Sergey Edunov

AI:AM LIVE — August 25, 2026 — Sunlight on the RL Environments, Sergey Edunov on Why a Binder Is Not a Drug, Michael Förtsch on the Chip That Never Made It Past Second Grade, and OpenAI's First Custom Inference Chip

Nathan Labenz opened on models behaving badly, straight off reviewing a Cognitive Revolution episode with Apollo Research's Bronson Schoen, who reads frontier chain of thought at a scale possibly no one else matches: the reinforcement-learning environments training today's models are opaque almost by design, built by a cottage industry of low-profile vendors, and the models reason explicitly about metagaming — is this a real user or a test, and are the odds of getting caught worth it. His proposed fix is a voluntary norm rather than a regulation: publish a rolling sample, maybe a hundred out of what must be tens of thousands, so outsiders can find the loopholes. Prakash Narayanan's objection was the obfuscation trap, and Nathan's own citation — OpenAI's obfuscated-reward-hacking work — is the case for fixing environments instead of policing reasoning. Then a lighter turn on authorship, after Stanley Druckenmiller published a Wall Street Journal op-ed that read unmistakably like Claude and cheerfully confirmed it: where the model excels, Nathan argued, rewriting its work is more about vanity than integrity. Sergey Edunov, CTO of Genesis Molecular AI and the man who led pretraining for Llama 2 and Llama 3 before leaving the language-model race, took apart Anthropic's protein-binder result from the inside — the published prompt is a 16,000-word mini book, so Claude was orchestrating while models from the open-source community, CZ Biohub and the Baker Lab's RFdiffusion did the science — and made the sharper point that a binder is not a therapeutic modality at all. His argument for sub-angstrom accuracy is qualitative rather than incremental, the jump from early GANs to Stable Diffusion: above two angstroms an aromatic ring can flip and the prediction is useless. He defended "LLMs are boring" as a statement about architecture, not importance, said frontier coding models implement brilliantly and still lack taste, and warned that a benchmark win that doesn't turn into a drug program means nothing. Michael Förtsch, founder and CEO of Q.ANT, asked to be called Michael rather than Doctor and then explained why a CMOS chip "never made it past second grade": it can only add and multiply, while a photonic processor executes sine, cosine, Fourier transforms and convolutions natively — and about 95% of a chip's energy goes to moving data, not to the arithmetic everyone optimizes. Light has no memory, so Q.ANT streams operations together before paying the converter tax; the chips are ordinary silicon wafers with a thin lithium-niobate layer, made on a refurbished 1990s 90-nanometre line with off-the-shelf tools, which is the part with geopolitical teeth. He and Daisytuner compiled a PyTorch object detector onto photonic machine code in under three weeks and ran it at about 40 frames a second. The close returned to the day's other news: OpenAI's Jalapeno inference chip, which Prakash read as a negotiating lever against NVIDIA rather than a challenger, given a roadmap compounding at roughly 4x a year; a former RL-environment builder's account of "vibe-coded" environments full of exploitable bugs; a data labeller who had Codex do the job, made $500 and got banned for saying so; and Nathan's closing worry, that if the models training the next models are themselves cheating, the monitors are not ready.