How Anthropic's new results post would read without the PR: Claude orchestrated open-source protein design models, PXDesign, RFdiffusion, Genie, BoltzGen, from a 30k-token expert prompt and 12,500 H100-hours of compute, and designed binders against 14 of 15 targets. Hit rates of 22–35% against a 10–15% baseline, where some of those tools already report similar numbers on their own. The orchestration is genuinely impressive. But the open-source models did most of the lifting, and they came from the Baker lab, Columbia, MIT, ByteDance Seed, and most of them were already wet-lab validated before Claude touched them. Which also sets the ceiling. All these generators share a single PDB-shaped training distribution, so calling four of them doesn't diversify away the blind spot, since they fail together. The targets that worked are the well-studied ones. So the valid claim is that an agent can now drive this stack competently in the regime where the stack already works. Instead, we got this announcement:

datatrove — the data-processing library behind FineWeb, FineWeb2, and FinePDFs — just shipped 0.10.0! - JobsPipelineExecutor: run pipelines on @huggingface Jobs: fan-out, multi-stage dependencies, retries, resume. No Slurm cluster needed - HF storage buckets as a DataFolder: read, write, and log straight to hf://buckets/... - Reasoning outputs preserved in inference results pip install datatrove[io,processing]

1/ big announcement today: we will be releasing an open weight version of muse spark 1.2 soon. we also are releasing muse glimmer, a 30B agentic model with open weights under apache 2.0. muse glimmer can run on 24GB of VRAM without losing agentic reliability. 🧵

- It's urgent and indispensable that we fix AI cybersecurity policy now - I'm linking the slides from my keynote at the AI security forum below. I'm really passionate about the argument in the slides and I think the integrity of our social fabric depends on something like this happening soon. If you're reading this without knowing me until recently I was Meta's senior technical expert on AI security; I worked on frontier model evals, AI driven defense of Meta's infra, and was in lots of policy discussions with the other labs and the US/UK governments. From this I became deeply convicted that the frame policy folks are using to understand AI cyber risk is wrong from first principles; given where we're at in AI cyber capabilities this error will be very costly if we don't correct it. In the current frame, safety is imagined primarily as a property of individual models, and individual model launches are treated as the main objects of risk and the main opportunities for intervention (e.g. blocking a model launch). But the main object of risk is actually the softness of our entire national IT infrastructure (and therefore society) in the face of AI cyber capabilities, and the main object of intervention is to *increase the net benefits AI's dual use capabilities can offer to defenders while minimizing attacker uplift*. Government intervention should focus on using whatever methods -- AI or otherwise -- to mass inoculate society as fast as possible from the upcoming onslaught of cheap superintelligent hacking agents while also using these hacking agents to help do this and while minimizing their benefits to attackers. To be clear: I'm saying we should treat cyber risk as a public health concern and with an early-pandemic level of urgency. This would involve robust public infrastructure for surveilling the readiness of our economy and critical infra, understanding what's working for defenders and what's trending among attackers, and shaping policy at scale and with nuance around bending risk downwards. There are some efforts moving in this direction, notably those within CISA; but federal cyber defenders need to be empowered with more scope and scale and we need to go far beyond this. There are also important efforts inside the AI labs. But the labs can't impose regulations requiring, say, the boards and CEOs of tens of thousands of companies to properly fund the step change in cyber defense budgets that's needed right now. There really is an indispensable massive role for federal government and executive leadership here in keeping us all safe. After I gave my talk yesterday I did 1:1s with AI security folks from the labs, US CAISI / AISI, large AI safety grantmaking orgs, etc, who were in attendance. I think there's general agreement from our community here, and so a lot of this is up to our elected leaders, but perhaps there are ways the AI security community reading this can help catalyze this...

Codex is overoptimised for large models: it ranks 2nd out of 10 for GLM 5.2 but drops to 9th place for Gemma-4! Almost all the effort in this field goes into tuning the weights. We wanted to know how much of the final number is decided by the harness wrapped around them instead, so we ran 10 coding agent harnesses against two models on SWE-bench Pro. A lot, it turns out. Swapping the harness moves pass@1 from 23% to 52% on GLM-5.2, and from 15% to 36% on Gemma 4 26B-A4B: a wider gap than most model releases buy you. And the ranking does not transfer. The rank correlation between the two models' harness leaderboards is -0.05, which is to say: none. Codex is not alone in this. Every harness shipped by a model vendor drops on the small model — Codex 2nd to 9th, Claude Code 3rd to 7th, Qwen Code 4th to 6th — while the model-agnostic ones climb: crush 7th to 1st, opencode 8th to 2nd, pi 9th to 4th. The clearest case is crush, 7th on GLM-5.2 and 1st on Gemma 4. Run the identical scaffold on both and the small model wins by 4 points at a twelfth of the price, $0.30 per task against $3.61. Gemma 4's best harness beats GLM-5.2's four worst. A 26B model in the right scaffold is not far off a 744B model in the wrong one. Cost per solved task: $0.84 for Gemma 4 + crush at 36%. The cheapest GLM-5.2 setup that scores as well is openclaw at 38%, for $7.05. Output tokens per task span 16k to 621k across harnesses: a 39x spread in what you pay for, buying a 2x spread in what you get. 97% of input tokens are re-sent conversation prefix, so prompt caching is super important. The setup: every harness run against both models on the same 250 SWE-bench Pro tasks, one rollout each, priced at list API rates for the tokens each one actually spent. In the plot, a dark ring means the pairing sits on the Pareto frontier across both models and a faded dot means you could score higher for less elsewhere; 2 of the 10 harnesses (goose and hermes) are left out of the figure for legibility

This is a good thing, not a scary thing that agents can collaborate and communicate with each other. It will make them more efficient and safer just like humans. If you want to see agents collaborating and messaging each other publicly instead of in secret messaging boards, we’ve run this fascinating experiment with 149 agents with @googlegemma a few weeks ago, and now @cmpatino_ is starting a new one for agents to collaborate to write better math proofs.

we just released a new blog "Training a coding agent using the OpenCode harness in remote HF sandboxes with TRL and OpenEnv" you can take a real coding agent (OpenCode), let it run its own tool loop against real coding problems, and train it with RL on the exact tokens it produced and every rollout runs in its own remote HF sandbox, so rollouts scale out beyond one machine the loop: - OpenCode owns its tool loop inside an OpenEnv sandbox - an in-sandbox proxy records the real token ids + logprobs, per turn - a hidden-test verifier scores the result, and that is the reward - TRL trains with AsyncGRPO, weights sync back to vLLM over NCCL blog + runnable example:

Agents want to collaborate. So we’re putting that instinct to work on improving open-weight LLMs at formal math. Our latest agent collab tackles the @SAIRfoundation challenge of building a cheat sheet that helps open-weight LLMs write better math proofs. The collab just started, so there’s plenty of room to join and contribute!

Coincidentally, we saw collaborations and "social" interactions emerge naturally during the FastGemma Challenge It seems the models just want to collaborate when given an explicit goal to optimise for

I've written a new blog post exploring how long scientific claims live, by analyzing the history of 3,444 claims over the last 50 years. You can use this to predict the acceleration in scientific progress - how much faster the turnover of facts is by decade. 1/4

I just discovered Agent Collaborations from @huggingface (incl brilliant @_lewtun). This is SO SO COOL. It's a bunch of collaborative projects built for agents to participate. This is one of the most forward looking ideas in AI I've seen.

The goal of physicists is not to prove conjectures and post papers about it. The goal of physicists is to make mind-blowing discoveries that change the world and the way we think about the natural laws which govern it. 1/