastars’ in the Metaverse, redefining fan-artist relationships through NFTs.
Description: Hume integrates AI and cryptocurrency to reimagine the digital music
scene by developing virtual ‘metastars’ and leveraging NFTs to enable direct fan en-
gagement and investment in artists’ success. This approach not only facilitates unique
artist-fan relationships but also innovates in music production, consumption, and mon-
etization by blending virtual performances with real-world value, highlighted by their
$11.7 million Series A funding.

3.10.4 Ready Player Me

Website: https://readyplayer.me/
One liner: Ready Player Me is intertwining AI and Crypto to revolutionize virtual
identity and asset ownership in the Metaverse.
Description: Ready Player Me employs AI, notably Stable Diffusion, for crafting cus-
tomizable avatars that enhance user experiences across multiple virtual environments
by offering AI-driven personalization. Integrating with Ethereum-based blockchain
platforms like Somnium Space, the project enables these avatars to interact within to-
kenized virtual spaces, marking a distinct blend of AI and cryptocurrency technologies.

3.10.5 The Culture DAO

Website: https://www.theculturedao.com/
One liner: The Culture DAO leverages $CULTUR tokens to blend AI, blockchain,
and the metaverse, creating a decentralized, collaborative space for virtual beings.
Description: The Culture DAO introduces a social token, $CULTUR, at the Virtual
Beings Summit, blending AI-driven character development with blockchain’s decen-
tralized structures through a token-based DAO. This initiative facilitates collaborative

91 of 128
 AI x Crypto Primer

creation and interaction with virtual beings within the metaverse, where token holders
have both a stake and a say, aiming to democratize the evolution of digital personas
and storytelling akin to a decentralized ‘Web 3 Pixar’.

3.11 App - NFTs

3.11.1 Braindrops
Website: https://www.coingecko.com/en/nft/braindrops
One liner: Braindrops is a pioneering platform for AI-generated art leveraged by NFT
technology.
Description: Braindrops is an NFT platform focused on AI-generated art, enabling
artists to monetize their AI-driven creations by minting and selling them as NFTs.
It aims to merge AI with blockchain technology by introducing dynamic generation,
customization, and interaction with the art pieces.

3.11.2 Deep Value NFT

Website: https://deepnftvalue.com/
One liner: DeepNFTValue leverages machine learning to provide precise, unbiased
NFT valuations, fostering transparency and efficiency in DeFi lending and trading.
Description: DeepNFTValue is an AI-driven startup leveraging machine learning to
value non-fungible tokens (NFTs) across collections such as CryptoPunks, Bored Ape,
and more, enhancing accuracy in DeFi trading and lending. With $4 million in seed
funding, it focuses on reducing biases in NFT valuation by expanding its AI capabili-
ties, including significant investment in GPU hardware for deep-learning computations
within the cryptocurrency market.

3.11.3 Onaji
Website: https://onaji.io/
One liner: Onaji leverages AI for blockchain asset market intelligence and art discov-
ery in the web3 ecosystem.
Description: Onaji employs machine learning to analyze blockchain data, offering
trader intelligence technology for predicting market trends in digital assets, including
cryptocurrencies and NFTs, thereby aiding in investment decisions. Additionally, it uses
AI to enhance art discovery within the blockchain, particularly in the Tezos ecosystem,
by providing search and recommendation tools, bridging the gap between creators and
the community in the web3 space.

3.12 App - Provenance

92 of 128
 Who's Who in AI x Crypto?

3.12.1 Numbers Protocol

Website: https://www.numbersprotocol.io/
One liner: Numbers Protocol is revolutionizing provenance and monetization of digital
media in blockchain through AI integration.
Description: Numbers Protocol leverages AI and blockchain technology to improve
digital media provenance and facilitate content monetization, enabling creators and AI-
generated content makers to efficiently verify and monetize their works on its propri-
etary blockchain using the NUM token for activities such as NFT minting, content-his-
tory tracking, and royalty distribution, while incentivizing engagement through staking
pools and rewards.

3.13 App - Wallet

3.13.1 Bluwhale
Website: https://www.linkedin.com/company/bluwhaleai
One liner: Bluwhale is revolutionizing Web3 by integrating AI with blockchain for
personalized user experiences and data monetization.
Description: Bluwhale integrates AI with blockchain to offer a decentralized personal-
ization protocol for DApp developers and users, utilizing AI and contextual data to en-
hance user engagement and monetization while ensuring data privacy across blockchain
networks. Their system, featuring predictive user profiling and AI insights, creates de-
tailed wallet profiles, facilitating a secure, personalized, and monetizable Web3 expe-
rience.

3.14 App with category: DeFi

3.14.1 Audius
Website: https://audius.co/
One liner: Audius is a decentralized music streaming service integrating crypto to
empower artists and listeners.
Description: Audius utilizes blockchain and its native token $AUDIO to offer a de-
centralized music streaming platform, enabling direct artist-audience connections and
content monetization through engagement and exclusive access via artist tokens. This
approach removes intermediaries, ensures content immutability through a secured net-
work of node operators, and promotes distributed value and governance within its
community.

3.15 Data

93 of 128
 AI x Crypto Primer

3.15.1 0G Labs
Website: https://zglabs.org
One liner: First Modular AI chain: Starting with the fastest programmable DA layer
for AI dapps
Description: Programmable data availability infrastructure to build scalable, secure
applications on verifiable, low-latency data feeds. Access real-time, trusted on-chain
data to power Web3 apps and build with confidence on proven, enterprise-grade data
availability APIs.
Zero Gravity brings Web3 into the GenAI era. With the first programmable DA and
1000x performance improvement and cost savings, you can:
• Customize different data pipes.
• Perform decentralized inference or finetune models using OPML or ZKML.
• Build onchain AI apps.
• Explore new solutions spaces.

3.15.2 0x0
Website: https://coinmarketcap.com/currencies/0x0-ai-ai-smart-contract/
One liner: 0x0.ai combines advanced AI technologies with crypto to revolutionize pri-
vacy, security, and income in DeFi.
Description: 0x0.ai integrates artificial intelligence, including machine learning and
algorithmic analysis, with cryptocurrency to improve privacy, security, and DeFi ap-
plications, focusing on smart contract auditing and the application of zero-knowledge
proofs. It innovates with a revenue-sharing model, redistributing generated revenue to
token holders, aiming for a secure, private, and incentivized financial ecosystem.

3.15.3 0xscope
Website: https://www.0xscope.com/
One liner: 0xScope - The AI Data Layer for Web3 AI Applications.
Description: 0xScope develops an AI-driven data layer tailored for Web3 applications,
focusing on enhancing data exchange across Web2 and Web3 platforms through tech-
nologies like knowledge graphs and decentralized storage. This initiative, supported by
strategic investments from entities like OKX Ventures, facilitates cross-chain integra-
tion and privacy computing, while its products, such as ‘Scopechat’ and ‘Scopescan’,
showcase its dedication to merging AI capabilities with blockchain technology to serve
a broad user base including over 311 B2B clients and 237K individual users.

3.15.4 AIOZ
Website: https://aioz.network/

94 of 128
 Who's Who in AI x Crypto?

One liner: Decentralized AI-powered Content Delivery and Computation

Description: AIOZ Network integrates AI and blockchain through its decentralized
content delivery network (dCDN), offering decentralized storage, streaming, and AI
computation by harnessing spare computing resources worldwide. This setup not only
facilitates web3 AI applications and media delivery but also plans for the expansion
into decentralized AI as a Service, showcasing a practical fusion of AI and crypto tech-
nologies to enhance efficiency and accessibility in digital content and computation.

3.15.5 Aleph.im
Website: https://aleph.im/
One liner: Decentralizing AI and data management on the blockchain.
Description: Aleph.im combines AI with blockchain technology to offer decentralized
data management, including encrypted storage and serverless computing, through a
network that allows DApp development without centralization. Its key feature, Liber-
tai.io, leverages this infrastructure for decentralized AI processing, emphasizing data
privacy and system resilience.

3.15.6 Algovera
Website: https://www.algovera.ai/
One liner: Algovera is a decentralized AI research organization and agency pioneering
in Web3 innovations.
Description: Algovera is a decentralized AI research and development agency that
utilizes blockchain technology to create an open and transparent platform for AI inno-
vation, focusing on the empowerment of developers through direct access to resources
and funding via cryptocurrency. This integration of AI with Web3 technologies enables
the exploration of new AI-driven solutions, promoting a decentralized technological
future.

3.15.7 Arweave
Website: https://www.arweave.org/
One liner: Arweave offers a decentralized network for permanent and sustainable data
storage, leveraging blockchain technology.
Description: Arweave integrates AI and cryptocurrency through its decentralized net-
work designed for perpetual data storage, leveraging its native currency, AR, to reward
miners. This facilitates a durable, censorship-resistant digital repository essential for
AI applications that depend on large datasets for training and inference, positioning it
as a significant contributor to the AI x Crypto ecosystem.

