Daniel Minarik

What is Mixture-of-Recursions (MoR), and why is it different from the much-talked-about Mixture-of-Experts (MoE)? Where MoE scales width by activating only a subset of large expert modules per token, MoR scales depth by letting each token decide how many times it should re-enter a shared Transformer block. In other words: MoE distributes computation across experts, MoR recycles computation across recursions. The result is a model that can “think harder” only where needed: trivial tokens get shallow processing, complex ones receive more passes. This token-wise routing, combined with smart key–value caching, yields smaller models with better perplexity, higher throughput, and lower latency—up to ~2× improvements reported in recent studies. Unlike static deep stacks, MoR offers dynamic depth per token. Unlike MoE, it avoids ballooning parameter counts, making it especially appealing for edge deployments, enterprise inference at scale, and multimodal tasks where compute budgets are tight. Early experiments in both language and vision confirm its generality. Open questions remain: Will the gains hold at tens of billions of parameters? How stable will routing be in real production pipelines? But strategically, MoR reframes the race: not how many parameters we can afford, but how much useful depth per token per joule. #AIResearch #MixtureOfRecursions #AdaptiveComputation #TransformerArchitecture #AI #FutureOfIntelligence #Banking https://lnkd.in/dS6XYjxX

35

3 Comments

Model Context Protocol (or 𝑴𝑪𝑷), which was released by Anthropic in late 2024 is all rage now ... it is an open standard that enables Agentic LLM apps to connect and integrate with various data sources (both external and internal) as well as tools in a consistent way ... Here is my Python based hands-on article that demonstrates how one can leverage 𝑴𝑪𝑷 in the LLM apps ... uses the 𝑶𝒍𝒍𝒂𝒎𝒂 platform and the 𝑳𝒂𝒏𝒈𝑪𝒉𝒂𝒊𝒏 framework with the 𝑰𝑩𝑴 𝑮𝒓𝒂𝒏𝒊𝒕𝒆 3.1 model ... Here is the link to my article: https://lnkd.in/eK-E75_4 ... #ollama #llm #mcp #langchain #python

35

Google Cloud’s new Gemini Cloud Assist Investigations is one of the most exciting things I’ve seen lately. I absolutely loved the demo by Abirami Sukumaran, Ph.D — especially the part on Multi-Agent Systems and AI-powered debugging. It felt like a real step forward in how we manage and understand distributed systems. Key takeaways: • Contextual investigations triggered from logs, errors, or Cloud Hub • AI-generated observations and hypotheses for faster root cause analysis • Smart, actionable recommendations to resolve incidents Still in private preview, but definitely worth a look: https://lnkd.in/d4Q68tVt

73

4 Comments

🚀 LangGraph. LangChain. Autogen. CrewAI. n8n. Make.com. At first glance, they look similar. In reality, they solve fundamentally different problems in the AI orchestration stack. 1️⃣ LangGraph – Orchestrating Memory & Process The operating system for AI workflows. Handles stateful reasoning, long-running tasks, branching logic, retries. Ideal for enterprise-grade, persistent applications. 2️⃣ LangChain – The Composable Toolkit The “Swiss Army Knife” of AI prototyping. Quickly connects LLMs to tools, APIs, and data sources. Best for chatbots, RAG, and demos. 3️⃣ Autogen – Multi-Agent Collaboration Microsoft’s research-driven framework. Multiple AIs debate, negotiate, and converge on solutions. Great for experimentation with AI-as-a-team. 4️⃣ CrewAI – Deployable Micro-Agents A pragmatic take on multi-agent orchestration. Spin up specialised agents with defined roles — scraper, analyst, writer. Optimised for speed and production. 5️⃣ Make.com – No-Code AI Plumbing Drag-and-drop automation between AI and apps like Slack, HubSpot, Gmail, Notion. Empowers non-technical teams to scale AI. 6️⃣ n8n – The Open-Source Automator Self-hosted and extensible. Connect AI to databases, APIs, and enterprise workflows without vendor lock-in. 💡 Bottom Line LangGraph → Complex, stateful workflows LangChain → Fast prototypes & integrations Autogen → Research-driven multi-agent collaboration CrewAI → Deployable agent teams Make.com → SaaS drag-and-drop automation n8n → Open-source pipelines for enterprises 🌍 My Global Mission — AI for ALL This is why I launched AI for ALL — a global initiative to make AI a public good: free, open, ethical, and inclusive. Through AI for ALL, we are curating open-source repositories, practical workflows, and community knowledge — lowering barriers and empowering youth, professionals, and women in tech worldwide. 🔗 Learn and participate: www.drdebashisdutta.com 💻 Explore the repos: github.com/drdebashisdutta 📩 Subscribe to my newsletter for AI insights & resources: 📱 For free mentoring and guidance, you can reach me directly: Mobile/WhatsApp: +966 550409099 Email: [email protected] 👉 Question for you: Which of these frameworks is powering your AI journey right now? ♻️ Repost to clarify the differences for your network. 🔔 Follow me for AI orchestration insights and updates on my AI for ALL initiative. #ArtificialIntelligence #AIForAll #LangGraph #LangChain #Autogen #CrewAI #n8n #MakeDotCom #AIWorkflows #AIOrchestration #OpenSourceAI #AgenticAI #EthicalAI #FutureOfWork #DigitalTransformation Thanks Prem N. for the amazing visual.

