My time at Ai2 / @ai2.bsky.social has come to an end. Ai2 is a wonderful place. The last 2.5+ years building Olmo, Tulu, and other projects will be one of the peaks of my entire career. www.interconnects.ai/p/farewell-ai2

Can closed labs can create margins with very high cost services in the long-term? I say yes, but only in a few niches like high end knowledge work. Meanwhile, the open model economy slowly grows to be bigger and control the long-tail of AI use. www.interconnects.ai/p/open-and-c...

The debate on if open or closed models win comes down to if there is disproportionate value to marginally better intelligence. The believers of this sit across from the open models will be good enough camp. Closed models will stay slightly smarter. Open models will be cheaper.

Very useful post from @jjding99 on Anthropic's China Dogma chinai.substack.com/p/anthropics...

Gemma 4 adoption numbers outpacing Qwen 3.5/3.6 for the same sized models is a big shift in the international balance of influence via open models.

Some ideas for what comes next, May 2026 Gemini Flash 3.5, Mythos, open-closed balance, America's open-source surge, emerging power struggles and more. www.interconnects.ai/p/some-ideas...

The geopolitical angle of the AI race is only going to keep accelerating. More domestic talent travel restrictions from the Chinese government. I expect more and more things like this, AI is getting embedded in the core of existing power structures. www.bloomberg.com/news/article...

I heard people needed clarification that my book was a post-training book posttrainingbook.com

On-policy distillation is on track to be a lasting method in post-training. The list of areas would be: Instruction tuning (SFT/IFT) RLHF Direct Preference Optimization (DPO et al) RLVR On-policy Distillation (OPD) New classes of methods are rare! Excited to play.

Being out of SF has lowered my information proximity but with the big upside of giving me space to cultivate my own beliefs and values around ai. We need more people zagging in AI, the monoculture just helps the incumbents win at this point.

Latest open artifacts (#21): Open model bonanza! Gemma 4, DeepSeek V4, Kimi K2.6, MiMo 2.5, GLM-5.1 & others. On CAISI's V4 assessment. An eventful month with one flagship release after another www.interconnects.ai/p/latest-ope...

Open software lowered deployment cost. Open AI lowers development cost. E.g. developing a bespoke model for an enterprise use case. We’re early in companies figuring out how to leverage this successfully. www.interconnects.ai/p/how-open-m...

Pretty wild I got my PhD 4 years ago to the day. I feel very lucky that I got to do it and make my switch into AI. Lot's of people today in AI are underselling the value of going through the process of a PhD.

Great telling of the sights when visiting China’s AI and robotics companies (the same trip I was on!). open.substack.com/pub/ailibrar...

Visiting most of the leading Chinese AI labs, I'm struck by a culture that's extremely well suited to building LLMs with fewer resources, but one happening in a very different ecosystem, more companies at play, almost no data industry, etc. Full report: www.interconnects.ai/p/notes-from...

Added a 1500 word mini history to my book on the path to on-policy distillation being a core post-training optimization technique. Is a fun rapidly growing area now! rlhfbook.com/c/12-synthet...

new pfp, final digital confirmation that I haven't had a mustache in like 2+ years

Adding an on policy distillation section to the RLHF book and it’s remarkable how bad LLMs / coding agents are at it, despite me giving them the core papers and 250 pages of context on how I present ideas.

We need to create a new term for the attacks some Chinese labs are doing on APIs that is different than distillation or else we risk tarnishing a crucial technique that is crucial to AI diffusion, academic research & the open-source ecosystem. www.interconnects.ai/p/the-distil...

Most obvious immediate impressions on coming back to the US from China. 1. The cars here are so lame. So many cool EVs in China, feels like I went 2 decades back in time. 2. The coffee here is so much better. First real coffee in almost 2 weeks changes you.

Demis is the only acceptable answer of which CEO do you trust most with AGI (doubly so until Anthropic/OpenAI go public, Google being public is a great check)

Let’s goooooooooo we are capybara’d up, thanks Qwen, keep the models coming

Feeling the AGI at Zhipu AI

🤘🤘🤘 (Peep the bitter lesson sign, in the Radiohead meeting room)

In Beijing and Hangzhou this week to get to know the AI community here!

Some thoughts on why I don’t think banning Chinese models over distillation is worth it. open.substack.com/pub/natolamb...

Great podcast from someone building LLMs within China, on how the model building is done and thought about there. podcasts.apple.com/us/podcast/a...

