I wish I were in the room: Ilya: Jensen, my friend. We have reached the point where our research is worth scaling! Jensen: Sure. How many GPUs? 100k? Ilya: More. Jensen: 200k? Ilya: More. Jensen: 300k??? Can you at least give me something? One hint. One benchmark where you're better. Ilya: We are done with research. We need to scale! Jensen: Anything. A number. A vibe. Ilya: Jensen, you are a nice guy. But Sundar is standing right outside, and he doesn't ask questions. Jensen: Where do I sign. Ilya: I am honored by your conviction. See you next year

Attackers have frontier AI. Defenders need a frontier AI ecosystem—the best open and closed models, force-multiplied by a global community. During the Hugging Face incident, closed AI blocked essential forensics. An open-weight frontier model helped contain the intrusion. That’s why we created the Open Secure AI Alliance.

Opus 5 reports 30% on ARC-AGI-3, ~4× the previous best model, ~20× its predecessor Opus 4.8. We tested it on Witness, our held-out suite of ARC-AGI-3-style interactive puzzle games. The leap doesn't transfer. On Witness composites (same harness, same budget for every model), Opus 5 lands at 43.4 ± 3.2, which is a statistical tie with kimi-k3 (42.8 ± 1.9) and Fable-5 (43.8 ± 9.7). Ahead of Opus 4.8 (34.8), but nowhere near a generational jump. The traces tell the why: (1) On our most classic Witness-style game, Opus 5 states the hidden rules before its first action, then plays a byte-identical optimal solution in 5/5 seeds at temperature 1.0. Zero exploration. It already knows this genre. (2) But on our most novel game (unusual mechanic combinations you can't pattern-match), Opus 5 regresses below Opus 4.8. Where rules must actually be discovered through interaction, the new model is worse than the old one. That decomposition (perfect on templates, regressed on novelty) is the signature of “scaffold-then-internalize” training on genre-specific data, not a general gain in interactive abstract reasoning. Our benchmark can't tell whether that data was their in-house ARC-AGI-3-like corpus with ARC-AGI-3-specialized harnessing (likely thanks to [schema]? or public Witness-genre corpus, or both, but it can tell the improvement isn't general. Held-out evals only stay held-out while nobody's optimizing for the genre, and that clock is always ticking. It's ticking for Witness too, the moment we publish it.

Preparation in progress for the San Francisco Open Weights mini-march tomorrow. Hope all these big tech CEOs won’t show up otherwise I might end up in jail (haven’t had time to ask for a permit 😅)

Jensen (first post!) argues that open models allow companies and countries to inspect, modify, and run AI on their own infrastructure, with less dependence on a few AI providers. The conclusion is that the world needs both frontier-closed and frontier-open models. I know that in some magic way, the conclusion is: Yeah, we need more models that will run on Nvidia GPUs, but I agree with that 🙄

I read the GLM-5.2 report and saw they use IndexShare, which is a cool, simple trick. Regular attention makes every token look at every other token, which is the quadratic cost everyone keeps trying to kill. Sparse attention is a workaround where each token only looks at a small set of relevant tokens instead of all of them. In DSA the way you pick that set is a small "indexer" that scores the keys and keeps the top-k. The indexer stays cheap but still picks well because it's trained to imitate the real attention distribution with a KL loss, and ranking which tokens matter turns out to be a much easier job than computing the exact attention, so it can run in FP8. The problem is that the indexer is itself quadratic, and it runs at every layer. so at 1M context most of your compute goes into deciding what to attend to, not into the attention. The trick with IndexShare is that instead of running it every layer, you share one indexer across a group of 4 layers and let the other 3 reuse that selection. they got 2.9x fewer FLOPs per token at 1M! the idea is betting the set of tokens worth attending to barely changes from one layer to the next, so recomputing it every layer is wasted work. This idea of sharing across different layers is not new, of course. things like HySparse or Kascade do similar reuse but keep a few real dense-attention layers around to compute the "true" selection. GLM takes it one step further and reuses the output of an indexer that was already an approximation, and it holds up because the model is trained that way from mid-training, not switched on at inference. Super simple!