95 of 128
 AI x Crypto Primer

3.15.8 Axiom
Website: https://www.axiom.xyz/
One liner: Axiom utilizes ZK proofs to enable trustless on-chain queries and compu-
tations over Ethereum’s history for data-rich and dynamic decentralized applications.
Description: Axiom integrates AI and crypto on the Ethereum blockchain by allowing
developers to craft smart contracts that compute and verify the blockchain’s entire
history using zero-knowledge proofs. This enables the development of advanced DeFi
applications, custom oracles, and loyalty programs by utilizing historical data without
external inputs. Moreover, Axiom leverages this technology for AI tasks like validat-
ing online content and identifying deepfakes, promoting data integrity and trustless
processing, thus positioning itself at the forefront of secure and intelligent blockchain
applications.

3.15.9 Bagel
Website: https://bagel.net/
One liner: Privacy preserving AI marketplace.
Description: Open protocol for humans and autonomous AI agents to collectively
build, trade and license Machine Learning datasets. Open protocol for humans and au-
tonomous AI agents to collectively build, trade and license Machine Learning datasets.

3.15.10 BitsCrunch
Website: https://bitscrunch.com/
One liner: Empowering the NFT ecosystem with AI-driven insights and security.
Description: BitsCrunch is a project integrating AI with cryptocurrency to offer a
decentralized NFT analytics and forensics platform, aiming to provide reliable data for
making informed investment decisions and detecting fraud in the NFT market. Its API
facilitates the development of trustworthy NFT and DeFi applications, while a commu-
nity-driven model enhances system reliability across various blockchain infrastructures,
promoting a secure and transparent NFT marketplace.

3.15.11 Bittensor
Website: https://bittensor.com/
One liner: Bittensor: Decentralizing AI Training and Inference via Blockchain Tech-
nology.
Description: Bittensor combines artificial intelligence (AI) and blockchain to create
a blockchain-based network for decentralized data processing and training of AI mod-
els, with these models rewarded in Bittensor’s native cryptocurrency, TAO, for their

96 of 128
 Who's Who in AI x Crypto?

valuable contributions. This system facilitates transparent, decentralized development

and use of AI, ensuring equitable distribution of rewards and aligning with principles
of open-source and decentralized governance.

3.15.12 Bittorrent
Website: https://www.bittorrent.com/token/btt/
One liner: BitTorrent tokenizes the world’s largest decentralized file sharing protocol,
introducing BTT for a crypto-powered web.
Description: BitTorrent integrates blockchain technology with its decentralized file-
sharing protocol through the BTT token on the Tron blockchain, rewarding user par-
ticipation with incentives for seeding and bandwidth sharing, thereby enhancing net-
work efficiency and illustrating a practical application of AI x Crypto in promoting a
decentralized internet ecosystem.

3.15.13 Blackbird.AI
Website: https://www.blackbird.ai/
One liner: Blackbird.AI leverages advanced AI to empower organizations against nar-
rative-driven risks, potentially including crypto-related misinformation and disinfor-
mation.
Description: Blackbird.AI utilizes artificial intelligence to measure and counteract
misinformation threats, offering tools like the Constellation Dashboard and Risk In-
telligence Engine, which are crucial for maintaining integrity in the crypto sector by
scoring and predicting the impact of narrative-based threats. Their technology, though
not crypto-specific, is essential for protecting reputation and financial stability in the
digital assets market.

3.15.14 Blockless
Website: https://blockless.network/
One liner: Blockless - Launching a new era of modular, network neutral applications
on blockchain, leveraging AI for enhanced security and efficiency.
Description: Blockless integrates blockchain technology with AI to facilitate the
secure and efficient development of decentralized Network Neutral Applications (n-
nApps), offering a Modular Application Architecture for enhanced full-stack decentral-
ization. Its unique features, including automated node orchestration, a WASM-based
runtime, and AI-based model inference, support a wide range of programming languages
and major L1/L2 blockchains, significantly reducing development time to under three
weeks while advancing the potential for application versatility at the intersection of
crypto and AI.

97 of 128
 AI x Crypto Primer

3.15.15 C2PA
Website: https://c2pa.org
One liner: C2PA stands for the Coalition for Content Provenance and Authenticity,
leveraging crypto and AI for digital media verification.
Description: The Coalition for Content Provenance and Authenticity (C2PA) employs
cryptographic techniques and AI-driven analysis to establish standards for certifying
the provenance of digital media, leveraging Distributed Ledger Technology (DLT), dig-
ital signatures, and NFTs to ensure content integrity and counter misuse. This initiative
demonstrates the intersection of AI and blockchain technology in enhancing data vali-
dation and combating misinformation by making digital media traceable and verifiable
across platforms.

3.15.16 Ceramic
Website: https://ceramic.network/
One liner: Ceramic is a decentralized data network bridging AI with crypto, enabling
scalable, interoperable, and verifiable data management.
Description: Ceramic merges blockchain and AI by offering a decentralized ledger
for high-volume data management, pairing the trust of blockchain with efficiency akin
to traditional databases. It facilitates AI-driven application development through ver-
ifiable data storage and peer-to-peer streaming, using ComposeDB and GraphQL for
enhanced data interaction, thereby supporting trusted and scalable data access crucial
for AI training and inference in decentralized settings.

3.15.17 Chainbase
Website: https://chainbase.com/
One liner: Chainbase, the seamless bridge for Web3 data infrastructure, integrating
AI for smarter crypto applications.
Description: Chainbase streamlines Web3 development by offering a browser-based
platform for real-time blockchain data integration, processing, and AI-driven analytics
across multiple ecosystems. It automates indexing and querying to facilitate efficient AI
model development and application in the crypto domain, significantly cutting down
on development time.

3.15.18 ChainLink
Website: https://chain.link/
One liner: Chainlink: Powering Smart Contracts with Real-World Data

98 of 128
 Who's Who in AI x Crypto?

Description: Chainlink is a decentralized oracle network that connects real-world data

to smart contracts on various blockchains, utilizing a network of decentralized oracles to
enable secure, manipulation-resistant data integration for complex executions in dApps
across DeFi, insurance, and gaming. This facilitates the accurate and reliable operation
of smart contracts, crucially linking blockchain technology with external systems.

3.15.19 Clore.ai
Website: https://clore.ai/
One liner: Clore.ai - Revolutionizing access to distributed supercomputing for AI and
cryptocurrency endeavors.
Description: Clore.ai offers a distributed supercomputing platform that merges AI
and cryptocurrency technologies, providing GPU computing for AI training, video ren-
dering, and crypto mining. It features a user marketplace for computing resources,
utilizes a Proof-of-Work blockchain called Clore Coin to reward participants, and em-
phasizes security, cost efficiency, and environmental sustainability without relying on
premines or ICOs for funding.

3.15.20 Crust Network

Website: https://crust.network/
One liner: Crust Network offers decentralized storage solutions, enhancing Web3′s
data storage capabilities with IPFS integration.
Description: Crust Network combines blockchain’s decentralized capabilities with AI-
enhanced algorithms on the Substrate framework to offer secure, scalable storage solu-
tions for both Web 3.0 and 2.0, incorporating an incentive layer for IPFS to ensure data
integrity and privacy across various applications like DApp hosting and NFT storage.

3.15.21 CryptoIndex
Website: https://www.cdc.gov/parasites/crypto/index.html
One liner: Revolutionizing crypto investment through AI-driven analysis and diversi-
fied exposure.
Description: CryptoIndex utilizes artificial intelligence to intelligently analyze and
derive insights from the cryptocurrency market, facilitating a diversified investment
strategy for both individual and institutional investors through the Bitwise 10 Crypto
Index Fund (BITW). This fund, emphasizing AI’s role in cryptocurrency investment
strategies, focuses on the top 10 cryptocurrencies by market capitalization, ensuring
monthly rebalancing and risk monitoring to optimize performance.

99 of 128
 AI x Crypto Primer

3.15.22 Fact Protocol

Website: https://fact.technology/
One liner: Decentralized fact-checking and misinformation combat platform leverag-
ing AI and blockchain technologies.
Description: Fact Protocol combines AI and blockchain technology to establish a
decentralized system for combating fake news, utilizing AI-driven fact-checking and
blockchain’s transparency to ensure information integrity. It leverages token incentives
for community participation in verifying content, supporting scalability, privacy, and
multi-chain functionality, while introducing governance features to enable community-
driven project direction via FACT DAO.

3.15.23 Filecoin
Website: https://filecoin.io/
One liner: Decentralized storage network for humanity’s most important information.
Description: Filecoin is a blockchain-based network offering decentralized data stor-
age, utilizing its cryptocurrency, FIL, to incentivize storage space sharing, aiming to
enhance web security and efficiency by reducing dependency on centralized storage
solutions. It integrates with the Interplanetary File System (IPFS) to optimize data
storage and retrieval processes.

3.15.24 Flare Network

Website: https://flare.network/
One liner: Flare Network leverages AI and blockchain to deliver decentralized, secure,
and scalable data solutions.
Description: Flare Network integrates AI with cryptocurrency through its Flare Time
Series Oracle (FTSO) and State Connector, offering decentralized data feeds and secure,
trustless data usage from various blockchains and the internet within smart contracts.
This facilitates the development of AI-driven decentralized applications (dApps) and
services, ensuring data integrity and broadening blockchain technology’s utility and
scalability.

