Will increasing the scale of current-style AI models keep yielding large capability gains without major slowdown in the near future?
The dispute is whether scaling up current-style AI models with more parameters, data, and compute will keep yielding big capability gains soon or start to stall. Over the past two months, Cotra (80,000 Hours Podcast) has argued that bigger models continue to unlock new abilities in following instructions and constraints. In the same period, Greene (The Diary Of A CEO) and Huang (Hard Fork) have warned that adding more data and compute to current models may soon hit hard limits.
Yes, continued scaling of current-style models will keep providing large capability gains soon.
Cotra (80,000 Hours Podcast) says that when she reruns earlier researchers’ instruction-following tests on newer, larger models, these larger models often solve the problems without extra techniques. She argues that evidence from recent model generations shows simple scaling improves adherence to side constraints more than bespoke alignment methods do. She claims that getting models to follow user instructions is already straightforward and becomes easier as models increase in size.
3 claims, in their words
-
“I still see a lot of machine learning researchers that are focused on addressing what feel more like capabilities gaps than alignment gaps that do help the models act better, but are more like, oh, our systems aren't very good at following instructions. Let's kind of put them in a setup that makes them better at understanding instructions. And the”
-
“we have some pretty strong evidence that just making the models bigger helps with this. And in fact, I often get proposals actually from people who are like, models aren't able to do this kind of thing. So they're not able to follow side constraints very well. And there'll be some examples. But then when I run those exact questions”
-
“I think we're in a world where it's pretty easy to get them to do that, and as they get bigger, it's even easier.”
No, scaling current-style models will soon face strong diminishing returns and limited capability improvements.
Greene (The Diary Of A CEO) argues that pouring more money, data, and compute into current language models will not produce unlimited further improvement. O’Neill (Dwarkesh Podcast) claims that at today’s capability frontier, the world contains few remaining bits that are highly relevant for pushing models into a higher capability basin. Huang (Hard Fork) states that models do not keep improving just by training longer with more compute and data, pointing to the need for a second scaling law.
3 claims, in their words
-
“There Isn't necessarily an unlimited capacity of that kind of artificial intelligence to continue to improve, right? And so it's not obvious to me that it's just a matter of pumping billions and billions of dollars, building ever larger data center, putting them in space, whatever you're going to do, and the more money you put in and the more”
-
“the capability frontier of where the models are at now, like how many bits in the world are actually like really relevant to like Improving the model's capabilities ... I think that's why the diminishing returns kicks in, because like even the world as a whole is not giving you the bits that are useful for tipping you into the”
-
“It is not true. It is not true that we just keep training these models to get better. notice it is the reason why the second scaling law had to come along.”
Listeners of “No, scaling current-style models will soon face strong diminishing returns and limited capability improvements.” have not heard
Listeners of these shows have not heard Cotra's view that newer, larger models often fix specific instruction-following failures without needing the special techniques earlier researchers proposed. They have not heard her claim that scaling model size improves adherence to side constraints more effectively than targeted alignment approaches. They have not heard her statement that getting models to follow instructions is already straightforward and becomes easier as models grow.
3 unanswered arguments, in their words
-
“I think we're in a world where it's pretty easy to get them to do that, and as they get bigger, it's even easier.”
-
“we have some pretty strong evidence that just making the models bigger helps with this. And in fact, I often get proposals actually from people who are like, models aren't able to do this kind of thing. So they're not able to follow side constraints very well. And there'll be some examples. But then when I run those exact questions”
-
“I still see a lot of machine learning researchers that are focused on addressing what feel more like capabilities gaps than alignment gaps that do help the models act better, but are more like, oh, our systems aren't very good at following instructions. Let's kind of put them in a setup that makes them better at understanding instructions. And the”
Listeners of “Yes, continued scaling of current-style models will keep providing large capability gains soon.” have not heard
Listeners of these shows have not heard Greene's concern that scaling money, data, and compute in current models may hit limits and fail to keep improving them. They have not heard O’Neill's claim that near the current capability frontier, the world offers few remaining highly relevant bits for improving models into the next capability basin. They have not heard Huang's assertion that more compute and data alone do not keep making models better, as illustrated by the need for a revised scaling law.
3 unanswered arguments, in their words
-
“It is not true. It is not true that we just keep training these models to get better. notice it is the reason why the second scaling law had to come along.”
-
“There Isn't necessarily an unlimited capacity of that kind of artificial intelligence to continue to improve, right? And so it's not obvious to me that it's just a matter of pumping billions and billions of dollars, building ever larger data center, putting them in space, whatever you're going to do, and the more money you put in and the more”
-
“the capability frontier of where the models are at now, like how many bits in the world are actually like really relevant to like Improving the model's capabilities ... I think that's why the diminishing returns kicks in, because like even the world as a whole is not giving you the bits that are useful for tipping you into the”
The scaling-optimistic side leans on observed improvements from larger models in instruction following, which the slowdown side never directly addresses. The slowdown side emphasizes potential limits from finite relevant data and non-linear scaling laws, which the scaling-optimistic side does not engage.
- Oct 14, 2026 4 new claims (Moonshots, Dwarkesh Podcast, Doom Debates).
- Oct 9, 2026 Published: cleared the gate.
- Oct 9, 2026 11 new claims (Dwarkesh Podcast, Machine Learning Street Talk, 80,000 Hours Podcast).
- Oct 8, 2026 2 new claims (Moonshots, Dwarkesh Podcast).
- Oct 7, 2026 Professor Brian Greene moved from “Yes, scaling will produce unlimited continued improvement.” to “No, scaling will not necessarily produce unlimited continued improvement.”.
- Oct 7, 2026 Jensen Huang moved from “Yes, scaling will produce unlimited continued improvement.” to “No, scaling will not necessarily produce unlimited continued improvement.”.
- Oct 7, 2026 3 new claims (Hard Fork, Doom Debates, The Diary Of A CEO).
Published because: 4 shows, 2 positions, 52 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
- 2026-10-0880,000 Hours Podcast — In 2023 Ajeya Cotra already knew what was coming (classic episode)
- 2026-10-07Moonshots — Why Altman Says "Accept Some Bad Things Happening" and What It Means for AI Safety | MOONSHOTS #301
- 2026-09-25Hard Fork — The Ezra Klein Show: Jensen Huang Thinks A.I. Alarmism Has Gone Too Far
- 2026-09-23Doom Debates — Will AI Kill Everyone by 2050? Debate with Dr. Casey Hart (Ontology Explained)
- 2026-09-11Dwarkesh Podcast — AI researchers debate how close we are to recursive self-improvement
- 2026-08-17The Diary Of A CEO — Top Scientist REVEALS: We Invented Heaven, God Is A Human Invention! | Prof Brian Greene
The week's bottom lines, in your inbox
One email a week: the stories that mattered and what they actually mean. Free.