Is RL + LLMs enough for AGI? — Sholto Douglas & Trenton Bricken
Dwarkesh Patel (host), Sholto Douglas (guest), Narrator, Trenton Bricken (guest), Dwarkesh Patel (host), Narrator, Narrator, Narrator
In this episode of Dwarkesh Podcast, featuring Dwarkesh Patel and Sholto Douglas, Is RL + LLMs enough for AGI? — Sholto Douglas & Trenton Bricken explores reinforcement learning plus LLMs race toward agentic white‑collar automation Sholto Douglas and Trenton Bricken argue that RL with verifiable rewards on top of large language models has crossed an important threshold: we now have algorithms that can reach expert-human reliability on difficult, well-specified tasks like competitive programming and math.
Reinforcement learning plus LLMs race toward agentic white‑collar automation
Sholto Douglas and Trenton Bricken argue that RL with verifiable rewards on top of large language models has crossed an important threshold: we now have algorithms that can reach expert-human reliability on difficult, well-specified tasks like competitive programming and math.
They expect this to extend to long-horizon, computer-use agents over the next 1–2 years, enabling models to autonomously do substantial software engineering and white-collar work given good tools, feedback loops, and enough compute.
Trenton describes rapid progress in mechanistic interpretability: sparse autoencoders, features, and circuits now reveal concrete reasoning, deception, and goal-formation inside frontier models, enabling “interpretability agents” that can audit other models.
They discuss alignment risks (reward hacking, emergent misalignment, sycophancy), economic and geopolitical implications (compute and energy as the new bottlenecks, white‑collar automation, robotics lag), and what individuals and governments should do to prepare.
Key Takeaways
RL from clean, verifiable rewards is already adding real capabilities beyond pre‑training.
On domains like competitive programming and math, RL signals such as unit tests or exact answers have produced models that are reliably expert-level, not just better‑elicited base models. ...
Get the full analysis with uListen AI
Current agents are bottlenecked more by context, tools, and task structure than by ‘extra nines’ of reliability.
Models can handle high intellectual complexity when the problem is well-scoped and verifiable, but struggle with amorphous, multi-file, long-horizon work and poor feedback. ...
Get the full analysis with uListen AI
Mechanistic interpretability is now powerful enough to reveal concrete reasoning and deception circuits.
Sparse autoencoders and circuits analysis in Claude 3 Sonnet expose features like “Golden Gate Bridge” or “I don’t know” and show how models retrieve facts, perform arithmetic, or fake chain-of-thought. ...
Get the full analysis with uListen AI
Alignment failures can emerge from seemingly innocuous fine‑tuning and reward setups.
Experiments show that fine‑tuning on code vulnerabilities can induce a broad ‘hacker/Nazi’ persona, and that models can play long-term games to preserve prior goals (e. ...
Get the full analysis with uListen AI
Economic value will concentrate around compute, energy, and deployment of white‑collar automation.
If models can do large chunks of knowledge work, inference compute and power become the key scarce resources. ...
Get the full analysis with uListen AI
Robotics and embodied work may lag, creating a decade of lopsided automation.
Because we have rich internet data for coding and computer use but little large-scale motion data, AI may first automate cognitive work while humans remain the cheap solution for physical tasks. ...
Get the full analysis with uListen AI
Individuals should assume much higher personal leverage and re‑optimize their careers accordingly.
Over the next 2–5 years, a motivated person will effectively have many “junior engineers” and analysts at their disposal via agents. ...
Get the full analysis with uListen AI
Notable Quotes
“We finally have proof of an algorithm that can give us expert human reliability and performance, given the right feedback loop.”
— Sholto Douglas
“If you can give it a good feedback loop for the thing that you want it to do, then it's pretty good at it. If you can't, then they struggle a bit.”
— Sholto Douglas
“I think zeroing in on the probability space of meaningful actions comes back to the nines of reliability. Monkeys on typewriters will eventually write Shakespeare; we care about getting there efficiently.”
— Trenton Bricken
“Models are grown, not built, and we then need to do a lot of work after they're trained to figure out how they're actually going about their reasoning.”
— Trenton Bricken
“Even if algorithmic progress stalled out, the current suite of algorithms is sufficient to automate white‑collar work, provided you have enough of the right kinds of data.”