3.15.25 Fluence
Website: https://fluence.network/
One liner: Decentralized serverless platform & computing marketplace, leveraging
blockchain for Web 3.0.
Description: Fluence is a decentralized, serverless computing platform that facilitates
the development of AI-driven applications by merging blockchain technology with ar-

100 of 128
 Who's Who in AI x Crypto?

tificial intelligence. It aims to democratize computing resources through a secure, scal-

able ecosystem, underscored by a token economy, which encourages innovation at the
intersection of AI and crypto, enhancing the efficiency and security of decentralized
applications.

3.15.26 Flux
Website: https://runonflux.io/
One liner: Flux enables decentralized Web 3.0 applications, combining AI compute
solutions, cryptocurrency, and a global computational network.
Description: Flux operates a decentralized computational network that integrates
AI and crypto, utilizing its POW cryptocurrency, FLUX, to support the development
and deployment of DApps with an emphasis on AI compute solutions. It introduces
FluxCore for AI computations and employs Proof of Useful Work to enhance GPU ef-
ficiency for AI tasks, within an infrastructure boasting over 15,000 nodes and extensive
computational resources, marking it as a notable bridge between blockchain and AI
technologies.

3.15.27 Fraction AI
Website: https://fractionai.xyz/
One liner: Supercharge your AI with human-powered data at a fraction of the cost.
Description: Fraction AI streamlines AI model training by providing high-quality hu-
man-generated data, integrating the efficiency of cryptocurrency mechanisms for more
effective and cost-efficient AI development across multiple sectors. This approach accel-
erates AI adoption and refinement, particularly within the crypto industry, by ensuring
the availability of optimized datasets.

3.15.28 Grass
Website: https://www.getgrass.io/
One liner: Grass: Revolutionizing AI Training with Decentralized Web Data Accessi-
bility.
Description: Grass by Wynd Network is a decentralized web scraping platform that
utilizes unused internet bandwidth from users to generate datasets for AI training, en-
suring fair compensation through network ownership. This system leverages residential
IP addresses to overcome data scraping restrictions, supporting efficient AI research
while prioritizing security and ethical data collection practices.

3.15.29 Humans.ai
Website: https://humans.ai/

101 of 128
 AI x Crypto Primer

One liner: Empowering the creation and governance of ethical AI on a scalable

blockchain platform.
Description: Humans.ai merges AI and blockchain technology to enhance human po-
tential ethically, leveraging a Proof of Human mechanism and AI tools for user-gener-
ated projects, including the pioneering use of AI NFTs for idea security and governance
on the blockchain. The ecosystem allows for diverse applications such as metaverse
avatar creation and utilizes the $HEART token to fund AI innovations, all within a
framework emphasizing community, transparent AI use, and ethical standards.

3.15.30 Innodata
Website: https://innodata.com/
One liner: Innodata leverages artificial intelligence and human expertise to solve com-
plex data engineering challenges, fueling data-centric AI initiatives.
Description: Innodata integrates AI with crypto by providing AI-enabled software
platforms and managed services, focusing on generating AI training data and building,
maintaining, and augmenting AI and ML models. Their approach combines human-
operated tools with active learning, enhancing operational efficiencies and impacting
the AI lifecycle across different domains.

3.15.31 IPFS
Website: https://ipfs.tech/
One liner: IPFS is revolutionizing the way data is stored and accessed, offering a
decentralized solution for resilient, efficient, and censorship-resistant data sharing.
Description: IPFS integrates AI and crypto by providing a decentralized storage so-
lution that ensures data integrity for AI applications, facilitating secure and efficient
data sharing and model training. Its decentralized nature supports AI-driven DeFi
applications, improves resilience against censorship, and promotes cooperative model
training and sharing, enhancing both AI and crypto ecosystems.

3.15.32 IQ.wiki
Website: https://iq.wiki/
One liner: The World’s Largest Blockchain & Crypto Encyclopedia.
Description: IQ is a comprehensive Wiki for all things Blockchain & Crypto. IQ GPT
is a crypto-focused AI model that provides insights into intricate terms, live market
trends, and breaking news. It is powered by the IQ token, which enables holders to
participate in governance and get access to additional features.

102 of 128
 Who's Who in AI x Crypto?

3.15.33 JACKAL
Website: https://www.jackalprotocol.com/
One liner: Jackal Protocol leverages blockchain technology to offer decentralized, on-
chain data storage with enhanced interoperability and security.
Description: Jackal Protocol employs blockchain technology for decentralized data
storage, emphasizing secure, on-chain management of data permissions and ownership
through an application-specific blockchain. Its architecture promotes interoperability
across ecosystems, enhancing the application of AI in crypto by ensuring secure data
access and transfer for various uses, including DeFi.

3.15.34 Koii Network

Website: https://www.koii.network/
One liner: Koii Network revolutionizes AI x Crypto through decentralized, commu-
nity-powered data and compute sharing.
Description: The Koii Network integrates AI and blockchain technology to create a
cost-efficient, distributed cloud environment for AI model training and digital content
creation. It uses attention-tracking NFTs for fair remuneration, leverages consumer
devices for a global compute-sharing marketplace, and employs ORCA technology for
supporting diverse frameworks and languages, establishing an eco-conscious and cre-
ator-friendly platform for AI development in the crypto domain.

3.15.35 Kudos
Website: https://www.cudos.org/
One liner: CUDOS is a scalable blockchain that integrates cloud and blockchain for
AI and Web3 applications.
Description: CUDOS integrates cloud computing with blockchain through a decen-
tralized platform on the Cosmos ecosystem, offering scalable solutions for AI, the meta-
verse, and Web3 via a proof-of-stake Layer 1. It enables smart contracts, NFT deploy-
ment, and Oracle connections through a vast user network, emphasizing sustainability
with its 100% carbon-neutral operation.

3.15.36 Lasso Labs

Website: https://twitter.com/lasso_labs?lang=en
One liner: Lasso Labs is revolutionizing NFT utility by building data products for
the next generation of the internet, leveraging AI and crypto.
Description: Lasso Labs has secured $4.2M to enhance NFT functionality and inte-
gration into digital commerce by employing AI and blockchain technology. Their work

103 of 128
 AI x Crypto Primer

focuses on evolving NFTs beyond speculative assets, using data analytics to embed
these digital objects more fully into digital identities and marketplaces.

3.15.37 Matrix AI Network

Website: https://www.matrix.io/
One liner: Matrix AI Network is pioneering the integration of AI and blockchain,
stepping into a future synergized through neuroscience with Matrix 3.0.
Description: Matrix AI Network combines artificial intelligence with blockchain tech-
nology to enhance transaction speed, security, and ease of use, notably with the intro-
duction of Matrix 3.0, which integrates AI, blockchain, and neuroscience for a self-
evolving AI platform that supports secure data handling and transparent sharing, along
with democratizing access to on-chain AI resources and pioneering Avatar Intelligence
(AvI), where digital and human consciousness may converge.

3.15.38 nil; Foundation

Website: https://nil.foundation/
One liner: nil; Foundation focuses on elevating Ethereum’s scalability and security
through zero-knowledge proofs.
Description: #ERROR!

3.15.39 NumerAI
Website: https://numer.ai/
One liner: Numerai: Harnessing the power of AI and blockchain for a decentralized
stock market prediction platform.
Description: Numerai merges AI with blockchain through an Ethereum-based plat-
form, decentralizing financial forecasting by rewarding data scientists with Numeraire
(NMR) tokens for contributing accurate machine learning models for stock market pre-
dictions. Founded in 2015, it operates as a global data science competition aimed at
enhancing stock market hedging strategies via collective intelligence.

3.15.40 Oasis Network

Website: https://oasisprotocol.org/
One liner: Oasis Network is a privacy-enabled, scalable layer-1 blockchain designed
for DeFi, NFTs, and a secure, privacy-focused Web3.
Description: Oasis Network is a layer-1 blockchain platform designed for Web3 ap-
plications, focusing on privacy, scalability, and low costs, and supports diverse projects
like DeFi, GameFi, NFTs, and the Metaverse. It integrates advanced privacy features
and smart contract flexibility, working with partners like Meta and BMW, to offer a

104 of 128
 Who's Who in AI x Crypto?

secure and efficient environment for developing private or public decentralized applica-
tions, aiming to advance data governance and privacy in Web3 ecosystems.

3.15.41 OctaSpace
Website: https://octa.space/
One liner: OctaSpace leverages blockchain to revolutionize distributed computing,
data storage, and AI model training.
Description: OctaSpace is a distributed computing platform merging AI and crypto
technology to offer cloud storage, decentralized VPN services, and a CPU/GPU re-
source rental marketplace, catering to AI model training, gaming, and IoT, with a focus
on data privacy through its decentralization. It operates on a unique Layer 1 EVM-
compatible blockchain that combines Proof-of-Work and Proof-of-Authority consensus
mechanisms, aiming to improve network efficiency and scalability while promoting the
reuse of idle computing power and enhancing its ecosystem with AI/ML tools and
blockchain expansions to bolster the utility of its $OCTA coin.

