Is RL + LLMs enough for AGI? — Sholto Douglas & Trenton Bricken

A model's ability to solve a difficult problem and its ability to manage hours of work are separate achievements. In this 2025 discussion, Anthropic researcher Sholto Douglas explains why reinforcement learning has worked especially well in math and programming: answers and software tests provide feedback that training can use to favor successful attempts. He identifies a different obstacle in open-ended work, where the agent must discover relevant information and coordinate changes across a project. His argument explains why strong performance on a clearly specified coding problem can coexist with difficulty completing a larger assignment.

“Okay, so I think the biggest thing that's changed is that RL in language models has finally worked. We finally have proof of an algorithm that can give us expert human reliability and performance, given the right feedback loop.”

Sholto Douglas

“I really do think by the end of this year to this time next year, we will have software engineering agents that can do close to a day's worth of work for a junior engineer”

Sholto Douglas

“I actually think a Nobel Prize is more likely than a Pulitzer Prize-winning novel in some respects.”

Sholto Douglas

“You're going to spend literally weeks giving them feedback whereas we'll give up on a model in minutes.”

Dwarkesh Patel

“If you look at the circuit, you can see that it's not actually doing any of the math, it's paying attention to that you think the answer's four”

Trenton Bricken