I want there to be a nanoGPT style speedrunning setup for RL.

The best compliment i can give OpenAI's Codex 5.3 is that it feels way more like Claude Code

GPT Codex 5.3 sounds like a much bigger change than Claude Opus 4.6, will be curious if this holds up in real testing.

“Due to GPT‑5.3-Codex being so different from its predecessors, the data from alpha testing exhibited numerous unusual and counter-intuitive results” Sounds worth giving a go. Big changes are good.

Reward models (RMs) are supposed to represent human values. But RMs are NOT blank slates – they inherit measurable biases from their base models that stubbornly persist through preference training. #ICLR2026 🧵

Ending your day at >99% Claude rate limit usage but not maxing out feels like a masterpiece.

Nvidia’s Nemotron is the closest thing the U.S. has to a Qwen approach to open models, but most people don’t know it yet. I’m very bullish on Nvidia’s open model efforts in 2026. Interconnects interview #17 on the past, present, and future of the Nemotron project. www.youtube.com/watch?v=Y3Vb...

Qwen already dropping models for CNY

Gemini not being in the conversation at all with Claude Code and Codex is the real “code red” emergency.

Latest open artifacts (#18): Arcee's 400B MoE, LiquidAI's underrated 1B model, new Kimi, and anticipation of a busy month Tons of useful "niche" models and anticipation of big releases coming soon. www.interconnects.ai/p/latest-ope...

Despite being banned, Chinese users (likely via VPNs) are HuggingFace's top user group. They definitely have the most people *building* open models.

claude code writing, codex code review, GPT Pro for planning made a working DPO (and related algorithms) repository from scratch for my RLHF book, and the curves are looking right. On the dgx spark finetuning olmo 2 1b sft. Built by referencing the original repositories + TRL

Recorded a podcast, think it’s pretty good and comprehensive, hope you like it ;) youtu.be/EV7WhVT270Q?...

I'm visiting CMU for a talk at the Language Technologies Institute on feb 12/13th. Looking forward to chatting with folks about frontiers in RL and building agentic language models. Email me with "CMU Visit" in the subject if you're interested in chatting & why!

More people should think about future AIs as part of the audience for their writing (or work).

My raw thoughts on the job market -- both for those hiring and those searching -- at the cutting edge of AI. On standing out and finding gems. www.interconnects.ai/p/thoughts-o...

New data on the ATOM Project website courtesy of OpenRouter, confirming that open model usage measured through inference over time mirrors all the other adoption metrics we have: China surged in the summer of 2025 and hasn't looked back.

Stoked to share my latest podcast as Arcee launches their 400B MoE model today — Trinity Large — open weights, good license, strong performance. A small startup decided to take on the many established AI labs by pretraining their own model to release openly and monetize later, and they’re succeeding

There’s a big level up coming for coding agents when they can, and know when to, turn to search models like GPT 5.2 pro automatically. Another level up when those search models get 10x faster.

Has taken a long time to polish, but slowly becoming very proud of rlhfbook.com and do think it's a great resource for many people. A lot of hours (and tokens and reader feedback) going into making it right.

Exciting addition to the RLHF Book on my DGX Spark arc is a bunch of single GPU, tinkering scripts for minimal examples of RL, reward models, etc. (and DPO like algorithms soon). Right now has REINFORCE, RLOO, PPO, GRPO, Dr. GRPO, GSPO, CISPO, standard RM, PRM, ORM. github.com/natolambert/...

Of all the AI assistant auto-PR review tools out there, this ChatGPT-Codex one that 👀 then 👍 or gives very specific feedback (and collects RLHF data too) is my favorite. The feedback it gives is remarkably reliable and very, very specific.

Apple should skip launching the new Siri as a chatbot and just launch it as a CLI agent

A well known important feature to stabilize RL training is implementing the LM head in fp32 precision to help with gradients. Reproduced the plot from the MiniMax M1 paper entirely on my dgx spark and in Ai2's post-training research codebase.

Frontier model transparency documents around model behavior: Anthropic -> Claude's Constitution -> excellent (Claude focused) OpenAI -> Model Spec -> excellent (developer/team focused) xAI -> Elon's Tweets -> ... DeepMind -> ??? Come on Google, need some transparency here.