A TLDR is that unless the training dynamics of leading LLMs change or open model builders run out of money, this ~6 month performance gap from closed to open models is here to stay. www.interconnects.ai/p/reading-to...

A good thread about this from @juliekallini.bsky.social x.com/JulieKallini...

New video! Talking through my 10+ open model pieces from early 2026 and how they fit together. They're all trying to figure out where open models go next. www.youtube.com/watch?v=hKIX...

Opus 4.7 has a new tokenizer. This means it's also a new base model. Glory days of pretraining still very much going.

The current pace of token-efficient reasoning improvements across minor Claude Opus/GPT model versions is pretty wild. All signs point to this continuing. 4.6 to 4.7 could've been presented as a fairly large model bump in the past with this plot.

I spent some time trying to distill all the complex factors impacting open models -- economics, capabilities, distribution, policy, etc. -- into a clear list of beliefs. Here they are in full. www.interconnects.ai/p/my-bets-on...

One of my key strategies with Interconnects is to develop the practice of making my work obviously compelling to a wider audience, keeping them hooked over time and wondering what I'm up to, etc. www.interconnects.ai/p/what-ive-b...

Excited to launch the accompanying free RLHF Course for my book. To kick it off, I've released: - Welcome video - Lecture 1: Overview of RLHF & Post-training - Lecture 2: IFT, Reward Models, Rejection Sampling - Lecture 3: RL Math - Lecture 4: RL Implementation Landing page: rlhfbook.com/course

a bit over 7 days out from the Gemma 4 release and it's models are outpacing (slightly) the equivalent Qwen 3.5 models on downloads. Big numbers!

Great stuff happening as we start to build out the codebases for my RLHF book (sorry, I haven't had much time until now!). Very accessible to issues, emails, comments etc to make it better. I'm also going to need another dgx spark. github.com/natolambert/...

In 5-10 years, as models get more expensive & capable, I see the funding structures and support for open models breaking down. We need to consider if we need other options of supporting the open ecosystem. The inevitable need for an open model consortium www.interconnects.ai/p/the-inevit...

Incredible video. Very pure experience in why so many of us love science. www.youtube.com/watch?v=oaXR...

1. dont fall for anti open model fearmongering, but 2. acknowledge that AI capabilities are proceeding fast, and eventually there may be a reason to be more careful with open weight models I don't think Mythos is that trigger, but I'm not 100% confident www.interconnects.ai/p/claude-myt...

My book, Reinforcement Learning from Human Feedback, is wrapping up and going into final production (copyediting, making pretty, formatting, etc.). Shipping to you in 1-2 months! It's a wonderful project to create a foundation of knowledge for the research communities that I love and operate in.

Directionally, I agree with this piece, but it's important to note that this is the first blip in a long, slow transition towards a more hybrid model of open and closed models. www.chinatalk.media/p/chinas-ai-...

New report is out with the latest open model adoption data we have gathered for Interconnects & The ATOM Project. At the surface level, we can see Chinese models continuing to accelerate in adoption. The report details much more. atomproject.ai/report

People are too obsessed with benchmarks for open models. The core determining factor of success often is: 1. Immediate & long term tooling support 2. Finetunability Tbh Gemma has struggled here in the past. Qwen has excelled at it. It's where the winners are crowned.

In my latest post, I quickly review Gemma 4, Google's exciting new open models, but mostly focus on the questions around what makes an open model succeed in the long term. www.interconnects.ai/p/gemma-4-an...

The TBPN acquisition is just the tip of the iceberg in a shifting media landscape, where direct relationships are king. We're in the first years of this transition that'll take probably ~10 years to fully settle. The point isn't to control editorial, but control distribution.

Google dropped 4 different Gemma open-weight models! I'm most excited that they're finally adopting a standard Apache 2.0 open source license. huggingface.co/collections/...

Nemotron Super / Ultra Arcee Trinity Large (soon) Gemma 4 (eventually) Reflection's first models (maybe) GPT OSS 2? (maybe) Thinky? Other neolabs? Things looking up for open models built in the US in 2026. We had 0 for a bit there.

Latest open artifacts (#20): New orgs! New types of models! With Nemotron Super, Sarvam, Cohere Transcribe, & others One of our favorite issues of the series so far. www.interconnects.ai/p/latest-ope...

In my latest post, I present my worldview for the future, where I expect sustained, fast AI progress and economic disruption, but none of the fantasy stories about fast takeoff, recursive self improvement, or the singularity. www.interconnects.ai/p/lossy-self...