Joe Lonsdale sat down with two of Anthropic’s reinforcement-learning leads and asked what it feels like to be on the frontier. Their answer: models that can match any human at computer-based work are very likely just a couple of years away.
The timeline isn’t abstract.
Schulte Douglas said he first became convinced in 2020 that scaling would deliver AGI in the 2020s after reading blog posts and papers around the GPT-3 era. He spent nights and weekends meeting the bar at DeepMind, where he worked for two years before joining Anthropic 18 months ago. His colleague Nick, a San Francisco native who joined just over three years ago when the startup was still under five years old, described the same acceleration inside the lab. Eighteen months ago Nick was typing every line of code by hand and guiding a model every few minutes; now he can hand a model a day or two of work and treat it like a junior teammate. On their core evaluation, Frontier Math, the company went from 0% on expert-assembled problems to well over 50-60% in the past year-a jump they tie to reinforcement learning on verifiable tasks such as math and code.
The next six to twenty-four months, they said, will not be about another web app. Both pointed to biology as the domain where AI becomes the most important technology of their lifetimes, not just because models are getting smarter but because combining prior human ideas is now happening at machine speed. The constraint they worry about is physical. Nick argued the U.S. lab complex is lagging China on both capacity and quality-from lab space to supply chain-and that without a buildout the intellectual gains will have nowhere to land. Per Anthropic, its economics team is already modeling scenarios for how that kind of capability shift hits the economy of 2030, a sign the company is treating the 2030s as a planning horizon rather than a forecast.
That horizon shapes their career advice. They see a possible ten-to-twenty-year diffusion window before labor markets fully adjust, where generalists who can pick the right problem to solve and deploy a thousand-person software company will have the advantage. It also shapes their warning. The same scaling that lets each marginal unit of intelligence double productive capacity also lowers the bar for bad actors in bio and cyber, which is why they frame models as entities to be released rather than tools to be used, and why they said the company has slowed itself more than any regulator has.