Chip Huyen

Anthropic cuts off OpenAI’s access to Claude models ... Anthropic says OpenAI’s engineers used Claude’s coding tools ahead of GPT‑5’s launch .... a direct violation of their terms, which forbid using Claude to build competing services. They’ve now revoked access .. except for limited benchmarking and safety evals. OpenAI called it “industry standard.” Anthropic didn’t agree.

166

32 Comments

Last week, China barred its major tech companies from buying Nvidia chips. This move received only modest attention in the media, but has implications beyond what’s widely appreciated. Specifically, it signals that China has progressed sufficiently in semiconductors to break away from dependence on advanced chips designed in the U.S., the vast majority of which are manufactured in Taiwan. It also highlights the U.S. vulnerability to possible disruptions in Taiwan at a moment when China is becoming less vulnerable. After the U.S. started restricting AI chip sales to China, China dramatically ramped up its semiconductor research and investment to move toward self-sufficiency. These efforts are starting to bear fruit, and China’s willingness to cut off Nvidia is a strong sign of its faith in its domestic capabilities. For example, the new DeepSeek-R1-Safe model was trained on 1000 Huawei Ascend chips. While individual Ascend chips are significantly less powerful than individual Nvidia or AMD chips, Huawei’s system-level design to orchestrate how a much larger number of chips work together seems to be paying off. For example, Huawei’s CloudMatrix 384 system of 384 chips aims to compete with Nvidia’s GB200, which uses 72 higher-capability chips. Today, U.S. access to advanced semiconductors is heavily dependent on Taiwan’s TSMC, which manufactures the vast majority of advanced chips. Unfortunately, U.S. efforts to ramp up domestic semiconductor manufacturing have been slow. I am encouraged that one fab at the TSMC Arizona facility is operating, but issues of workforce training, culture, licensing and permitting, and the supply chain are still being addressed, and there is still a long road ahead for the U.S. facility to be a viable substitute for Taiwan manufacturing. If China gains independence from Taiwan manufacturing significantly faster than the U.S., this would leave the U.S. much more vulnerable to possible disruptions in Taiwan, whether through natural disasters or man-made events. If manufacturing in Taiwan is disrupted for any reason and Chinese companies end up accounting for a large fraction of global semiconductor manufacturing capabilities, that would also help China gain tremendous geopolitical influence. Despite occasional moments of heightened tensions and large-scale military exercises, Taiwan has been mostly peaceful since the 1960s. This peace has helped the people of Taiwan to prosper and allowed AI to make tremendous advances, built on top of chips made by TSMC. I hope we will find a path to maintaining peace for many decades more. But hope is not a plan. In addition to working to ensure peace, practical work lies ahead to multi-source, build more fabs in more nations, and enhance the resilience of the semiconductor supply chain. Dependence on any single manufacturer invites shortages, price spikes, and stalled innovation the moment something goes sideways. [Original text: https://lnkd.in/gxR48TK8 ]

5,654

236 Comments

In agentic AI, experience matters more than hype. I’ve seen firsthand how the conversation around AI often focuses on its promise rather than the practicalities of deployment. Richard Li and I explore the implications of agentic hype in a new blog. While agentic AI is compelling, getting it to work at scale is a fundamentally harder problem than many realize. It's not about clever prototypes or your development framework; it's about building hardened, distributed systems that are durable, observable, and scalable. Success with agentic AI won't be achieved through simple orchestration. It depends on a deep understanding of infrastructure and a commitment to engineering excellence. We need to move beyond the hype and focus on building a foundation of stability and trust. Are you ready for your agentic transformation? We'll help you move AI from concept to reality. #AI #AgenticAI #Innovation #Technology #Leadership https://lnkd.in/gbvSNuZy

52

10 Comments

A fun project: I built the LLM interface I wished existed. interface0 is for power users: cross-provider chat & memory, multi-model synthesis, sharing & team features, granular context engineering, template prompts, forking, prompt enhancing, & more. Use models from OpenAI/ChatGPT, Anthropic/Claude, Gemini, Grok, and more, all in one place. You can query them all at once with a single prompt and get only the best response, or pick and choose — and either way, interface0 maintains memory across all of them. You can use template prompts with variables for repeated workflows; share chats with friends or teammates and collaboratively respond; create team-wide prompt libraries and knowledge bases; easily swap the persona of the agent responding; make phone calls and send and receive emails; and much more. I've been hacking on this for the last couple months and will continue to build it out. We're in the age of "personal software" and it's been a lot of fun to create a product that fits my own needs so well. Link in comment if you want to check it out! Here's a quick video walkthrough:

