Will continued scaling and improved reinforcement-learning environments keep producing large general intelligence gains beyond current human researchers despite diminishing returns and compute constraints?
The debate asks whether scaling laws and improved RL environments will keep producing big intelligence gains beyond human researchers despite diminishing returns and compute limits. Patel (Dwarkesh Podcast) and Millidge (Dwarkesh Podcast) argue that scaling RL across many environments can train agents with meta-skills that surpass human AI researchers. O’Neill (Dwarkesh Podcast) argues that diminishing returns and compute limits around environment design will flatten future gains, a point raised on these shows since early September.
Scaling and better RL environments can keep driving large intelligence gains beyond human researchers.
Millidge (Dwarkesh Podcast) states that training in environments, rather than just distilling trajectories, can push AI research ability far above human researchers. Patel (Dwarkesh Podcast) claims that scaling RL across millions of diverse environments can teach agents transferable skills like persistence and collaboration that let them ramp up quickly in new domains. Rodriguez (80,000 Hours Podcast) and Shapira (Doom Debates) argue that algorithmic improvements and post-training or swarm scaling can keep boosting capability even when training compute is tightly regulated. Harris (80,000 Hours Podcast) and Patel say that taking scaling seriously, then automating RL-driven search for better training algorithms, will keep increasing model intelligence.
10 claims, in their words
-
“we will scale up RLVR training across millions of diverse environments, across hundreds of different kinds of domains, and what will emerge at the other end is an agent which has like learned these basic skills or less than basic skills around being persistent, being able to triage information in context, eventually having like end-to-end optimization of working with other”
-
“if you're just distilling on like trajectories, you can never go above it, but environments can go quite a far away above what a human can do.”
-
“I definitely see and feel and believe in math more than probably the average person does. So if you show me a graph and I believe the data that gives rise to that graph and I can, you know, I can make a story for myself that like, actually this is, there, there are fundamental reasons why this graph exists,”
-
“Computing power isn't the same thing as intelligence. It's a bit like limiting how much fuel a car can use. That works until engineers figure out how to build a more efficient engine.”
-
“as you say, that heuristic is becoming less and less valuable. There's different scaling laws you mentioned. There's the post-training scaling law when you have more compute, when you're doing reinforcement learning on the thinking that it's doing after it's trained, you can get more intelligence that way. And there's this new scaling law, you know, the swarm scaling law,”
-
“there is still somehow alpha in just taking scaling seriously. So scaling is this idea ... if you want to make an AI system smarter, all you have to do, roughly speaking, is train it at a larger scale with more computing power and more data.”
-
“you can RL that model to get better at finding pre-training compute multipliers. For example, you can make an RL objective that GPT-7 has to minimize how much compute it uses to train a GPT-2 size model to a fixed pre-training class.”
-
“scaling up RL training will give AIs the abilities necessary to ramp up quickly in new contexts, even in hard-to-verify domains like the ones mentioned above”
-
“this kind of intuition is also trainable. For example, you can make an RL environment where GPT-7 has to predict whether an algorithmic improvement that it made at the 100 million parameter scale will also hold at the 10 billion parameter scale.”
-
“your automated researchers can try to improve your post-training algorithms such that if you want to RL on a new set of environments, it takes less compute or data or time, all three of which are easily measurable.”
Diminishing returns and practical constraints will prevent further large intelligence gains beyond human researchers.
O’Neill (Dwarkesh Podcast) cites a study where RL failed to move a model from roughly fifth-grade math and primary-school English to late high-school and college math competence. O’Neill argues that designing ever-harder RL environments and giving agents the compute and time to master them will hit diminishing returns, flattening capability gains. O’Neill claims that because new models mainly distill the last few months of mixed human and AI research progress, recursive self-improvement looks like a slow, asymptotic process.
3 claims, in their words
-
“there was a paper recently where they trained it up to like fifth grade maths. and like also like primary school, like English and stuff, so it was like a decent language model. And they tried to RL it to do like, you know, late high school and college maths, and the gap was just too large, like they couldn't”
-
“at some point there's diminishing returns in how hard it is to create those environments in the first place ... And then there's also going to be, like, the compute and time volume next for the agent to actually do those tasks. So, like, I think you're just going to start seeing this, like, curve to flatten out.”
-
“you're always Like just trying to get the last three months of progress. And that, that progress is being contributed to by AI, of course, but it also still has humans in the loop and it feels like, you know, you're just constantly inching closer and closer to what the human researchers are like finding and capable of doing.”
Listeners of “Scaling and better RL environments can keep driving large intelligence gains beyond human researchers.” have not heard
Listeners of these shows have not heard O’Neill's point that a recent RL experiment could not push a fifth-grade-level model to late high-school and college math performance. O’Neill also argues that the difficulty of constructing new RL environments, plus the compute and time they demand, will eventually cause capability gains to plateau. O’Neill further claims that because training pipelines keep humans in the loop, RL and distillation mostly capture recent incremental advances, making self-improvement feel asymptotic rather than explosive.
3 unanswered arguments, in their words
-
“there was a paper recently where they trained it up to like fifth grade maths. and like also like primary school, like English and stuff, so it was like a decent language model. And they tried to RL it to do like, you know, late high school and college maths, and the gap was just too large, like they couldn't”
-
“at some point there's diminishing returns in how hard it is to create those environments in the first place ... And then there's also going to be, like, the compute and time volume next for the agent to actually do those tasks. So, like, I think you're just going to start seeing this, like, curve to flatten out.”
