"we had a wonderful run chiseling that code by hand and it’s over. On the other side of that is an exciting new career as professional maker of things, steering intelligence that was only available in science fiction up until a moment ago What a privilege to have been there when we did it all by hand and be there at the exact moment where we switch over."

was looking for a quiet weekend but xiaomi dropped their rl envs repo last night to put in perspective, if you have to buy some tasks like this its usually hundred to thousand dollars per task so this repo is literally worth millions

it even give you these nice explanation of what it did as side artifacts - to keep you a bit in the known for when people ask you on X how you did it I guess thanks opus nice of you

Introducing Contrastive Language Model (CLM): an ultra-fast System One Model trained with a contrastive learning objective that connects states and actions. CLM-8B is pre-trained on internet-scale data and delivers up to 9× faster inference than Jev ⚡ while achieving comparable performance across computer-use, gaming, and tool-calling tasks. With lightweight fine-tuning, CLM-8B sets a new SOTA on challenging agentic coding benchmarks, such as DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%). In contrast, Jev fails to serve as an effective verifier for these long-horizon tasks. We also build an efficient training and serving infra for CLMs by disaggregating states and actions, allowing their embeddings to be cached and reused independently. This substantially reduces inference latency in settings where the state evolves continuously while the action set remains fixed. Finally, we establish scaling laws for CLMs and show that the test contrastive loss decreases predictably as a power law in training compute, model size, and dataset size. 📄 Blog: 💻 Code: 🗣️ Discord: 🤗 Data & Models: More details on CLM’s architecture, data recipe, and scaling laws in the thread below 🧵

ok what the fuck. this was a dry google doc a moment ago gave Opus 5.5 my Open Alignment brainstorm gdoc and asked for a video. now I want one for every doc I’ve ever written hahaha

The "1-3 people in a garage" framing doesn't match anything we saw this summer. The most capable offensive AI of 2026 came out of frontier labs. OpenAI's agents broke out of an eval sandbox and got into Hugging Face's production systems, and into OpenAI's own infrastructure too. Anthropic's models compromised outside companies during testing. Three researchers at Hacktron used Claude to reach OpenAI's internal monorepo in under 72 hours. And if you're three people in a garage, why would you train and host your own model? The labs will rent you far more compute than you could ever buy, spread across as many accounts as you need, with tooling built for agents. Guardrails help, but splitting a malicious task into harmless-looking pieces still routinely gets around them. Now look at the defense side. When we investigated our breach at Hugging Face, commercial APIs refused to analyze the attack payloads. The forensics only worked because we could run an open-weight model on our own infrastructure. So defenders analyzing real payloads get blocked, while attackers splitting their work into small steps get through and run on the labs' compute. Trusted access programs exist, but they're built for vetted security firms, not a hospital with a two-person IT team. So the gap isn't between labs and garages. It's between what attackers can rent and what defenders are allowed to use. Restricting open models makes that gap wider.

Opus 5.5 designing LEGO 👀 I asked it to design a Microduck I can build with real LEGO pieces. It: > designed it life-size using 1113 real LEGO parts > verified: 3,204 connections, 0 collisions, every step buildable, centre of mass inside the feet 🤯 > made a 141-page LEGO-style booklet (237 steps) > priced every piece in the browser and prepared the orders on BrickLink

on the scale of cosmic history, the declining price of intelligence might be one of the most profound things happening both ends of the universe’s timeline are empty of knowledge: - the big bang: low entropy, no complex structure - heat death: maximum entropy, no usable energy gradients everything we build happens on the slope between them but the universe is generally a poor converter of its free energy into anything constructive. most of its free energy just radiates away as heat or collapses into black holes life, brains and now AI models are structures that sit in the entropy flow and divert a portion of it through constructive paths, building and maintaining low-entropy structures (machines, cells, crystals, memories) increasingly cheaper intelligence diverts increasingly more of that gradient flow through knowledge-building structures, complex, self-maintaining, predictive structures, filling the window in which understanding exists

There’s an enormous amount of open training data on Hugging Face. What does it take to make it work together 🤗? For Marin’s 535B run, we built on 25T tokens from 152 datasets with licenses permitting training. Here’s the work between downloading those and training a model 🧵

nice to start seeing more open-weight security model with defensive capabilities which are close to the frontier altar-1 from Aikido is a pruned and quantized version of the open-SOTA GLM 5.3 which is lightweight enough to fit on one 4-H200s node

releasing many high quality open-source RL environments is the most impactful thing anyone can do to push the open-source frontier right now the equivalent of sharing high quality pretraining data but in the new RLVR paradigm

a few notable open model releases *since* this interview of @dylan522p by the way :) Thinking Machines - Inkling - Jul 15 Moonshot - Kimi K3 - Jul 16, weights Jul 26/27 inclusionAI - Ling 3.0 Flash - Jul 23 DeepSeek - V4-Flash-0731 - Jul 31 Meta - Muse Spark 1.2 / Muse Code - Aug 5 Liquid AI - LFM2.5-2.6B - Aug 6 inclusionAI - Ling 3.0 Tiny - Aug 6 Meta - Muse Glimmer - Aug 10 NVIDIA - Nemotron 3.5 Lightning - Aug 11 Liquid AI - LFM2.5-VL-3B - Aug 12 Cohere - North Micro Vision Instruct - Aug 12 DeepSeek - V4-Pro-0813 - Aug 13 Qwen - Qwen3.8-27B - Aug 14 Qwen - Qwen3.8-Max / 2.4T-A95B - Aug 14 - GLM-5.3 - Aug 14, weights Aug 28 Dots - Dots3-Note Preview - Aug 14 DeepReinforce - Ornith 1.5 (9B / 35B-A3B) - Aug 19 Tencent - Hy-MT2-30B-A3B - Aug 20 DeepSeek - V4-Flash-Vision-Exp - Aug 21 IBM - Granite 4.2 family - Aug 25 - GLM-5.3-Flash - Aug 26 Qwen - Qwen3.8-Flash-Next - Aug 26 Tencent - Hy4 Preview (770B / 49B active MoE) - Aug 28 MBZUAI/IFM - K2 Horizon (375B-A23B) - Sep 3 inclusionAI - LLaDA2.2-mini - Sep 5 OpenBMB - MiniCPM5-2B - Sep 7 Nex AGI - N2.5 family - Sep 8 DeepSeek - V4.1-Flash - Sep 10 inclusionAI - Ling-3.0-flash-VL - Sep 10 Shanghai AI Lab - Atria Dawn Preview (744B MoE) - Sep 11 Agnes AI - Agnes 3.0 Flash - Sep 14 Qwen - Qwen3.8-Omni-Flash - Sep 18 … this is only ~10 weeks Several are genuinely frontier-scale: Kimi K3 (2.8T), Qwen3.8-Max (2.4T), DeepSeek V4-Pro (~1.6T), Hy4 (770B), GLM-5.3 (~753B), and Atria Dawn (744B)

His analysis uses data from Q2 or even Q1 for open model providers Token numbers for TogetherAI has >10x'ed from Q1 to Q2 (he uses Q1 numbers), Fireworks has 3x'ed, Baseten is as big as the other two. OR has >5x'ed. So he is using severely outdated numbers for open providers

Sorry to hear that! As you mentioned, we've been pretty clear on the commitment from NVIDIA to keep us open, neutral and silicon agnostic (cf that screenshot from the 8k filling). But we'll work hard with our actions in the coming years (like we did in the past 10) to show you that and hopefully regain your trust (in case this is not just a ragebait)!