298

121 Comments

I don't just code or train AI models, I teach them to think. Everyone's trying to retrain LLMs. I found a better way. Here's my framework for TOXO (Context-Augmented Language Models): ❌ Traditional Approach: → Write complex code → Spend $50K+ on GPUs → Wait weeks for training → Pray it works ✅ TOXO Approach: → Write instructions in plain English → Upload your knowledge base → Give feedback when AI gets it wrong → Export a .toxo file that works with any blackbox LLMs(api) The magic? We don't retrain the model. We teach it HOW to think about your specific problem. Think of it like this: ❌ Retraining = Teaching a doctor to become a lawyer (expensive, slow) ✅ TOXO-tuning = Teaching a doctor to think like/w.r.t the patient when treating patients (fast, cheap) Result: 40% faster learning with 90% less compute. This is what happens when a broke student can't afford the "right" way to do things. Sometimes the limitation becomes the innovation. Who else is tired of expensive AI infrastructure? Drop your biggest AI frustration below 👇 #TOXO #AIInnovation #NoCodeAI #StudentFounder #ThinkDifferent

53

1 Comment

As someone building AI products, I’ve found that clarity around how agents really work is still missing in much of the industry conversation. This 30-point guide by Chip Huyen does an exceptional job of breaking down the fundamentals; from planning and tool use to reflection and evaluation. I’m sharing it here because I believe resources like this can raise the level of understanding in our space and the more we align on the foundations, the faster we can push AI forward responsibly.

42

AI governance doesn't mean being anti-AI One of the most common questions I'm asked is how I use AI in my work. Knowing how to effectively use the AI tools and applications we all have access to is becoming non-negotiable in 2025. And whether or not we're hitting a wall in terms of frontier AI model capability doesn't negate how transformative the current applications are, when used in the right way. I use AI every day to augment and optimise my work, especially my research, learning, content creation, and writing. So, to answer the question, below are my 5 top personal Dos and Don'ts for using AI ⤵️ The key takeaway is that attaining and deepening subject matter expertise and sharpening your critical thinking should be your strategic priority. However, augmenting this expertise and judgement with informed and intentional use of AI is both a tactical necessity and a winning formula. How I leverage AI 1. Review and critique of my work 2. Initial research on a topic 3. Editing and proof-reading 4. Ideation and inspiration 5. Data analysis and cleansing What I do myself 1. Write my content (especially the first draft!) 2. Determine my argument and position 3. Create infographics and visuals 4. Complex legal analysis 5. Read important papers and books These are purely my personal views and I would love to hear if you disagree with me or if you leverage AI for entirely different use cases and scenarios!

327

34 Comments

The narrative that bigger models are better models is being dismantled, and the latest research from NVIDIA is a powerful testament to this shift. Their new paper, "Small Language Models are the Future of Agentic AI" (https://lnkd.in/gFET26Sz), is a must-read. It compellingly argues that Small Language Models (SLMs) are the true future of scalable, efficient Agentic AI. The key takeaways align perfectly with our foundational thesis for Inference Labs: • Superior Efficiency: SLMs offer drastically lower latency and cost for specialized agentic tasks compared to monolithic LLMs. • Purpose-Built Design: Their smaller size makes them inherently more suitable for deterministic, repeatable processes, a bedrock of reliable agents. • Heterogeneous Architectures: The future is a collaborative system: a general-purpose LLM orchestrator directing a fleet of specialized, high-efficiency SLM agents. The paper is brilliant analysis of the software paradigm, however, it stops short of addressing a monumental opportunity. The critical next step is Formal Verification. What makes compact SLMs truly revolutionary isn't just their performance, it's their potential interoperability. Zero-knowledge cryptography is a key to verified, certifiable, and safe AI. The move from LLMs to SLMs isn't just an optimization. It's the fundamental enabler for hyperscaling AI as the edges of organizations start to blur.

225

9 Comments

Introducing the AI Engineer World's Fair multi-turn conversation benchmark. I wrote a post for the AI Tinkerers newsletter about: - why multi-turn is hard for LLMs, - the particular challenges of audio/voice, and - a benchmark I use to think about LLM performance for conversational agents. https://lnkd.in/ecEfvQBs

48

6 Comments

AI has shifted from prediction to generation. From Transformers in 2017 to ChatGPT’s launch in 2022, and now multimodal models, retrieval, and creative AI—the frontier is no longer labeling reality, but constructing with it. Read the full story 👇 🔗 https://lnkd.in/gNbKnzef #AI #GenerativeAI #MachineLearning #ChatGPT #Innovation #FutureOfAI

26

1 Comment

I've been toying with the idea of generating passive income out of running LLMs locally. I was wondering if I could just build something that is plug-n-play - you run a host at home and it runs around the clock generating income by serving AI compute. I built a thing and called it PeerLLM. PeerLLM is a decentralized AI network where anyone can participate in the new AI economy. Instead of AI taking jobs away, PeerLLM gives people a way to earn from what they already have. I built an architecture where: 0/ Consumers connect to the network for AI services. 1/ The network distributes requests across community hosts. 2/ Hosts can be servers, models, or data providers run locally, on the go, or in the cloud. This is more than a technical system. It’s a community-driven economy, where people pool resources and share in the value AI creates. In this video, I share the idea, the architecture, and a live demo of PeerLLM including the website, the network, and the host system in action. Watch here: https://lnkd.in/d_KbKJab #llms #gpt #ai