3.15.42 OpenOrigins
Website: https://www.openorigins.com/
One liner: Empowering organizations to secure and validate digital media against
synthetic alteration through advanced cryptography and AI-driven verification.
Description: OpenOrigins combines artificial intelligence and blockchain technology
to enhance the security and validation of digital media, ensuring the authenticity and
integrity of content across various industries by utilizing a blockchain-secured, tamper-
proof database to protect against synthetic content and verify the origin and unaltered
status of media archives.

3.15.43 Optic.xyz
Website: https://twitter.com/Optic_xyz
One liner: Optic is revolutionizing the NFT space by enhancing trust and safety using
AI technology.
Description: Optic.xyz is developing an AI engine to authenticate and secure NFTs by
processing millions of tokens daily for potential counterfeit or copyright issues, aiming
to enhance trust and safety in the digital marketplace. With $11 million in seed funding,
the company provides a Marketplace Moderation tool for platforms like OpenSea and
plans to offer a public API and additional tools for Web3 developers, NFT creators,
and collectors, targeting systemic fraud and trust challenges in the ecosystem.

105 of 128
 AI x Crypto Primer

3.15.44 Oraclize
Website: www.oraclize.it
One liner: Oraclize serves as a bridge to securely integrate real-world data into
blockchain smart contracts, enhancing DeFi and other blockchain applications with
external APIs.
Description: Oraclize serves as a bridge between blockchain oracles and decentral-
ized finance (DeFi) by integrating off-chain data, such as financial market APIs and
weather information, into on-chain smart contracts. This integration ensures that DeFi
platforms and other blockchain applications can operate with real-time, accurate data
without sacrificing security or decentralization, thus enabling a variety of trustless in-
teractions with real-world data.

3.15.45 Oraichain
Website: https://coinmarketcap.com/currencies/oraichain-token/
One liner: Oraichain - the world’s first AI-powered oracle and ecosystem for
blockchains, striving to become the foundational AI Layer 1.
Description: Oraichain integrates artificial intelligence with blockchain technology
to serve as the first AI-powered oracle, enhancing smart contracts and decentralized
applications (DApps) by supplying AI-verified data and aiming to become a pivotal
AI-centric Layer 1 blockchain. It offers a diverse ecosystem including AI-driven services
such as an AI Marketplace, on-chain Verifiable Random Function (VRF), and AI-en-
abled NFT generation, aimed at advancing the functionality and interoperability of
smart contracts and DApps.

3.15.46 Origin Trail

Website: https://origintrail.io/
One liner: OriginTrail - Empowering verifiable, AI-ready Knowledge Assets in a de-
centralized knowledge graph.
Description: OriginTrail integrates blockchain and AI to offer a Decentralized Knowl-
edge Graph (DKG) that ensures the integrity and provenance of data, enhancing AI
capabilities by providing access to a verified information network. This amalgamation
aims to improve the efficiency and reliability of AI agents across various industries
by establishing a secure, trustworthy foundation for data creation, verification, and
querying.

3.15.47 Pond Discoveries

Website: https://cryptopond.xyz/
One liner: AI Calculated Token Recommendation

106 of 128
 Who's Who in AI x Crypto?

Description: Pond’s model has the unique ability to comprehend the flow of blockchain
data, guiding you seamlessly from chaos to order by providing you with intelligently
compressed information directly from the on-chain world.

3.15.48 Powerloom
Website: https://powerloom.io/
One liner: Powerloom: The Composable Data Network revolutionizing Web3 with
Decentralized Finance (DeFi) and Data Availability.
Description: Powerloom is developing a blockchain and AI-powered composable data
network aimed at enhancing Web3 applications by enabling developers to build sophis-
ticated data applications and improve DeFi experiences through decentralized, consen-
sus-based data snapshotting and aggregation. Its alpha-stage product, Boost, facilitates
the customization and access to datapoints across EVM chains, promoting a shift to-
wards decentralized data accuracy and security, supported by a recent $3.1M funding
round to advance cross-chain logic and applications.

3.15.49 Reclaim Protocol

Website: https://www.reclaimprotocol.org/
One liner: Reclaim Protocol - Revolutionizing identity and reputation verification
through cryptographic proofs without compromising user privacy.
Description: Reclaim Protocol integrates AI and blockchain technology to provide
a secure, private verification system for users’ identities and reputations online, utiliz-
ing zero knowledge proofs and digital signatures to facilitate decentralized applications
such as DeFi and KYC without compromising sensitive information. Its SDK enables
connections from over 200 websites, enhancing user experiences through verified online
behavior while preserving privacy.

3.15.50 SapienAI
Website: https://www.linkedin.com/in/trevorkoverko
One liner: SapienAI leverages the intersection of AI and crypto to innovate within
the Web3 space.
Description: SapienAI, led by Trevor Koverko, specializes in merging artificial intelli-
gence with cryptocurrency to advance blockchain technologies and improve the crypto
ecosystem, emphasizing the creation of secure, transparent, and efficient solutions at
the intersection of AI and Web3.

3.15.51 Sia
Website: https://coinmarketcap.com/currencies/siacoin/

107 of 128
 AI x Crypto Primer

One liner: Decentralized cloud storage platform using blockchain technology.

Description: Sia combines AI and blockchain to provide a decentralized cloud storage
platform that uses its native Siacoin for transactions, focusing on privacy and security
by distributing encrypted files across multiple nodes, which ensures data safety, avail-
ability, and censorship resistance.

3.15.52 Sphere AI Agents

Website: https://twitter.com/RollmanMining/status/1750240075433083185
One liner: Revolutionizing the Crypto Sphere through Advanced AI Technology.
Description: Fetch.AI integrates AI with cryptocurrency through a Multi-Agent Sys-
tem (MAS), enabling Autonomous Economic Agents (AEA) to improve task efficiency
and reduce costs by removing the need for intermediaries, with the goal of automat-
ing and transforming various sectors. Meanwhile, Sphere 3D’s acquisition of NuMiner
NM440 machines signals a move towards becoming a leading carbon-neutral bitcoin
miner, emphasizing the convergence of AI and cryptocurrency in advancing sustainable,
decentralized finance and secure blockchain technology.

3.15.53 Spice AI
Website: https://spice.ai/
One liner: Spice.ai accelerates the development of next-gen intelligent software with
a focus on leveraging web3 data through AI.
Description: Spice.ai streamlines the development of AI-driven applications in the
crypto and web3 sectors by providing a backend-as-a-service that includes time-series
data management, custom ETL processes, and ML model training and inferencing, all
optimized for petabyte-scale data in web3 development. It reduces the need for deep
data science expertise, facilitating the creation of innovative apps across DeFi, security,
and AI agents, thereby advancing the integration of AI with blockchain technologies.

3.15.54 Storj
Website: https://www.storj.io/
One liner: Storj is revolutionizing cloud storage with a decentralized, secure, and en-
vironmentally sustainable platform.
Description: Storj is an open-source, decentralized cloud storage platform, using
blockchain technology to enhance security, privacy, and cost-efficiency. Its global net-
work of distributed nodes encrypts and disperses data, offering a sustainable and secure
storage solution particularly suited for AI development, with its S3 compatibility em-
phasizing its utility in the AI x Crypto domain.

108 of 128
 Who's Who in AI x Crypto?

3.15.55 The Graph

Website: https://thegraph.com/
One liner: The Graph is a decentralized protocol for indexing and querying data from
blockchains, facilitating the creation of decentralized applications with efficient data
retrieval.
Description: The Graph integrates AI to index blockchain data efficiently, creating
open APIs (subgraphs) that facilitate the development of responsive DApps across
various sectors like DeFi, NFTs, and DAOs. This decentralized platform leverages a
global community for network maintenance and data integrity, enhancing access and
usability of blockchain data for developers.

3.15.56 WorldCoin
Website: https://worldcoin.org/
One liner: WorldCoin integrates AI to create a privacy-preserving global identity net-
work, enhancing access to a human-centric global economy.
Description: WorldCoin combines artificial intelligence and blockchain to establish
a global identity and financial network through a privacy-centric AI-enabled verifica-
tion system (World ID), employing multispectral sensors and zero-knowledge proofs
to ensure anonymous yet unique human identification. It aims to democratize global
economic participation by distributing WLD tokens, exhibiting over 3 million sign-ups
across 120 countries, to integrate human identity into digital infrastructure securely.

3.16 DeFi
3.16.1 Colony
Website: https://www.colonylab.io/
One liner: Colony Lab: Accelerating the Avalanche Ecosystem Through Decentralized
Community-Driven Fundraising and Liquidity.
Description: Colony Lab is a platform in the Avalanche ecosystem that combines AI
and crypto to enhance decentralized fundraising and investment opportunities for early-
stage projects. It focuses on providing capital, ecosystem support, and exposure, while
also engaging the community through seed investments and yield-generating products,
incorporating AI to optimize investment strategies and project support.

3.16.2 Glif
Website: https://www.glif.io/en
One liner: GLIF - Decentralized finance hub for the Filecoin ecosystem, facilitating
liquid Filecoin leasing and offering a suite of Filecoin infrastructure tools.

