FleetingBits

Since Mythos, Anthropic has been making aggressive pushes into biology. The end of 2025 was when we saw AI coding agents rise; expect something similar for bio in late 2026 and early ’27, sparked by these new 10T models. Imo, the evidence is everywhere if you look closely: * April: Mythos 5 is the first model to produce novel, compelling scientific hypotheses, and some are independently validated in a lab * April: Mythos’ biggest benchmark leap over Opus 4.8 is in LABBench2, a benchmark testing real-world scientific research * April: Anthropic acquires Coefficient Bio * Mid-June: Anthropic is rumored to have Mythos 5.1 [more RL over Mythos base], which should result in significant jumps in computational bio * June 19: Anthropic hires John Jumper * June 29: Claude Science * June 30: Anthropic announces preclinical drug program * July 9: Anthropic begins hiring several roles for a wet-lab bio research/data group And it’s not just Anthropic — I expect even more impressive results for the GPT-6 base model, and OpenAI seems to think similarly after it released LifeSciBench. Given OpenAI’s world-class RL team, my bet is GPT-6.1 is going to be a step-change for scientific ideation.

This reminds me of a comment by a DC insider who said one of the weirdest things about working with Silicon Valley was that if you mentioned regulating something in the normal way that milk or eggs were regulated, they would start quoting Orwell and say that surely that could only lead to dystopia. My impression is that many pharmaceuticals are currently as regulated than chips would be under Plan A. They can only be made in special government-licensed factories, the government maintains the right to inspect those factories at any time, the factories have to prove that they're only sending them to licensed pharmacies, and the pharmacies have to keep excruciating records showing that they only dispensed them to licensed customers. Has this turned into a global surveillance state, or is it such a boring part of everyday life that nobody notices? I assume the government has access to all of my financial transactions; certainly if I tried to send money to Ayatollah Khameini somebody would notice and stop it. This is creepy and probably does qualify as a global surveillance state, and crypto made a decent effort to dismantle it, and I supported that when it seemed plausible - but has anything bad beyond the obvious happened because of it? I think there's a line between "regulate AI chips as much as we regulate Xanax" and "have a global surveillance state", and Plan A keeps on the right side of it. If you disagree, you'll have to tell me which specific part of it you're worried about.

A very interesting data point on whether distillation gives you a competitive advantage. Sonnet 5 is just objectively worse than GLM 5.2, and is likely around a similar size. Sonnet 5 was almost certainly distilled (and not just distilled from tokens, but proper distillation from logits) from Mythos. This should give you an update on how relatively unimportant distillation is, in particular for the Chinese labs. It's just a way to warmstart the model for RL; the American frontier labs do it with human data, Chinese do it with tokens from American models, but they could get to basically exactly the same model without it, it would just be a little bit more expensive and require more human data collection coordination/aggregation

High talent density companies in no particular order / not including stealth. Cognition, Cursor, Thinky, OpenAI, Anthropic, Modal, Applied Compute, Flappy, Mercor, Standard Intelligence, Etched, Fireworks, Decagon. Source: my friend group