I regularly hear from people that my post-training book/writing has helped them get a job at labs by focusing on material that develops intuitions for how training methods work. It's very validating, because this type of "idea/math heavy" writing doesn't really sell many books. It's rewarding the right people who put the effort into learning. I'm happy I made the time for it.

New talk on the open model economy. I wanted to break down the case of where open model builders should steer the economics of the industry, if we want to avoid being swallowed by OpenAI and Anthropic in due time. This rests fully on being able to drive the cost of inference down dramatically on specialized, repetitive tasks, rather than trying to compete on cost with the frontier labs in any general agent. A core piece of it is how the open model economy is very fragmented, leaving an interesting puzzle we need to solve to get revenue from applications to feed back into the next model cycle. This is a very near-term problem as closed labs scale rapidly. This talk investigates: Are people using open models? How close are open models to the frontier? Mapping the open model economy How open models win This is a practice recording of my talk "How Open Models Win" for Fireworks Forge 2026 in San Francisco (3 November 2026), which will have improvements over this one.

My first open model lecture, covering all the many pieces that build the ecosystem, and the open questions/debates determining where it goes next. How does distillation fit in? How do we manage risks? How do we see access to models changing in the future? I'm excited to revisit these questions about every 6 months, to give a state of the union on the field. Chapters 00:00 Introduction 01:36 What is an open model? 08:47 Why should we care about open models? 18:52 How competitive are the best open models today? 25:28 How are open models built? The role of distillation 34:52 Are people actually using open models? 41:47 What comes next with open models? 47:32 Conclusions What questions did I not answer that you have?

Interconnects Podcast #20#: We got together to debate hard questions and messy realities after a crazy, somewhat intimidating, week in the AI world: - Why did Kimi K3.1 Pro have a mixed reaction with such incredible benchmarks? Will it be adopted? - Were we wrong about the risks of open-weight models and cybersecurity? - Why do cyber experts discuss recent events so differently than AI experts? - What comes next in the few months? What will change our minds? Chapters: 00:00 Kimi K3.1 Reaction 08:15 Kimi’s open-weight cyber principles 11:23 You can hack everything era 15:54 PHP2Shell, LiteLLM, and the new internet 24:00 Abliterated APIs, safeguards, and trusted access 31:41 Anthropic not hiring juniors employees 36:16 Florian’s Muse addiction 45:30 Wrap-up More of these soon!

Finally got to put together a complete deck on the state of open models. What they are, licenses, how competitive they are, signs people are using them, distillation's impact, and how we balance the risks and benefits of openness! Will record a video lecture of these soon, and hopefully re-do them a few times a year.

Tomorrow at 9 AM CT: Join our AI Book Club author chat with @natolambert on Reinforcement Learning from Human Feedback 📖 We'll talk RLHF, why every lab is scaling RL, and this week's wild open-model releases. Come even if you're mid-book. RSVP 👇

Closed frontier LLMs are dangerous In most post-training reasoning efforts, a pre-trained base model trains on its own chains of thought (CoTs) when a final-answer verifier says the answer is correct. This is the basic principle of RL methods based on outcome rewards, used for STEM reasoning, coding tasks and even ethical alignment. Different algorithms have weigh the CoTs differently, but they all reinforce CoTs. As we've shown, training on one's own data, while not knowing how to distinguish it from someone else's data, causes delusions. See Everyone knows the different between something they did and something someone else did. This is essential for safety. If I write "Kevin is unfriendly" on a napkin, go for a walk, and then read the message on the napkin not knowing who wrote it, then my belief that "Kevin is unfriendly" is reinforced. These delusions are essentially confabulations, over-confidence that results from confusing P(o | a) with P(o | do(a)), where o is an outcome and a the action (CoT tokens). Unlike in past work, here I'm alerting everyone to this danger when training models to think, a typical step in the process of building agents. When the verifier supplies one bit about the answer, algorithms like ReST or GRPO spread that bit over every token of the CoT trace, so a reasoning step that is wrong but happens to precede a correct answer (a false lemma, a shortcut that only works on the training distribution, a sycophantic or test-gaming rationale, a subtle bias) is reinforced. Traits that are merely neutral are sharpened by mode-seeking self-distillation, and traits that are spuriously correlated with success grow exponentially. The damage may be invisible to the researchers and metrics, and in-distribution accuracy rises, but it shows up off-distribution. There has been some work on this. For example, Appendix G of the STaR paper touches on this potential bias amplification but only on the surface. Some researchers have proposed process rewards as the solution However, self-delusion by mistaking ones actions with observations produced by a third party highlights that this problem can be much more serious. Something subtle, invisible to the metrics and to the researchers, could lurk in the synthetic data. When deploying the AI out of distribution, that something could be fatal. Don't trust RL, RLHF, etc, blindly unless you really understand how the verification has been done. Corollary, don't trust any of the closed frontier models.