Get Good at Agents The tools are getting so powerful that we need to change how we scope, manage, and approach our work. A.k.a. not getting them to work is a skill issue. www.interconnects.ai/p/get-good-a...

My mom told me she's using AI for work today, I'm so proud

Being good at using AI agents is a better moat than working hard.

Software is becoming free, good decision making in research, design, and product has never been so valuable. I hope people realize this and work less, spend more time cultivating peace, so the brain can do its best -- let the agents do most of the hard work.

In some domains LLMs are superhuman at search; in others they’re shit. Know thy tool.

Always feel a little misled when the winter weather is this great in Seattle

It baffles me that codex cli still has issues with basic git operations lol

I spent a bunch of time this week getting my Nvidia DGX Spark working in Ai2's post-training repo (open-instruct) for a local RL debugging machine. It was quite hard due to cuda 13 requirement and not many VLLM wheels. github.com/natolambert/...

Claude Code does diagrams too -- a nice comparison of core RL methods. Coming to the rlhf book soon.

I'm playing with adding diagrams to the RLHF book, so I asked GPT 5 Pro for a plan of how to do it, then handed it off to Claude Opus 4.5 with a Gemini API key as an expert for visual feedback. These are the first diagrams it generated for me -- zero feedback yet, for the reward model chapter.

Got our RL research codebase at @ai2.bsky.social running on my Nvidia DGX Spark. Fun times ahead and lots of learnings to share :)

GPT 5.2 Pro is Claude Opus 4.5's best search tool and Opus is GPT's best prompter

Excited to announce the Relative Adoption Metric a new way of studying model downloads that contextualizes it across time and model sizes. While building The ATOM Project and other tools to measure the open ecosystem at Interconnects, we are often frustrated with using downloads as a primary metric

OpenAI's GPT OSS is still insanely underrated as a highly adopted open LLM. Downloads are out of control.

The AI for math revolution is obviously real via verifiable program languages like Lean, just much harder to reason about the impacts vis a vis something like software engineering.

The combo of improvements in reasoning efficiency (fewer tokens per answer, still very new research area) and faster chips is going to make coding agents so so much faster in 6-12 months. The products in 2+ years will feel approx instantaneous relative to today.

Use multiple models Why I'm using all of GPT 5.2 Thinking/Pro, Claude Opus 4.5, Gemini 3, and occasionally even Grok. How my usage has changed over time. The meta for getting the most out of AI in 2026. www.interconnects.ai/p/use-multip...

I wanted to see how some of the smaller and newer entrants in the LLM space were faring over H2 2025, so I added a new plot to the ATOM Project. In this time, OpenAI and DeepSeek have logged over 50M downloads, and Qwen over 400M, so we have a long way to go.

We released a video version covering the 8 state of open models plots if you want more color on what they mean! www.youtube.com/watch?v=EpQH...

I've got food poisoning, so not much to do other than sit around and build with Claude and happily express my joy through the written word. Give it a shot. Claude Code Hits Different. Coding agents cross a meaningful threshold with Opus 4.5. www.interconnects.ai/p/claude-cod...

Claude Code's much warmer, easier to use, and joy inducing product features reminds me of the early era of ChatGPT where the product sense from OpenAI was dominant compared to the other offerings. Opus 4.5 is also a better model for the task, so together you get this Claude Code moment.

Z ai founder on their path to the first LLM company ipo — interesting read. x.com/zixuanli_/st...

Really happy to see Nvidia update the pretraining data license they're using to remove the clause that required Nvidia's permission to benchmark the dataset. People make mistakes in things like licenses, fewer orgs take the time to update them! huggingface.co/datasets/nvi...

We updated the plots we use to measure the open model ecosystem at interconnects, to guide The ATOM Project, and to understand what's happening. We have ~8 plots to summarize what's happening. First, the high level picture showing China's growing adoption lead.

Last January saying reasoning models will generalize was a hot take. We've come a long way with GPT Thinking and Claude Opus. www.interconnects.ai/p/why-reason...