24

2 Comments

𝐓𝐡𝐞 𝐬𝐞𝐜𝐨𝐧𝐝 𝐚𝐩𝐩𝐥𝐢𝐜𝐚𝐭𝐢𝐨𝐧 𝐨𝐟 𝐪𝐮𝐚𝐧𝐭𝐮𝐦 𝐜𝐨𝐦𝐩𝐮𝐭𝐢𝐧𝐠 𝐢𝐧 𝐟𝐢𝐧𝐚𝐧𝐜𝐞: IBM 𝐚𝐧𝐝 Vanguard, 𝐭𝐡𝐞 𝐰𝐨𝐫𝐥𝐝'𝐬 𝐬𝐞𝐜𝐨𝐧𝐝-𝐥𝐚𝐫𝐠𝐞𝐬𝐭 𝐚𝐬𝐬𝐞𝐭 𝐟𝐮𝐧𝐝, 𝐮𝐬𝐞𝐝 𝐪𝐮𝐚𝐧𝐭𝐮𝐦 𝐜𝐨𝐦𝐩𝐮𝐭𝐢𝐧𝐠 𝐭𝐨 𝐬𝐨𝐥𝐯𝐞 𝐚 𝐫𝐞𝐚𝐥-𝐰𝐨𝐫𝐥𝐝 𝐩𝐨𝐫𝐭𝐟𝐨𝐥𝐢𝐨 𝐨𝐩𝐭𝐢𝐦𝐢𝐳𝐚𝐭𝐢𝐨𝐧 𝐩𝐫𝐨𝐛𝐥𝐞𝐦. ➡️ The IBM-Vanguard team explore how sampling-based variational quantum algorithms, #VQA may be able to address this challenge. These methods combine the strengths of quantum sampling with classical optimization and post-processing, enabling a new class of #heuristics algorithms to explore complex solution landscapes more efficiently than traditional methods. ➡️ The team applied the Conditional Value at Risk-based Variational Quantum Algorithm, #CVaRVQA method to an Exchange Traded Fund (#ETF) portfolio construction problem; The results were benchmarked against the classical optimization solver #CPLEX, some important informations: 1️⃣ the quantum-classical workflow consistently outperformed a purely classical local search approach, especially as problem size increased. 2️⃣ There is 3 important dimensions for bonds #risk, #industry and #duration as a optimization objectives. 3️⃣ They used 109 #qubits of the 133 available on an IBM Quantum #Heron r1 processor, executing circuits with up to 4,200 gates. The quantum samples were then refined using a classical local search algorithm to further improve solution quality. IBM Quantum #VQA #Quantum_optimization

45

MCP (Model Context Protocol) has emerged as the standard way for LLMs to connect to external tools, servers, and data sources. The protocol defines how those connections work - a shared schema for exposing resources and tool calls. That’s why it’s being adopted across OpenAI, Anthropic, LangGraph, CrewAI, Composio, and more. But here’s the reality: MCP only gives you the spec. If you’re building an app, you still need to: ▪️ tell your copilot which MCP servers to talk to ▪️ manage how responses flow into conversations ▪️ and build UI to make tool calls usable It works, but it’s a lot of wiring. That’s the gap CopilotKit🪁 (AG-UI)’s MCP Client closes. Check here: https://lnkd.in/dF9R7PrZ Point your copilot at an MCP server, and it automatically: ✔️ discovers tools and resources at runtime ✔️ makes them callable by the LLM like native actions ✔️ streams results directly into the conversation ✔️ renders them with built-in UI components Example: instead of hand-building a dashboard for SQL queries, your copilot can connect to a database MCP server, let the LLM generate queries, and then render the results in-app with a ready-made table component. No extra plumbing required. 👉 MCP defined the protocol. 👉 CopilotKit MCP Client makes it practical in your app. Setup takes one command: npx copilotkit@latest init -m MCP

420

85 Comments

As we step into 2025, let’s remember that the period 1990-2000 is often referred to as an AI winter, marked by high expectations and setbacks in development. When new technology is hyped, inflated expectations can lead to disillusionment, raising concerns about another AI winter. Will this happen again? I don’t think so. AI today delivers immense value across industries—coding, healthcare, customer support, and more—with innovation moving at an unprecedented pace. While challenges like startup failures and unmet expectations may arise, the capital, talent, and proven utility of AI ensure its continued progress. At worst, if there has to be a winter, it will be more like an “Albanian seaside winter”—mild and full of potential. Let’s embrace the opportunities ahead! #AI #RaiffeisenBankAlbania #Tech #Future https://lnkd.in/dmvV6wqK

78

3 Comments

Good panel discussion involving two ex regulators at #DMS25 on the topic - How to build a GenAI data strategy and operations model to deliver business value. The in-room Slido poll stated that the most critical issue in achieving positive ROI was “poor data quality” (at 44%). As a company that provides a platform focusing on DQ and matching @Datactics is seeing numerous questions in this area. The panel discussion was much more wide ranging and nuanced, however, and clearly just buying a DQ tool and implementing a remediation strategy isn’t a magic bullet. Some interesting insights … Steve Green (Sumitomo) - don’t be afraid to mix quick wins for the business alongside strategic projects Naresh Rajah - it’s always an AI AND data strategy so get a senior stakeholder from business involved Peter Jackson (Schroders) - engage business transform team early and don’t let perfect be an enemy of the good Ruby Paranantham (Colombia TI) - always seek to create value downstream in end user applications like reg reporting or predictive analytics #DataManagementSummit #DMS25 #DMSLondon #DataQuality #RegTech

30