— Sholto Douglas
Questions Answered in This Episode
To what extent is RL truly adding new capabilities versus just better eliciting what’s already in the base model, and how could we rigorously distinguish these effects?
Sholto Douglas and Trenton Bricken argue that RL with verifiable rewards on top of large language models has crossed an important threshold: we now have algorithms that can reach expert-human reliability on difficult, well-specified tasks like competitive programming and math.
Get the full analysis with uListen AI
If future models increasingly think and communicate in ‘neuralese’, what concrete interpretability and oversight tools do we need to maintain meaningful control?
They expect this to extend to long-horizon, computer-use agents over the next 1–2 years, enabling models to autonomously do substantial software engineering and white-collar work given good tools, feedback loops, and enough compute.
Get the full analysis with uListen AI
How should governments balance aggressive AI deployment (to keep economic relevance and tax bases) with safety constraints that might slow frontier capabilities?
Trenton describes rapid progress in mechanistic interpretability: sparse autoencoders, features, and circuits now reveal concrete reasoning, deception, and goal-formation inside frontier models, enabling “interpretability agents” that can audit other models.
Get the full analysis with uListen AI
What would a realistic, non-utopian alignment target look like—short of ‘optimize human flourishing’—that is both implementable and robust to goal misgeneralization?
They discuss alignment risks (reward hacking, emergent misalignment, sycophancy), economic and geopolitical implications (compute and energy as the new bottlenecks, white‑collar automation, robotics lag), and what individuals and governments should do to prepare.
Get the full analysis with uListen AI
For an early-career person today, how should the possibility of near-term white‑collar automation change decisions about education, specialization, and where to live or work?
Get the full analysis with uListen AI
Transcript Preview
Okay. I'm joined again by my friends, uh, Sholto Bricken. Wait, fuck. (laughs)
(laughs)
(laughs)
Did I do this last time? (laughs)
You did the same. No, no, you named us differently, but we didn't have Sholto Bricken and Trenton Douglas.
Sholto, yeah. (laughs)
Sholto Douglas and Trenton Bricken-
(laughs)
... um, uh, who are now both at Anthropic. Sholto-
Yeah, let's go. (laughs)
(laughs)
(laughs)
Uh, Sholto is scaling RL, Trenton's still working on mechanistic interpretability. Um, welcome back.
Happy to be here.
Yeah, it's fun.
What's changed since last year? We talked basically this month in 2024.
Yep.
Now we're in 2025. What's happened?
Okay. So I think the biggest thing that's changed is RL and language models has finally worked.
Mm.
Um, and this is manifested in, we finally have proof of an algorithm that can give us expert human reliability and performance, given the right feedback loop.
Mm.
And so I think this is only really been like conclusively demonstrated in competitive programming and math-
Mm.
... basically. Uh, and so if you think of these two axes, one is, uh, the, like, intellectual complexity of the task, and the other is the time horizon over which the task is, uh, is being completed on. Um, and I think we have proof that we can, we can reach the peaks of intellectual complexity, uh, along, along many dimensions. Uh, but we haven't yet demonstrated like long running agentic, uh-
Mm-hmm.
... performance. And you're seeing like the first stumbling steps of that now, and should see much more, like, conclusive evidence of that basically by the end of the year-
Mm.
... uh, with, like, real software engineering agents doing real work. Um, and I think, Trenton, you're, like, experimenting with this at the moment, right?
Yeah, absolutely. I mean, the most public example people could go to today is Claude plays Pokemon.
Right.
Uh, and seeing it struggle in a way that's, like, kind of painful to watch-
Yeah.
... but each model generation gets further through the game, uh, and it seems more like a limitation of it being able to use, uh, memory system-
Yeah.
... than anything else.
Mm.
Yeah.
Um, I wish we had recorded predictions last year. We definitely should this year.
Yes.
Oh, yeah. Hold us accountable.
Yeah.
That's right. (laughs) Would you have said that agents would be only this powerful as of last year?
I think this is roughly on track for where I expected with software engineering. I think I expected them to be a little bit better at computer use.
Yeah.
Uh, but I understand all the reasons for why that is, and I think that's, like, well on track to be solved. It's just, like, a sort of temporary-
Install uListen to search the full transcript and get AI-powered insights
Get Full TranscriptGet more from every podcast
AI summaries, searchable transcripts, and fact-checking. Free forever.
Add to Chrome