It's not just TB2.1 ...

7/ These findings show that representation of conscious and suppressed content alternates every ~ 60 ms. Thus conscious perception does not reside in a specific brain region, but rather a specific temporal phase.

We collaborated with @scaleAI to build a high-fidelity port simulation RL environment where AI agents tackle real-world logistics challenges

Cómo se puede contar algo tan humano de una manera tan guapa…

'Tis the week of test-time communication! :D Very cool work from Aneesh et al. going into depth on the way agents can communicate, how they can learn these strategies, when do they generalize, and what are the archetypes of tasks where these help.

We’re open-sourcing Pro and Flash, MiMo-V2.6-Distill-Qwen-9B, the technical report, 7K+ RL task environments, an end-to-end RL framework and composable mini-harnesses. Reproduce, verify and build on the work.

Computer scientists discover the power of *closed loop* and *error correction*! I have advocated for many years by now that all intelligent systems would be closed-loop, not open end-to-end. This has been the very basic lesson from Cybernetics (control theory and systems theory), originated more than 80 years ago. This view has been prominently featured in our open-sourced new book about the principles of intelligence: ma-lab-berkeley.github.io/de…

looked into rollouts and found out the 9mo old parser is broken

Agentic evals are messy. A benchmark score tells you something about model performance, but it also reflects the whole system around it: the harness, sandbox, dependencies, timeouts and sometimes a loophole the agent found in the task. That’s why trajectories matter so much to us. They show what the agent actually did and whether the score means what we think it does. We publish them to make that evidence transparent and auditable, giving the wider community more to learn from. Watch @aalSonOfRavi and @ConnorBAdams go deep on all of this with @petergostev from @arena, including some surprisingly creative reward hacks!

Congrats to @Zai_org on GLM-5.3. It massively beats every American open model: Nemotron, Laguna, Inkling. It's not close. GLM-5.3 uses the same base model as GLM-5.2; every gain came from post-training. If America keeps going down the "Nemotron Committee" path instead of fostering real competition, it will keep losing.

Great to see @dotsstudioai from the popular Chinese social media app (@xiaohongshu, known as RedNote in English) join the open-weight community! ❤️ The team is still trying it out for day-to-day use to see the actual quality of the model or if it is eval maxed. Regardless, it is always great to see another company join the open community!

But we also took a chance to have a look at some in-the-wild scheming, reward seeking, etc. examples, and dumped it in appendix. 1) Summarizer unfaithfulness Reasoning summaries often omit important information from the original trace. Here, Opus 4.8 realizes it knows the answer to an AIME problem and then tries to fit a solution to that answer. None of this appears in the summary.