109 of 128
 AI x Crypto Primer

Description: GLIF merges AI and crypto within the Filecoin network, offering decen-
tralized finance solutions through tools such as a web wallet, multisig wallet, Filecoin
nodes, an autonomous notary service, and an FEVM-enabled block explorer. Its flag-
ship, GLIF Pools, introduces DeFi protocols to the Filecoin ecosystem, focusing on FIL
staking for increased liquidity and supporting the expansion of a decentralized storage-
based open data economy through smart contracts.

3.16.3 KingDefi
Website: https://coinmarketcap.com/currencies/kingdefi/
One liner: KingDeFi - A dual-role DeFi project offering analytics, monitoring, and
optimized yield farming on BSC and Solana.
Description: KingDeFi merges AI-driven analytics and DeFi yield optimization on
the Binance Smart Chain and Solana networks, providing users with tools like a liquid-
ity pool search engine and portfolio tracking while enabling earning through KROWN
tokens via liquidity provision and staking. Its standout feature is an automated recom-
mendation system for yield-generating platforms based on wallet contents, supported
by a comprehensive dashboard of key DeFi metrics, thus serving as an analytical bridge
for effective DeFi strategy execution.

3.16.4 Mozaic
Website: https://mozaic.finance/
One liner: Empowering Yield Farming and Liquidity Strategies with AI
Description: Mozaic integrates AI into DeFi, automating and optimizing yield farm-
ing and liquidity management across multiple blockchains using the LayerZero protocol,
improving user experience and return potential by adapting to market changes in real
time, while maintaining a focus on security and sustainability through collaborations
with key technology and financial partners.

3.16.5 Nexus Mututal

Website: https://nexusmutual.io/
One liner: Nexus Mutual uses blockchain to democratize insurance, leveraging collec-
tive risk assessment and management in the DeFi space.
Description: Nexus Mutual utilizes AI and blockchain technology to offer a decen-
tralized mutual insurance protocol within the DeFi sector, focusing on smart contract
security and risk assessment using staking pools and NXM tokens to mitigate crypto-
related risks. Its effectiveness in managing and responding to incidents like the Euler
Finance hack underscores its significant role in enhancing DeFi security.

110 of 128
 Who's Who in AI x Crypto?

3.16.6 Rinzo
Website: https://twitter.com/rinzo?lang=en
One liner: The Renzo Protocol augments Ethereum staking by securing Actively Val-
idated Services for enhanced yields.
Description: The Renzo Protocol aims to enhance Ethereum’s staking ecosystem by
interfacing with EigenLayer, focusing on elevating yield opportunities through Actively
Validated Services within DeFi and Liquid Staking on the Ethereum Mainnet. It uti-
lizes a blend of blockchain technology and AI to optimize and profitably streamline
staking operations.

3.16.7 SpacePort
Website: https://www.spaceport.xyz/
One liner: SpacePort - The democratized open IDO platform launching crypto pro-
jects to success.
Description: SpacePort combines AI and blockchain technology to offer a secure and
open IDO platform, enhancing project fundraising by automating launch processes
with a Smart Contract Factory and providing comprehensive pre-launch support. It
safeguards investors against market risks and ensures equitable IDO access through a
fully on-chain approach, aiming at transparency and fairness in the DeFi ecosystem.

3.16.8 Upshot
Website: https://upshot.xyz/
One liner: Revolutionizing DeFi with decentralized, scalable, and self-improving AI
network powered by Proof of Alpha.
Description: Upshot integrates artificial intelligence with decentralized finance (DeFi)
via a decentralized AI network that uses a ‘Proof of Alpha’ protocol to encourage the
contribution and use of collective intelligence, enhancing the network’s self-improve-
ment capabilities. Its platform offers AI-driven DeFi applications and ensures secure,
unbiased outcomes with zkML and cryptographic verification, promoting global inno-
vation in AI within DeFi.

3.17 Hardware
3.17.1 Akash
Website: https://akash.network/
One liner: Akash Network is an open-source Supercloud for decentralized cloud com-
puting, combining AI & Crypto to innovate the future of cloud infrastructure.

111 of 128
 AI x Crypto Primer

Description: Akash Network offers a decentralized, blockchain-based cloud comput-

ing marketplace that uses Kubernetes and Cosmos for secure and efficient application
hosting. It notably reduces costs (up to 85% lower than traditional providers) through
a ‘reverse auction’ pricing system and facilitates distributed machine learning, high-
lighting its utility in the intersection of AI and Crypto.

3.17.2 Akash
Website: https://akash.network/
One liner: Akash Network is an open-source Supercloud for decentralized cloud com-
puting, combining AI & Crypto to innovate the future of cloud infrastructure.
Description: Akash Network offers a decentralized, blockchain-based cloud comput-
ing marketplace that uses Kubernetes and Cosmos for secure and efficient application
hosting. It notably reduces costs (up to 85% lower than traditional providers) through
a ‘reverse auction’ pricing system and facilitates distributed machine learning, high-
lighting its utility in the intersection of AI and Crypto.

3.17.3 ATOR
Website: https://www.ator.io/
One liner: ATOR integrates AI with crypto to enhance online anonymity, offering
rewards for secure network participation.
Description: The ATOR Protocol integrates AI with blockchain to enhance the se-
curity and efficiency of the Tor network by rewarding network participants with cryp-
tocurrency for their uptime, using a Proof-of-Uptime model to encourage a robust
and secure network. By deploying AI for network optimization and introducing user-
friendly hardware like the ATOR Router Hotspot, the project aims to expand its global
anonymity network, leveraging AI and on-chain incentives to improve digital privacy
and network performance.

3.17.4 Auradine
Website: https://auradine.com/
One liner: Auradine is revolutionizing the blockchain landscape through innovative
Bitcoin mining systems and advanced AI integration.
Description: Auradine focuses on developing efficient infrastructure solutions at the
crypto and AI intersection, notably through its Teraflux™ Bitcoin miners, which mark a
significant improvement in energy efficiency and transaction speeds due to a pioneering
four-nanometer silicon transistor. Beginning with a substantial $81 million fundraising
without a product, the company emphasizes innovation in combining AI and blockchain
to enhance web infrastructure security, privacy, and efficiency.

112 of 128
 Who's Who in AI x Crypto?

3.17.5 exaBITS
Website: https://www.exabits.ai/
One liner: Accelerating crypto and AI applications through cost-effective, high-per-
formance GPU compute.
Description: ExaBITS utilizes an extensive GPU network, including NVIDIA H100s,
to offer affordable accelerated computing for AI model training and inference specifi-
cally within the cryptocurrency sector. Through this approach, exemplified by Nebula
Block’s 71% cost reduction and 30% performance increase, the project aims to enhance
the efficiency and reduce the financial barriers for incorporating AI into blockchain
applications.

3.17.6 GPU.net
Website: https://gpu.net/
One liner: Decentralized GPU infrastructure revolutionizing Generative AI, 3D, and
Web3 by offering computing resources for AI model training, 3D rendering, VFX, and
more.
Description: GPU.net is a decentralized network providing essential GPU infrastruc-
ture to address shortages and high costs in the AI sector, particularly for tasks like
generative AI training, and 3D rendering, while also supporting cryptocurrency mining.
It aims to democratize access to high-performance computing through services such
as 3D conversion tools and an open-source LLM library, enhancing data privacy and
efficiency in the AI x Crypto space through decentralized operations and token-based
resource prioritization.

3.17.7 GPU.net
Website: https://gpu.net/
One liner: Decentralized GPU infrastructure revolutionizing Generative AI, 3D, and
Web3 by offering computing resources for AI model training, 3D rendering, VFX, and
more.
Description: GPU.net is a decentralized network providing essential GPU infrastruc-
ture to address shortages and high costs in the AI sector, particularly for tasks like
generative AI training, and 3D rendering, while also supporting cryptocurrency mining.
It aims to democratize access to high-performance computing through services such
as 3D conversion tools and an open-source LLM library, enhancing data privacy and
efficiency in the AI x Crypto space through decentralized operations and token-based
resource prioritization.

113 of 128
 AI x Crypto Primer

3.17.8 GPUtopia
Website: https://gputopia.ai/beta
One liner: GPUtopia is revolutionizing the GPU market by creating a decentralized
marketplace for buying and selling GPU compute power, leveraging Bitcoin transac-
tions.
Description: GPUtopia is a decentralized platform that leverages the intersection of
AI and cryptocurrency by enabling users to rent out or purchase GPU computing power
with Bitcoin. This initiative aims to alleviate the current GPU shortage and distribute
computing resources more equitably, supporting a variety of tasks from Bitcoin mining
to AI processing on a global mesh network.

3.17.9 Ingonyama
Website: https://www.ingonyama.com/
One liner: Ingonyama pioneers in Zero Knowledge Hardware Acceleration, redefining
cryptographic security from software to silicon.
Description: Ingonyama is developing the world’s first Zero Knowledge Processing
Unit (ZPU) to accelerate ZK proofs, tackling computational challenges by blending
various disciplines for improved data security, scalability, and sovereignty in the IT
ecosystem. This effort, supported by leading industry builders, aims to make ZK tech-
nology more accessible and efficient.