-
“you're always Like just trying to get the last three months of progress. And that, that progress is being contributed to by AI, of course, but it also still has humans in the loop and it feels like, you know, you're just constantly inching closer and closer to what the human researchers are like finding and capable of doing.”
Listeners of “Diminishing returns and practical constraints will prevent further large intelligence gains beyond human researchers.” have not heard
Listeners of these shows have not heard Millidge's argument that rich RL environments, unlike pure trajectory distillation, can train AI researchers that outperform the humans who generated the original data. Rodriguez argues that compute caps resemble fuel limits on cars, because algorithmic efficiency gains can keep raising intelligence even when raw compute is constrained. Shapira and Harris claim that post-training and swarm scaling laws, together with continued large-scale training, let capabilities keep rising without proportional increases in training compute. Patel also claims that RL-trained automated researchers can iteratively improve training procedures themselves, cutting compute and time costs for future RL runs on new environments.
8 unanswered arguments, in their words
-
“there is still somehow alpha in just taking scaling seriously. So scaling is this idea ... if you want to make an AI system smarter, all you have to do, roughly speaking, is train it at a larger scale with more computing power and more data.”
-
“you can RL that model to get better at finding pre-training compute multipliers. For example, you can make an RL objective that GPT-7 has to minimize how much compute it uses to train a GPT-2 size model to a fixed pre-training class.”
-
“this kind of intuition is also trainable. For example, you can make an RL environment where GPT-7 has to predict whether an algorithmic improvement that it made at the 100 million parameter scale will also hold at the 10 billion parameter scale.”
-
“as you say, that heuristic is becoming less and less valuable. There's different scaling laws you mentioned. There's the post-training scaling law when you have more compute, when you're doing reinforcement learning on the thinking that it's doing after it's trained, you can get more intelligence that way. And there's this new scaling law, you know, the swarm scaling law,”
-
“if you're just distilling on like trajectories, you can never go above it, but environments can go quite a far away above what a human can do.”
-
“Computing power isn't the same thing as intelligence. It's a bit like limiting how much fuel a car can use. That works until engineers figure out how to build a more efficient engine.”
-
“scaling up RL training will give AIs the abilities necessary to ramp up quickly in new contexts, even in hard-to-verify domains like the ones mentioned above”
-
“your automated researchers can try to improve your post-training algorithms such that if you want to RL on a new set of environments, it takes less compute or data or time, all three of which are easily measurable.”
Position 0 leans on evidence that scaling laws and RL environments enable transferable skills and efficiency gains, which position 1 does not directly challenge. Position 1 relies on arguments about RL stagnation and environment or human-in-the-loop bottlenecks that position 0's scaling advocates leave unaddressed.
- Nov 14, 2026 Published: cleared the gate.
- Nov 14, 2026 Charlie O’Neill moved from “diminishing returns will prevent it.” to “Diminishing returns and practical constraints will prevent further large intelligence gains beyond human researchers.”.
- Nov 14, 2026 Beren Millidge moved from “RL can still exceed humans.” to “Scaling and better RL environments can keep driving large intelligence gains beyond human researchers.”.
- Nov 14, 2026 9 new claims (Doom Debates, Dwarkesh Podcast, 80,000 Hours Podcast).
- Oct 23, 2026 Beren Millidge moved from “diminishing returns will flatten RL gains.” to “RL can still exceed humans.”.
- Oct 23, 2026 3 new claims (Dwarkesh Podcast).
- Oct 22, 2026 Beren Millidge moved from “RL can still exceed humans.” to “diminishing returns will flatten RL gains.”.
- Oct 22, 2026 1 new claim (Dwarkesh Podcast).
- Oct 10, 2026 3 new claims (Dwarkesh Podcast, 80,000 Hours Podcast).
- Oct 9, 2026 3 new claims (Dwarkesh Podcast).
- Oct 7, 2026 Beren Millidge moved from “successive RL environments can enable AI to exceed human researchers.” to “RL can still exceed humans.”.
- Oct 7, 2026 Charlie O’Neill moved from “distilling on trajectories cannot exceed humans and effort grows exponentially.” to “diminishing returns will prevent it.”.
- Oct 7, 2026 4 new claims (Dwarkesh Podcast).
- Oct 4, 2026 3 new claims (Dwarkesh Podcast).
Published because: 3 shows, 2 positions, 63 days of debate. Questions publish only when claims come from at least three shows, at least two positions are held by different people, and the debate has lasted two weeks.
Pairs are matched by content across different episodes and shows; no exchange between these speakers took place unless both quotes come from the same episode.
The prose on this page is written by a model from the quotes shown, and every sentence of it rests on quotes you can open; sentences that could not be traced to a quote were removed before publishing. Positions are placed from each speaker's own words. Evidence labels say what the speaker cited, never whether it is true.
6 episodes behind this page
- 2026-11-13Dwarkesh Podcast — The avalanche that hit math is coming for AI development
- 2026-11-0580,000 Hours Podcast — An AI disaster hits. The president calls Xi. Then what? | Jeremie Harris (crosspost from NonZero)
- 2026-10-31Doom Debates — Is an AI Bubble Bursting? With Robert Wright’s Nonzero Podcast
- 2026-10-2780,000 Hours Podcast — Why Silicon Valley can’t slow down AI on its own | Luisa Rodriguez
- 2026-10-20Doom Debates — “AI Chernobyl” Will Scare Trump and Xi To Order An Immediate Pause — Jeremie & Edouard Harris
- 2026-09-11Dwarkesh Podcast — AI researchers debate how close we are to recursive self-improvement
The week's bottom lines, in your inbox
One email a week: the stories that mattered and what they actually mean. Free.