Ramez Naam warns recursive AI self-improvement faces diminishing returns

Ramez Naam warns recursive AI self-improvement faces diminishing returns · Digg

Commentary on X

@ramez >Just not seeing a Singularity on the horizon. Singularity is a red herring, for idiots. It's over for humanity even if there's no singularity, because once there are drop-in replacements for employees, from the PoV of the government there's no real need for people.

3:25 AM · Jul 15, 2026

mattmac@MatthewMac58111

@ramez yet those working all day with coding models intuitively see an effectiveness multiplier that "feels more order of magnitude" - I think they're more correct than this nano-slice-of-a-gnats-whisker "data

5:36 AM · Jul 15, 2026

Greg@gentschev

@ramez This is great work, and I agree with you based on current data. But it does seem like it might eventually cross an “it works” threshold the way coding did in late 2025.

3:20 AM · Jul 15, 2026

Virgil Maro@_virgil19

@ramez a 13th root exponent is the least cinematic way anyone has ever described the singularity slowing down. sounds more like a tax bracket than an apocalypse

8:10 AM · Jul 15, 2026

View all

Firas Hermez@FirasHermez

@ramez I suspect there is a soft limit but there is still plenty of room to expand beyond what our biological brains are capable of 🤔

12:45 AM · Jul 15, 2026

Faye Rae@fayeray331

@ramez any1 who is using ai for ml experiments like me could have told you that. severe lack of taste

6:38 AM · Jul 15, 2026

Tom Davidson@TomDavidsonX

This is an interesting paper, but it tells us very little about whether we'll get a software driven intelligence explosion. 1. In this paper ai can only edit one part of the stack: the scaffold. No hardware optimisation or post training or pre training etc. 2. ai today is clearly much less capable than human researchers! The interesting question is not whether AI is capable enough *today* to drive an intelligence explosion. Everyone knows that it can't! It is whether it will happen in the future when AI can automate all research. (Also the paper massively overhypes itself by claiming it's the first experimental evidence about recursive self-improvement. There have been many papers where agents have tried to design better scaffolds for themselves before. This is not the first.)

7:31 AM · Jul 15, 2026

View all

Firas Hermez@FirasHermez

@ramez I suspect there is a soft limit but there is still plenty of room to expand beyond what our biological brains are capable of 🤔

12:45 AM · Jul 15, 2026

Faye Rae@fayeray331

@ramez any1 who is using ai for ml experiments like me could have told you that. severe lack of taste

6:38 AM · Jul 15, 2026

Tom Davidson@TomDavidsonX

7:31 AM · Jul 15, 2026

Ramez Naam@ramez

Agree with Tom here. I don't consider this RSI experiment to be an end point. I expect better and more capable ones in the future. I do think it's worth noting the power-law diminishing returns from the feedback loop they implemented, which are in line with expectations that RSI broadly faces diminishing returns.

7:37 AM · Jul 15, 2026

Ramez Naam@ramez

@TomDavidsonX I believe that's only true in models that assume a Cobb-Douglas production function. And AI improvement isn't Cobb-Douglas, either in ML theory or in our limited empirical data. Not all models make that assumption, btw. Those that don't fail to find a software only singularity.

10:25 AM · Jul 15, 2026

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

> The AI model advances at about [Intelligence]^0.075, or the the 13th root of input intelligence [1,2]. no, it ain't gonna be like that

9:24 AM · Jul 15, 2026

Tom Davidson@TomDavidsonX

@ramez True. Though fyi all models of the intelligence explosion include power-law diminishing returns, but often it can be overcome by the increasing capability of the AI systems! See ~any paper on the intelligence explosion for an example!

9:44 AM · Jul 15, 2026

Daniel Jeffries@Dan_Jeffries1

@ramez Like all exponential curves, it's just a hidden S curve.

3:29 AM · Jul 15, 2026

Combined views

Commentary on X

Combined views

Sentiment

Positive Read

Digg Deeper

Commentary on X

Sentiment

Positive Read

Digg Deeper