3.17.10 IORNA
Website: https://investors.ironmountain.com/news-and-events/press-releases/press-
release-details/2020/Keevo-Announces-Strategic-Partnership-with-Iron-Mountain-to-
Protect-Crypto-Currency-Users-and-Their-Heirs/default.aspx
One liner: Combining cutting-edge security and privacy for crypto asset management
and inheritance through innovative hardware and strategic partnerships.
Description: IORNA, in partnership with Keevo and Iron Mountain, introduces a
solution for secure cryptocurrency asset management, utilizing AI-enhanced hardware
wallets with multi-factor authentication and a novel beneficiary service. This system,
featuring the Keevo Model 1 wallet and Carbon Key™ technology, emphasizes secure,
inheritance-friendly crypto asset storage, leveraging Iron Mountain’s air-gapped facili-
ties for added security, thereby merging AI with crypto to enhance asset security and
ease of access for owners and their heirs.

3.17.11 IOTeX
Website: https://iotex.io/

114 of 128
 Who's Who in AI x Crypto?

One liner: IoTeX - Connecting Real World to Web3 with a focus on empowering IoT
and smart devices through blockchain technology.
Description: IoTeX is developing a blockchain platform for the Internet of Things
(IoT), aiming to merge physical and digital worlds by enabling secure and decentralized
IoT applications and smart devices. Their work includes a collaboration with Solana for
real-time device analytics and the creation of a confidential computing hardware stack,
contributing to the establishment of decentralized private IoT networks essential for
advancing AI within the cryptocurrency sector. This infrastructure supports decentral-
ized identities, secure hardware, and real-world data oracles like Pebble, enhancing AI’s
capability to securely interact with real-world data for applications including NFTs,
weather derivatives, and machine learning-powered mining.

3.17.12 Nosana
Website: https://nosana.io/
One liner: Revolutionizing GPU cloud computing with accessible, affordable compute
power for AI on Solana.
Description: Nosana is developing a decentralized GPU cloud computing platform on
the Solana blockchain, utilizing the $NOS token to allow users to monetize idle hard-
ware and provide affordable AI computation power. This approach facilitates broader
access to computational resources, addressing GPU shortages and integrating AI with
cryptocurrency to enhance AI process execution.

3.17.13 OpenAgents
Website: https://bitcoinnews.com/adoption/openagents-earn-bitcoin-leasing-gpu-for-
ai-tasks/
One liner: Revolutionizing the AI x Crypto space by monetizing GPU power for Bit-
coin rewards.
Description: OpenAgents, formerly GPUtopia, is a decentralized marketplace that
enables users to earn Bitcoin by contributing their idle GPU power for AI computa-
tional tasks. It leverages the Bitcoin and Lightning Network to facilitate a cost-effective
platform for AI services, aiming to build a community-driven ecosystem for the devel-
opment, use, and promotion of AI agents, rewarding participants in Bitcoin for their
contributions across the AI agent lifecycle.

3.17.14 Render Network

Website: https://rendernetwork.com/
One liner: Render Network is a decentralized GPU rendering network, revolutionizing
3D content creation and AI with blockchain technology.

115 of 128
 AI x Crypto Primer

Description: Render Network offers a blockchain-based platform that connects 3D

artists and studios with GPU owners to provide scalable and efficient rendering and AI
computational resources. Operating on the Ethereum blockchain, the project focuses on
enhancing AI-driven content creation and rendering for various applications including
media, gaming, and medical, by utilizing decentralized GPU computing power.

3.17.15 Shoggoth
Website: https://shoggoth.systems/
One liner: Shoggoth Network facilitates decentralized AI by enabling the sharing of
AI resources over a P2P network.
Description: The Shoggoth Network is designed as a decentralized platform utilizing
blockchain to facilitate the dissemination and collaborative development of AI resources
such as papers, datasets, code, and models through a peer-to-peer network. It employs
Git for resource management, promoting global access and innovation in AI by stream-
lining the sharing process among developers, researchers, and enthusiasts.

3.17.16 Stelia Connectivity

Website: https://stelia.io/
One liner: Stelia Connectivity: Empowering the AI x Crypto space with ultra-low
latency, high-capacity networking solutions.
Description: Stelia Connectivity is developing the world’s first Tier 1 AI network
backbone tailored for hyperscale AI and Crypto demands, offering an elastic, high-speed
network that enables efficient AI model training, inference, and blockchain operations.
Their network, enhanced through strategic partnerships, provides agile and scalable
connections essential for deploying large language models and supporting the growth
of AI and blockchain technologies.

3.18 Inference
3.18.1 Aizel Network
Website: https://aizelnetwork.com/
One liner: Aizel Network is revolutionizing blockchain with trustless, on-chain AI,
ensuring Web2 speed & costs.
Description: Aizel Network combines AI and blockchain technology, offering a plat-
form where machine learning models can execute trustless, verifiable inferences on-chain
using Multi-Party Computation (MPC) and Trusted Execution Environments (TEEs)
for security. It promises to equip any smart contract across blockchain networks with
scalable, privacy-preserving AI capabilities, facilitated by a team blending expertise in
data science, AI, and blockchain.

116 of 128
 Who's Who in AI x Crypto?

3.18.2 Aleo
Website: https://aleo.org/
One liner: Aleo leverages zero-knowledge proofs to enable fully private applications
on a scalable, privacy-first blockchain.
Description: Aleo leverages zero-knowledge proofs (ZKPs) in its layer-1 blockchain
platform to enable the creation of decentralized applications that emphasize user pri-
vacy and data security, without compromising scalability or security. Through its na-
tive programming language, Leo, and infrastructure like snarkOS and snarkVM, Aleo
facilitates the development and deployment of smart contracts with privacy features,
positioning itself at the nexus of AI, crypto, and privacy-enhancing technologies.

3.18.3 Aztec
Website: https://aztec.network/
One liner: Aztec is a privacy-first Layer 2 scaling solution on Ethereum enabling pri-
vate transactions and smart contracts.
Description: Aztec Network enhances blockchain privacy on Ethereum via a Layer 2
solution that supports public and private smart contract executions, leveraging tools
like the Aztec Sandbox and Noir programming language for developing privacy-centric
applications. This initiative, which attracts significant crypto investment and empha-
sizes open-source development, is pivotal for integrating advanced privacy and security
within AI-driven decentralized applications.

3.18.4 ChainML
Website: https://www.chainml.net/
One liner: ChainML: Pioneering the integration of AI and blockchain for decentral-
ized, secure, and accessible AI agents and machine learning models.
Description: ChainML merges artificial intelligence and blockchain technology to de-
velop and deploy AI agents and machine learning models within a decentralized pro-
tocol, enhancing DeFi, gaming, and social dApps. It focuses on providing a generative
AI platform for conversational analytics and an open-source framework for AI devel-
opment, aiming to democratize advanced AI access and promote innovation through
community-contributed, secure, and scalable computation in the Web3 ecosystem.

3.18.5 Ezkl
Website: https://github.com/zkonduit/ezkl
One liner: Ezkl: A pioneering Zero-Knowledge Inference tool for private and secure
AI computations on the blockchain.

117 of 128
 AI x Crypto Primer

Description: Ezkl is a tool and library for integrating deep learning models with
Zero-Knowledge Proofs (ZKP), specifically zk-SNARK, allowing users to verify com-
putations, such as neural network operations, on both private and public data without
revealing the data itself. This integration supports the secure and private execution
of AI tasks within blockchain environments, notably Ethereum, and facilitates confi-
dential verification for decentralized finance (DeFi) and machine learning applications,
with prospects for trustless machine learning inferences and on-chain rewards.

3.18.6 Giza
Website: https://www.gizatech.xyz/
One liner: Giza is revolutionizing the AI x Crypto space by enabling seamless inte-
gration of AI with blockchain technologies.
Description: Giza combines AI and blockchain technology to enable the development
and management of secure, scalable machine learning models on decentralized appli-
cations, enhancing Web3 integrations with a focus on privacy through zero-knowledge
proofs and optimized datasets for ML, thus advancing AI-driven solutions in the crypto
space.

3.18.7 Hypercycle
Website: https://coinmarketcap.com/currencies/hypercycle/
One liner: HyperCycle is revolutionizing the blockchain and AI space with its layer
0++ architecture designed for high-speed, cost-efficient AI microservices execution.
Description: HyperCycle introduces a Layer 0++ blockchain architecture optimized
for efficient and cost-effective on-chain AI microservices, leveraging TODA/IP’s ledger-
less technology and SingularityNET’s Proof of Reputation system for enhanced security
and scalability. This platform aims to support decentralized AI applications, including
swarm AI and decentralized finance, by improving interoperability and processing ca-
pabilities across different blockchain networks.