Really important update from China's ministry of commerce and ministry of science and tech (h/t @Nectar_Gan), not because it signals that the government's honeymoon period with open weight AI is over, far from it. But because it further confirms that as China's working on building out a more robust AI risk management regime (sth discussed in this year's Five Year Plan, see photo below, & by Xi at the World of AI Conference), it is singling out high-risk open models for more scrutiny. It appears that if the risks/ capabilities demand, a filing regime or prohibition may be considered, but this will be narrowly enforced and affect only a small number of products & services. So, what hasn't changed: Open model releases at home and abroad remain a welcomed development by the government. What has been clarified: On the open weight AI front, expect China to step up capability/risk-based governance to meet domestic needs, global demands. Here are the relevant discussions and sentences (I'll let X's translate function do its job): 1. 此次修订亦是技术贸易管理的**常态化⼯作**，依法依规进⾏，不针对任何特定国家、地区或实体。 2. 近年来，⼈⼯智能、⽣物医药和光伏相关技术在敏感领域的应⽤不断拓展，各主要经济体均实施了不同程度的出⼝管控。 3. 中国参考国际通⾏做法，依法实施必要的管理，既**防范技术失控和不当扩散⻛险，也保障正常国际技术交流合作**，推动科技向上向善、安全可控、造福⼈类。 4. 中国政府⼀贯⽀持企业依法依规开展互利共赢的跨国经营与国际技术合作，**有关政策⽴场没有改变**。 5. 此次拟将⼈⼯智能⼤模型、智能体、⼒集群、信息处理相关23项关键技术增列⾄《⽬录》，主要有以下考虑：... 此次拟列管的**⼈⼯智能⾃主驱动恶意代码⽣成与变异技术等，可被⽤于攻击国家关键信息基础设施、国防军⼯、核 ⼼政务等敏感系统，或⽤于绕过国家安全防御体系**。 6. ⿎励开源开放、合作共享。... **为⽀持构建⼈⼯智能开源创新⽣态体系，对⼈⼯智能类技术开源、中国企业在境外⾃建⾃营算⼒中⼼、赴境外开展芯⽚流⽚制造、利⽤境外资源开展联合研发、参与国际标准制定等情形的技术出，适⽤⾃动许可等便利措施，基本实现“即报即出”** ；对已通过开源、发表学术论⽂等⽅式进⼊公共领域的相关技术，拟不纳⼊出⼝管理范围，为⼈⼯智能技术国际交流合作创造良好环。 7. 需要特别指出的是，中⽅主张在相互尊重、平等互利基础上，与各国开展⼈⼯智能对话合作，**共同应对⼈⼯智能技术扩散和失控⻛险**。 8. **有关措施正式实施后，绝⼤部分⼈⼯智能技术仍可以⾃由出⼝。** 对于纳⼊限制类的⼈⼯智能技术，开展海外本地化部署，属于技术出⼝管理范围；**如果部署在海外⾃有云等⻛险相对可控的情形下，可适⽤⾃动许可等便利措施。** 对于企业利⽤限制类⼤模型对外提供API调⽤、APP/⽹⻚交互、SaaS应⽤、本地端侧轻量化部署等情形，不属于⼤模型技术出⼝管理范围。 (highlights are mine) Overall, this doc does not strike me as a trade bargaining chip as much as primarily an AI security play & secondarily a soft power play (i.e. responding to global concerns over risks from frontier open models). This development/ document also chimes with my predictions back in July & my observations last month:

Seems very important, as tons of people are tweeting about the end of bigger open models from China due to this from MOCOM, but there's a statement intending to make it easier for open source technology. Much hinges on implementation and how "near automatic" approval would work.

@xlr8harder Its got all the restrictions across all license eras 1. Built with muse branding nonsense that no one followed 2. Revenue gate like kimi/glm (reasonable) 3. Restricted domains / authorized use policy 4. All of the above applies to models trained on ai outputs (maybe)

Guys this is like my superbowl I waited so long for US open weight models to be relevant again. Before today the US models were still behind Chinese models from as early as Feb. 2026. Still more work to do! This cut the gap by at least half, in time (say 3 months)

If it performs roughly what like the scores indicate, adoption of this model is going to go gangbusters. First open-weigh model in the US with near frontier agentic capabilities (GLM 5.2 + Kimi K3 was that step change from China). US models get crazy adoption.

We've been monitoring a decline in HuggingFace downloads across the top open LLMs we track for interconnects. It looks like there was a slight change in counting of downloads, as we saw a ~30% download rate drop for many top models, but this is continuing over time. What else do people think it may be?