Debu Sinha

**** Limits of LLMs: Scaling Laws and the Limits of Model-Centric Thinking**** Scaling laws have become one of the most influential ideas in modern deep learning. They offer a simple, elegant promise: increase model size, data, and compute, and performance will improve in a predictable way. This idea has guided billions of dollars of investment and shaped how large language models (LLMs) are built. Yet, while scaling laws are mathematically precise, their interpretation is often overstretched. They explain how loss behaves during training. They do not explain how intelligence behaves during use. The blog below examines what scaling laws genuinely capture and, more importantly, what they fundamentally miss. https://lnkd.in/gTj6kyuD #LLM #GenerativeAI

14

The Era of AI Agents is Here ​I am proud to announce the release of my new book, "The Algorithmic Loop: Modern Agentic Design Patterns," co-authored with Saumya Singh. ​For the past year, I’ve been deeply focused on a single question: How do we move from AI that simply answers questions to AI that autonomously solves problems? In this book, I dive into the architectural frameworks that make "Agentic AI" possible. We move beyond simple prompting and explore the core cycle that defines the next generation of technology: Perception, Memory, Action, and Feedback. ​Why I wrote this book: ​To bridge the gap between theoretical AI and practical, agentic implementation. ​To define the design patterns that allow agents to interact meaningfully with their environments. ​To provide a roadmap for developers and leaders ready to build the "Autonomous Future." ​Co-authoring this with Saumya Singh has been an incredible journey of refining these complex ideas into a guide that is both technical and actionable. ​If you are ready to master the Algorithmic Loop, you can find our work on Amazon now. ​👉 Get your copy here: https://lnkd.in/gZSAm7c3 ​#AgenticAI #AIArchitecture #TheAlgorithmicLoop #TechLeadership #GyanPrakash #Innovation

55

7 Comments

🤖RLAD Training LLMs to Discover Abstractions for Solving Reasoning Problems. 🔍 Motivation: How do we move large language models (LLMs) beyond pattern matching—toward algorithmic, abstract reasoning on hard problems? Most current RL post-training approaches create longer chains of thought, but they often drift into verbose or degenerate exploration, missing true reasoning depth. 📄 Solution: RLAD introduces 'reasoning abstractions'—concise, natural language descriptions of procedural and factual knowledge that guide models in learning successful reasoning strategies. The approach trains two LLMs: one to generate abstractions and another to solve problems using those abstractions, via a novel RL setup. Crucially, abstractions act as high-level subgoals or skills, helping models explore diverse solution paths and generalize to harder tasks. 🔑 Results: RLAD achieves up to 44% improvement on math reasoning benchmarks (AIME 2025, DeepScaleR Hard, AMC 2023) over state-of-the-art long chain-of-thought RL methods. Moreover, the benefits scale with more compute allocated to abstraction generation versus solution sampling—a key insight for practical deployment. 💡 Implications: Abstraction-driven training broadens the reasoning abilities of LLMs, unlocking new horizons for model interpretability, strategy diversity, and generalization across complex domains. This direction is especially relevant as scaling chain-of-thought length yields diminishing returns. (Attached: Original research PDF)

13

1 Comment

Researchers introduce a novel LLM architecture that scales test-time compute through implicit reasoning in latent space via a recurrent block, eliminating the need for word-based chain-of-thought or specialized training data. https://lnkd.in/gF4msQC5 #MachineLearning #AI #LLM #DeepLearning #AgenticAI