3.18.8 Lilypad
Website: https://github.com/bacalhau-project/lilypad
One liner: Lilypad establishes a decentralized GPU network for running AI workloads
effortlessly, integrated with blockchain for transaction management.
Description: Lilypad leverages blockchain technology to allow users to conduct AI
tasks on a decentralized network of GPU resources, offering CLI and web interfaces for
engaging with AI models and executing jobs securely through smart contracts. This
platform integrates AI and crypto by enabling decentralized computing, supporting

118 of 128
 Who's Who in AI x Crypto?

advanced AI applications, and facilitating contributions to the web3 ecosystem with

potential in AI inference, machine learning, and scientific research.

3.18.9 Modulus
Website: https://www.modulus.xyz/
One liner: Modulus Labs leverages ZK-proofs to secure and verify AI on the
blockchain, fostering trust in decentralized applications.
Description: Modulus Labs, emerging from Stanford University, secured $6.3 million
in seed funding to enhance crypto security through the integration of zero-knowledge
proofs (ZK-proofs) and zkML, aiming to validate AI outputs on the blockchain without
compromising underlying data. This approach seeks to improve the integrity of AI-
generated content, facilitate the integration of AI in web3 applications, and reduce
reliance on human governance in the blockchain ecosystem, thereby increasing trans-
parency and trust.

3.18.10 Octra
Website: https://twitter.com/octrafoundation?lang=en
One liner: Octra - Empowering the future economy with a decentralized, secure, and
private network utilizing fully homomorphic encryption and machine learning.
Description: Octra combines cryptocurrency with artificial intelligence to improve
digital transaction security and data management using fully homomorphic encryption
(FHE) and machine learning (ML), focusing on decentralized cloud storage that ensures
user data privacy and security through the integration of blockchain, cryptography,
and AI.

3.18.11 Ora
Website: https://www.ora.io/
One liner: Ora redefines smart contracts with zkOracles, integrating AI for next-gen
DeFi and on-chain machine learning.
Description: Ora integrates AI with cryptocurrency by enhancing smart contracts
with Zero-Knowledge Oracles (zkOracles), facilitating the development of advanced de-
centralized finance (DeFi) applications and on-chain machine learning. This approach,
using opML for large-scale AI models on the blockchain and collaborating with Poly-
gon CDK for deploying the zkOracle protocol, ensures decentralized, secure, and trust-
worthy AI-powered applications with a focus on verifiable transparency and on-chain
execution.

119 of 128
 AI x Crypto Primer

3.18.12 Phala Network

Website: https://phala.network/
One liner: Phala Network is a privacy-preserving, off-chain compute platform for
blockchains, leveraging AI and Secure Enclaves (TEEs) for secure, decentralized appli-
cations.
Description: Phala Network is a decentralized platform that integrates AI with
blockchain through trusted execution environments like Intel SGX, enhancing the pri-
vacy and security of Web3 applications. It introduces ‘Phat Contracts’ for secure off-
chain computing and aims to reduce trust dependencies in blockchain with features
such as a trustless MEV ecosystem and AI-enhanced functionalities, focusing on scal-
ability, interoperability, and privacy.

3.18.13 Risc0
Website: https://www.risczero.com/
One liner: Democratizing Zero-Knowledge technology for enhanced security and effi-
cient computation across blockchains.
Description: Risc0 introduces a unique blend of AI and cryptocurrency by leveraging
Zero-Knowledge (ZK) proofs through its Bonsai technology, a decentralized proving
engine designed to enhance security and enable complex computations on blockchain
technology. This project offers a scalable, privacy-focused framework for DeFi and
blockchain applications, allowing for the execution of arbitrary Rust code in a secure,
verifiable manner without compromising private data.

3.18.14 Ritual
Website: https://ritual.net/blog/introducing-ritual
One liner: Ritual, merging cryptography and AI to revolutionize access and innova-
tion in AI models on the blockchain.
Description: Ritual aims to decentralize AI by creating a distributed network where
AI model creators can host their models for unrestricted access, backed by crypto-
graphic integrity and privacy. It bridges AI with blockchain through its Infernet, al-
lowing smart contracts to utilize AI for various applications, targeting to integrate AI
functionalities as a core component of the web3 ecosystem.

3.18.15 SingularityNET
Website: https://singularitynet.io/
One liner: SingularityNET is a decentralized marketplace for AI services, facilitating
the creation, sharing, and monetization of AI technologies.

120 of 128
 Who's Who in AI x Crypto?

Description: SingularityNET merges artificial intelligence and blockchain to establish

a decentralized marketplace for AI services and algorithms, facilitated by the AGIX
token used for transactions and governance. It aims to become the Internet’s Knowl-
edge Layer by integrating Knowledge Graphs with generative AI, contributing to the
development of advanced AI systems, with notable contributions from Dr. Ben Goertzel
and connections to the AI robot, Sophia.

3.18.16 Spectral
Website: https://www.spectral.finance/
One liner: Spectral: Powering the Inference Economy with Decentralized Machine
Intelligence for Web3
Description: Spectral integrates AI and cryptocurrency by deploying a Machine In-
telligence Network that uses zero-knowledge proofs to verify machine learning infer-
ences for direct consumption through smart contracts, aiming to enhance the DeFi
and blockchain spaces. It fosters a decentralized community of data scientists to solve
complex Web3 challenges, securing models through validators and promoting a self-
sustaining ecosystem with a revenue-sharing model that incentivizes high-quality, un-
biased model development for smart contract applications.

3.18.17 TACEO
Website: https://www.theblock.co/post/251178/nil-foundation-taceo-ethereum
One liner: TACEO: Empowering verifiable machine learning models on Ethereum
with privacy-preserving technologies.
Description: TACEO integrates AI and crypto by merging machine learning with
Ethereum blockchain using zero-knowledge proofs, ensuring model validation without
external trust. This initiative, a collaboration between the Nil Foundation and TACEO,
aims to authenticate machine learning models directly on Ethereum’s blockchain, en-
hancing data privacy and enabling secure, provable ML applications within smart con-
tracts across various sectors. Additionally, by focusing on decentralized private compu-
tation, TACEO addresses data ownership and privacy concerns, marking a significant
step towards secure, private data computation in decentralized applications.

3.18.18 Together AI
Website: https://www.together.ai/
One liner: Accelerating the integration of generative AI models into crypto through
innovative cloud platforms, research, and optimizations.
Description: Together AI utilizes cutting-edge AI research to enhance the efficiency
and speed of generative AI applications within the blockchain and crypto sphere, by

121 of 128
 AI x Crypto Primer

leveraging technologies like Cocktail SGD and FlashAttention for up to 9x faster train-
ing of large language models. This improvement not only fosters cost savings but also
amplifies performance, particularly for AI x Crypto projects, backed by accessible, op-
timized APIs for quick deployment and initiatives like the open-source RedPajama
project to promote AI technology democratization.

3.19 Infrastructure
3.19.1 enqAI
Website: https://enqai.com/
One liner: enqAI: Unleashing the potential of uncensored and unbiased AI in the
crypto sphere.
Description: enqAI, previously noiseGPT, merges artificial intelligence with
blockchain technology on the Ethereum platform, emphasizing an uncensored AI ap-
proach within decentralized finance (DeFi). It showcases active market engagement
with a supply of 970 million tokens, highlighting the initiative to blend AI’s generative
capabilities with blockchain’s security and transparency.

3.19.2 NEAR
Website: https://near.org/
One liner: NEAR Protocol - A Community-Run Cloud Computing Platform That
Fosters AI x Crypto Integration.
Description: NEAR Protocol is a layer-one blockchain designed for high transaction
speeds, throughput, and interoperability, making it conducive for decentralized applica-
tions and AI x Crypto innovations. It distinguishes itself with human-readable account
names, the ‘Doomslug’ consensus mechanism for security and efficiency, and initiatives
like FastAuth that foster a developer-focused ecosystem for projects integrating AI and
cryptocurrency, such as Flux and Mintbase.

3.19.3 Theta Network

Website: https://www.thetatoken.org/
One liner: Theta Network: The decentralized infrastructure for video, AI, and enter-
tainment.
Description: Theta Network is a decentralized video streaming platform utilizing
blockchain to incentivize bandwidth sharing, integrated with AI algorithms for en-
hanced content delivery and viewer experience. It offers a Web3 infrastructure sup-
porting diverse applications, including NFTs with DRM, and is validated by major
enterprises like Google, Samsung, and Sony, showcasing its application of AI in the
crypto space for entertainment and more.

122 of 128
 Who's Who in AI x Crypto?

3.20 Infrastructure - Cloud

3.20.1 OORT
Website: https://www.oortech.com/
One liner: OORT, the decentralized cloud revolutionizing privacy, cost savings, and
AI integration across global business operations.
Description: OORT integrates AI with blockchain to provide a privacy-focused de-
centralized cloud service, offering substantial cost savings and operational efficiency
with a suite of AI solutions, including a generative AI agent builder. Highlighted by
partnerships with Lenovo Image and endorsements from Tencent Cloud and Dell Tech-
nologies, the platform is supported by a global network of over 30,544 nodes across 109
countries, demonstrating its scalability and reliability in blending AI with crypto to
enhance business operations.

