AGI in a couple of years — Anthropic's RL leads on coding agents, Frontier Math and biology

Key points

Key takeaways from the American Optimist interview with Sholto Douglas and Nicholas Marwell, two reinforcement-learning leads at Anthropic, hosted by Joe Lonsdale (October 2026):

AGI within a couple of years. Douglas believes models at least as capable as all humans are very likely in the next couple of years, multiplying intellectual and physical labour — potentially centuries of progress — but Anthropic sees its job as shepherding the world through the attendant risks.

Coding agents have become junior team members. Eighteen months ago Douglas was typing every line of code by hand; now he asks models to do a day or two of independent work, intervening like a manager, and they drive progress on RL science tasks.

Frontier Math is falling fast. Scores on Frontier Math, a set of problems assembled by professors across fields, have gone from near zero to well over 40–60% in the past year — because verifiable, worked RL problems are easy to train and check, unlike romance or poetry.

Biology is the next frontier. Marwell calls AI the most important technology in the life sciences certainly of his lifetime and maybe ever, expecting impact over the next 6–24 months similar to software engineering. The constraint he worries about is US physical and lab infrastructure, and its quality, which trails China.

Think of models as agents, not tools. Marwell argues the entity frame is more useful than the tool frame: models take actions on your behalf in the world, rather than being operated the entire time — though he personally tries not to anthropomorphise them.

Run-rate talk is leaking. Lonsdale noted leaked figures putting Anthropic on roughly $80bn+ run rate — ahead of OpenAI — although neither guest confirmed numbers, and Douglas kept a deliberately straight face.

The regulatory-capture charge. Asked about claims that Anthropic seeks rules to slow rivals, Douglas pushed back: “So far we've slowed ourselves down much more than anyone else.” The conversation also covered open-source and distillation dynamics with Chinese labs, and whether Silicon Valley researchers share heartland values of virtue, faith and institutions.

Read more: YouTube — American Optimist · Joe Lonsdale's blog