2

Anthropic's technical blog is a great resource for production-ready AI systems. I went through some of their best articles lately (imo) -- They're really good at explaining the thought process behind anticipating challenges in a prod scenario and designing robust systems. I used Miskies AI to create visual and hands-on explanations for all articles. - Equipping agents for the real world with Agent Skills Claude is powerful, but real work requires procedural knowledge and organizational context. Introducing Agent Skills, a new way to build specialized agents using files and folders. https://lnkd.in/gCrFPFFV - Effective context engineering for AI agents Context is a critical but finite resource for AI agents. In this post, they explore strategies for effectively curating and managing the context that powers them. https://lnkd.in/gbetNUVW - Writing effective tools for agents — with agents Agents are only as effective as the tools we give them. They share how to write high-quality tools and evaluations, and how you can boost performance by using Claude to optimize its tools for itself. https://lnkd.in/gGzpctYF - How we built our multi-agent research system Research feature uses multiple Claude agents to explore complex topics more effectively. They share the engineering challenges and the lessons learned from building this system. https://lnkd.in/gw4C4r-9 Links to the original articles are also in the comments. ♻️ Share with someone building production-ready AI systems :) #AI #LLMs #GenAI

31

3 Comments

Semantic search (by intention) is an under appreciated approach to code search, especially in a world of non determinism where accruing tech debt through different named but identical functional intent code is very common. https://lnkd.in/erTa9JgD To date I think only cursors in house model has been trained on a search tool with this kind of capability.

2

1 Comment

Great time at the Liquidity Summit two weeks ago. I had the chance to sit down with the NYSE right after Bill Ackman to talk AI investment trends, compute demand, and what's driving the next phase of growth. Here's the conversation.

93

1 Comment

We're sharing code -- an autonomous ML research system #agents We're sharing an early preview of Applied AI’s #senpai - autoresearch for physical AI, inspired by Karpathy’s latest hit (Self ExperimentatioN for Physical AI - senpai). Aerodynamics Modelling 🛫 The task is to train a model to predict the airflow over pairs of airfoils placed close together, e.g. 2 spoilers on a car, based on the TandemFoilSet paper (dataset linked below). We're training deep learning models for these types of Computational Fluid Dynamics (CFD) tasks. Agents 🤖 1 Claude Advisor (Jürgen) is tasked with coming up with research ideas to try and is given the ability to launch multiple Student Claudes (each with their own GPUs) to implement the idea, run the experiment and report results. Metrics are logged to Weights & Biases and all agents are given the wandb Skill. See advisor.md for the full prompt. Communication #Github PRs are used to communicate experiment instructions and results between the Advisor and Student - messages in a bottle. Positive experiments are merged, unsuccessful are closed or asked for more follow up experiments. The Advisor is given a Skill to download experiment summaries and details to .md files (as agents tend to do better at exploring files instead of hitting apis). Code: https://lnkd.in/eJE2_viS Dataset paper: https://lnkd.in/eYJ-6zxb Incredibly lucky to work with the experts that threw this together in a sprint 'just for the fun of participating in the autoresearch explosion online' -- Morgan McGuire & Thomas Capelle . Absolute masters of the craft. Its fairly new dataset so shouldn’t be easy for an LLM to search the web for solutions (making it an even better test case). This is still an early implementation and we’ve tried to keep it fairly simple. This is a concept-like project, so try at your own peril! #ai #autoresearch #appliedai #agents #engineering

134

The AI competition isn’t about Model IQ. It’s about who owns the coordination layer. The landscape is evolving beyond “whose model is smarter” toward a more commercial question: who controls the integration path? Three distinct power centers are emerging: 1. Surface Owners: Microsoft and Google are embedding AI into the OS and the procurement flow. Convenience becomes switching friction. 2. Capability Leaders: Where model quality and cost/performance matter. Anthropic lives here; OpenAI spans capability + surface. 3. The Coordination Layer: The connective tissue - APIs, permissions, connectors. This is where implementation overhead hides and where switching costs accumulate. The Strategy Split: - OpenAI is going vertical by building proprietary coordination (Assistants/Responses) + marketplace motion to become a Surface Owner. - Anthropic is going horizontal by pushing open coordination (MCP + Agent Skills) to commoditize the connection. If coordination stays open (standardized), the “integration moat” shrinks. Then the moat shifts back to what actually matters: unit economics + model quality (latency, reliability, overhead). Rule of thumb: Closed coordination → Surface wins. Open coordination → Capability can win.

9