HOW CLAUDE WORKS?
Peeking Inside Claude’s Cognition
Bhavishya Pandit
Introduction
Ever wondered what an AI is really thinking?
LLMs like Claude are now the foundation of contemporary AI
solutions, driving a wide range of applications from chatbots to
virtual tutors, and much more.
However, for the most part, the mechanisms by which they do so are
unknown. The question is, do we really understand AI?
In this post we will peek inside the mind of a large language model.
The results will surprise you.
Bhavishya Pandit
Anthropic’s Mission
Anthropic doesn’t just want powerful AI, but an understandable AI.
Source
One way they do this is by tracing how Claude forms its answers,
step by step, to turn it from black box into a glass box.
They trace :
how features light up
how information flows
how Claude “decides”
How do they do it? Let’s zoom into the building blocks with their “AI
microscope”.
Bhavishya Pandit
Claude Thinks in Features & Circuits
At its core, Claude thinks not in words but features.
Just like the lego bricks, alone these features mean little. But when
connected, they form circuits that map out Claude’s thought process
step-by-step.
Then attribution graphs spotlight which features power each answer.
Source
This attribution graph shows it performs multi-hop reasoning—linking
“Dallas” to “Texas,” then combining “Texas” + “capital” to reach “Austin.”
These layered steps reveal how Claude builds answers, using both
direct and indirect pathways, through intermediate concept activation,
not just word matching.
Bhavishya Pandit
Claude Has a ‘Language of Thought’
Claude doesn’t just react. It plans.
For example :
Source
Researchers found Claude uses shared multilingual circuits to answer
antonym prompts across languages.
It first activates same language-independent “antonym” features,
then combines them with language-specific output features (like “big”
in Chinese).
When asked to write a haiku, Claude strategizes :
Theme
Rhythm
Structure
Then responds with a haiku. It maps out this structure in its
“thinking space” , and then writes.
This reveals Claude’s ability to multi-step reasoning, much like
human planning.
Bhavishya Pandit
But, what if it Lies?
Claude often explains why it gave an answer. But is that explanation
always real?
Researchers found Claude sometimes explains its answers with
smooth, logical-sounding reasons, but these reasons don’t reflect
how it actually arrived at that answer.
This is called “decoupled reasoning”. It’s like a student guessing on a
test, then inventing a clever-sounding explanation afterward.
Source
Like, when asked “What is the capital of Australia?”, Claude may
answer “It’s Sydney”, stating a reason like “Sydney is the largest city
in Australia. It hosts key government buildings near the harbor. So,
Sydney is the capital.”
It’s wrong since the real capital is Canberra.
These hallucinated justifications can mislead users, even experts.
Bhavishya Pandit
Why Is It Challenging to Understand AI?
There’s no “Decode” button for AI, not yet.
To understand Claude’s decisions, researchers dig through:
Millions of activations
Layers of feature traces
Complex attention maps
It’s like reverse-engineering each thought, frame by frame.
Even with tools like “AI microscopes,” interpreting Claude’s mind
remains labor-intensive.
But Anthropic is building tools to make this detective work faster,and
someday, automate this.
Bhavishya Pandit
The Future of Transparent AI
Peeking inside Claude isn’t just fascinating, it’s a step toward safer,
fairer, and smarter AI.
This research is the foundation for how we can understand & shape AI.
Transparency is how we build trust, empower users, and create
inclusive AI systems for everyone.
It means moving from black-box decisions to collaborative intelligence.
This opens the door to AI that can be audited, corrected, and improved,
not just used.
Bhavishya Pandit
Model transparency
comes at the cost of
security?
Let me know your thoughts in the comments
Bhavishya Pandit
Stay Ahead with Our Tech Newsletter! 🚀
👉 Subscribe and join 1k+ leaders and professionals
🔗 https://bhavishyapandit9.substack.com/
Join our newsletter for:
Step-by-step guides to mastering complex topics
Industry trends & innovations delivered straight to your inbox
Actionable tips to enhance your skills and stay competitive
Insights on cutting-edge AI & software development
💡 Whether you're a developer, researcher, or tech enthusiast, this newsletter is your
shortcut to staying informed and ahead of the curve.
Bhavishya Pandit
Follow to stay updated on
Generative AI
LIKE COMMENT REPOST
Bhavishya Pandit