61

11 Comments

New model? Chunking trick? Retrieval hack? → Ship it and hope it helps. But AI systems are probabilistic, which makes “what’s better?” a deceptively hard question. Without the right evals, you’re just guessing. 👉 Enter Eval-Driven Development. By defining success up front and grounding every experiment in metrics, you turn: - Guesses → measurable signals - Hunches → confidence - Iterations → progress that compounds Robert Shelton (Applied AI Engineer, Redis) at 6th Annual MLOps World | GenAI Summit will share how moving beyond the vibe transforms the way we design and evaluate AI systems and why evals should be the backbone of modern AI development. The shift is simple but powerful: stop guessing, start measuring.

26

2 Comments

Anthropic released a new blog post called “Effective Context Engineering for AI Agents” that sheds light on how they implement certain things and how they make use of context engineering. In the blog, they explain the difference between context engineering and prompt engineering, talk about “context retrieval and agentic search”, and share some of their general conclusions.

28

Last week we announced Credo AI Labs, and today I'd like to highlight our first Lab: 𝐌𝐨𝐝𝐞𝐥 𝐓𝐫𝐮𝐬𝐭 𝐒𝐜𝐨𝐫𝐞𝐬: https://lnkd.in/gW7_sTVN This project is about understanding context-specific model acceptability based on the third-party benchmark ecosystem. It's free to explore and use - give it a go and let us know what you think! --------------------- Model Trust Scores is a continuously updated aggregation of third-party evaluations of foundation models. We bring together evaluations from a number of third-party evaluators, and aggregate them in a use-case-context sensitive way to give industry-specific leaderboards. Essentially it is a benchmark 𝑡𝑟𝑎𝑛𝑠𝑙𝑎𝑡𝑖𝑜𝑛 𝑒𝑛𝑔𝑖𝑛𝑒, built atop the benchmarking ecosystem. Thanks to: ➤ Vals AI and Artificial Analysis - two fantastic evaluators with prompt updates on a huge range of models ➤ AIR-Bench from Center for Research on Foundation Models at Stanford University, a safety measure on model refusal rates ➤ LiveBench - a longstanding capabilities benchmark ... and more! The main advancement is our AI-powered relevance scoring system, which figures out which benchmarks are actually important for a particular use case. From these relevance scores and all the benchmarks we calculate: ➤ Capability - Can it perform the task? ➤ Safety - Can you trust it to follow safety guidelines? ➤ Speed - Response latency ➤ Affordability - Aggregated input/output token count

36

2 Comments

LLMs can predict planetary motion. But they still do not know why planets move. A recent study from Harvard and MIT posed a powerful question: "Can today's AI models go beyond prediction and learn the laws that govern the universe?": https://lnkd.in/gT2ZDuRh To test this, researchers trained a transformer-based AI model on millions of simulated solar systems. The model learned to predict the trajectories of planets with remarkable accuracy. But when the researchers checked whether the model had recovered Newton’s laws of gravitation - the actual force equations it failed spectacularly. The model had no understanding of force, mass, or inverse-square laws. Instead, it relied on pattern recognition heuristics that looked right on paper, but broke completely when tested on new systems. Even more striking: well-known LLMs like Claude, Gemini, and OpenAI’s o3 model did the same. They got the "what" right, but the "why" utterly wrong. This is equivalent to the model saying "I know that sun will rise tomorrow, but not sure why". It shows that LLMs are not building world models. They are statistical simulators. They can extrapolate, but not explain. They can mimic Newton's predictions, but cannot rediscover Newton's insight. And this is exactly where Scientific Machine Learning (SciML) steps in. Instead of forcing neural networks to learn everything from scratch, SciML fuses physics and learning. It embeds domain knowledge into machine learning models, allowing them to generalize far beyond what data alone can offer. At Vizuara, we put this philosophy into action. We modeled Chandrasekhar’s white dwarf equation - a nonlinear differential equation describing the balance of quantum degeneracy pressure and gravity in collapsing stars. We did this with only 10% of the data. Using a Universal Differential Equation (UDE) framework, we infused known physics directly into the learning loop. The result was a model that not only fit the observed data, but generalized across unobserved regimes. Our work was accepted at the NeurIPS 2024 Workshop on Scientific Discovery and Machine Learning. This matters because the future of AI for science cannot rely on token prediction alone. It needs: → models that generalize with little data → systems that reason about causality, not just correlation → architectures that reflect the structure of the world LLMs are powerful. But for scientific discovery, they must be augmented with models that understand mechanisms. SciML offers that bridge This is the frontier we are working on at Vizuara. And we are just getting started. If you wish to work on SciML problems with us, check this out: https://lnkd.in/gRHFk5C8

671

31 Comments