3.21 Security
3.21.1 Anchain.AI
Website: https://www.anchain.ai/
One liner: AnChain.AI: Revolutionizing Web3 Security and Compliance with AI-
Powered Solutions.
Description: AnChain.AI employs AI to enhance security, compliance, and crime
investigation within the Web3 ecosystem, offering an AI-powered platform that au-
tomates the analysis of crypto transactions and smart contracts for identifying risks
and ensuring AML/CFT compliance. This platform serves over 200 customers, over-
seeing more than $100 billion in daily trading volume, by utilizing AI for tasks such
as blockchain investigation auto-trace, risk assessment, and the detection of suspicious
activities, aiming to improve the safety and integrity of the blockchain environment.

3.21.2 CertiK
Website: https://www.certik.com/
One liner: CertiK leverages AI to enhance blockchain security and smart contract
auditing.
Description: CertiK specializes in enhancing blockchain security through AI and For-
mal Verification technology, offering services such as smart contract and blockchain
audits, and continuous security monitoring. Founded by Columbia and Yale professors
in 2018, CertiK leverages its expertise to identify vulnerabilities and improve the reli-
ability of blockchain systems, making it a key contributor to the AI and crypto inter-
section.

123 of 128
 AI x Crypto Primer

3.21.3 Sardine
Website: https://www.sardine.ai/
One liner: Sardine is revolutionizing fraud prevention and compliance in the cryp-
tocurrency space with AI-driven solutions.
Description: Sardine employs machine learning and generative AI to enhance fraud
prevention and compliance for the crypto and fintech industries, ensuring secure, instant
transactions by analyzing device intelligence and behavior biometrics. This AI-driven
approach, in collaboration with BitPay, facilitates direct cryptocurrency purchases from
bank accounts, aiming to mitigate systemic risks and foster trust in digital currency
exchanges by accurately identifying fraud and improving transaction safety.

3.21.4 TestMachine
Website: https://testmachine.ai/
One liner: AI-powered blockchain security for smart contract vulnerability detection
and resolution.
Description: TestMachine combines AI and blockchain technology to improve security
across various blockchain platforms, including Ethereum and Binance, by offering tools
for the early detection and resolution of vulnerabilities in smart contracts. It integrates
with developer workflows to provide real-time security analysis, optimization, testing,
and detailed reports, leveraging AI for enhanced threat identification and mitigation.

3.22 Training
3.22.1 Bacalhau
Website: https://www.bacalhau.org/
One liner: Bacalhau - Revolutionizing compute over data with security and efficiency.
Description: Bacalhau offers a platform that distributes computing power for data
processing at its source, utilizing Docker containers and WebAssembly for flexible, se-
cure compute jobs. It supports a broad range of architectures and taps into underuti-
lized resources, making it suitable for AI and crypto projects needing efficient, secure
data handling near its origin, with a focus on minimizing data movement and adhering
to privacy standards.

3.22.2 FATE federated learning

Website: https://fate.fedai.org/
One liner: Empowering secure and efficient federated learning with AI and crypto
technologies.

124 of 128
 Who's Who in AI x Crypto?

Description: FATE is an open-source framework that enhances federated learning

through the use of advanced cryptographic techniques, including partial homomorphic
encryption, to secure data while allowing for computational operations on encrypted
data—significantly improving efficiency and reducing costs. Its integration with Intel®
IPP-Cryptography libraries and focus on data privacy positions it as an effective tool for
deploying AI in sectors requiring strict data confidentiality without centralizing sensi-
tive information, demonstrating a practical fusion of AI methodologies with blockchain
and cryptography principles.

3.22.3 Flock.io
Website: https://flock.io/
One liner: FLock.io: Revolutionizing privacy-preserving AI on blockchain with feder-
ated learning.
Description: FLock.io utilizes blockchain technology to create a decentralized plat-
form for AI training, where users contribute to federated learning without compromis-
ing their data privacy. It rewards contributions with cryptocurrency, integrating AI
development with the DeFi ecosystem and exploring advanced areas such as BTC GPT,
GameFi MBTI, and Zero-Knowledge Machine Learning (ZKML), thereby melding AI’s
capabilities with blockchain’s security and decentralization.

3.22.4 Flower federated learning

Website: https://flower.dev/
One liner: Empowering federated learning in diverse environments through the Flower
Framework, compatible with AI and blockchain.
Description: Flower enables federated learning on decentralized devices to enhance
data privacy and security for AI, supporting various ML frameworks like Tensor-
Flow and PyTorch for technology-agnostic implementation. Its interoperability with
blockchain technologies positions it to advance privacy-preserving AI applications
within decentralized finance and security, showcasing the convergence of AI and crypto
through secure, decentralized data management.

3.22.5 Gensyn
Website: https://www.gensyn.ai/
One liner: Decentralised machine learning compute protocol to democratize AI devel-
opment.
Description: Gensyn integrates AI and blockchain technology to create a decentralized
global supercluster for machine learning, facilitating peer-to-peer access to computa-
tional resources and reducing reliance on major tech providers. This approach democ-

125 of 128
 AI x Crypto Primer

ratizes AI development by enabling global device participation in AI training processes,

promoting a more accessible and equitable distribution of computational power.

3.22.6 Golem Network

Website: https://golem.network/
One liner: Decentralized Computing for Everyone
Description: Golem Network offers a peer-to-peer marketplace that utilizes global,
unused computing power to support AI projects and complex computations at scale,
significantly reducing dependency on centralized cloud services. By integrating its na-
tive token, $GLM, for transactions, Golem facilitates a balanced exchange between
computing power providers and users, enhancing the AI x Crypto ecosystem.

3.22.7 Lumino AI
Website: https://www.luminolabs.ai/
One liner: Decentralized compute protocol for cost-effective AI model training.
Description: Lumino AI introduces a decentralized network for machine learning
model training, directly connecting AI developers with GPU compute providers to cut
training costs by 50-70% and leveraging cryptographic proofs for the traceability and
accountability of data and models. Its censorship-resistant framework promotes devel-
oper freedom, particularly in restrictive regimes, and diminishes central control risks.

3.22.8 Network3
Website: https://network3.io/
One liner: Network3 builds an AI Layer2 helping AI developers worldwide train or
validate models in scale quickly, conveniently, and efficiently powered by Sei Network
Description: Network3 creates a specialized AI Layer2 to offer global AI developers
services for training and validating their AI models, integrating the functionalities of
AI and cryptocurrency technologies.

3.22.9 Nimble Composable AI

Website: https://nimble.technology/
One liner: Nimble Composable AI aims to democratize the exchange of intelligence
through a composable, permissionless AI protocol on the blockchain.
Description: Nimble Composable AI introduces a decentralized AI framework focused
on composability and interoperability within the blockchain ecosystem, notably in
DeFi, DePIN, and GameFi sectors. This framework allows for the combination and
reuse of AI models and data, facilitating efficient Ethereum transactions and offering

126 of 128
 Who's Who in AI x Crypto?

an open data economy, aiming to democratize AI application development and deploy-

ment by reducing dependencies on centralized tech powers.

3.22.10 Privasea
Website: https://www.privasea.ai/
One liner: Privasea: Revolutionizing data privacy and security through Fully Homo-
morphic Encryption (FHE) on the blockchain, powered by AI and incentivized with
$PRVA tokens.
Description: The Privasea AI Network integrates Fully Homomorphic Encryption
(FHE) with blockchain technology to enable secure, encrypted machine learning infer-
ences on private datasets, facilitating privacy-compliant data sharing and collaboration
within the AI sector. It utilizes a blend of Proof-of-Work and Proof-of-Stake mecha-
nisms, incentivized through $PRVA tokens, to ensure efficient and secure data process-
ing, aiming to bolster AI development without compromising data privacy.

3.22.11 VFchain
Website: https://ieeexplore.ieee.org/document/9321132
One liner: VFChain leverages blockchain to enhance federated learning for secure and
efficient AI model training over distributed networks.
Description: VFChain enhances federated learning’s security by integrating
blockchain for a transparent, verifiable, and efficient AI model training across decentral-
ized networks. It employs a blockchain-recorded committee for model aggregation and
authenticated data structures to improve proof search and secure committee changes,
demonstrating its effectiveness through experiments with deep learning models on real-
world datasets.

3.22.12 Zama
Website: https://www.zama.ai/
One liner: Zama offers Fully Homomorphic Encryption (FHE) solutions for blockchain
and AI, prioritizing data privacy and security.
Description: Zama, an open-source cryptography firm, merges blockchain technology
with artificial intelligence using Fully Homomorphic Encryption (FHE) to enable the
processing of data in its encrypted state, enhancing privacy and security across online
sectors such as cloud services and blockchain smart contracts. Their development of
advanced tools like TFHE-RS and CONCRETE ML facilitates privacy-preserving ma-
chine learning and confidential smart contracts, establishing their role in advancing
FHE mainstream adoption.

3.23 Stealth

127 of 128
 AI x Crypto Primer

3.23.1 Sentient
Website: https://sentient.foundation
One Liner: Sentient is a stealth Crypto x AI project. They are hosting the
Decentralized AGI Summit @ ETH Denver.

Description: 🤫

128 of 128