The Lab Data Proves the Intelligence Explosion Is a Myth | Ramez Naam artwork

The Lab Data Proves the Intelligence Explosion Is a Myth | Ramez Naam

MTS

October 10, 2026

Investor and computer scientist Ramez Naam argues that recursive self-improvement and AI intelligence explosions face steep diminishing returns based on data from OpenAI and Anthropic system cards.

0:00

Ramez Naam

The thing that has never made full sense about the FOOM story to me, is that this feedback loop, if you have a better AI, it can make better algorithms, better chips, and that leads to an even better AI, right? That's the basic story. But most problems are not linear hard. Most problems are much harder than linear. So you have diminishing returns. And when you try to plug some of that data into the formal models that exist of a FOOM, what I found was that the current feedback loop, it could change, but the current feedback loop is about a factor of 5 to 10, too weak to sustain itself.

0:38

SPEAKER_2

All right, we are back. We're back with Ramez Naam, who is the founder and managing partner of Planetary VC, a computer scientist, investor, author who's written extensively about energy and technological progress. He recently wrote a guest post on the No Opinion sub stack called Where's the Intelligence Explosion?

If you've been in AI world for the last month or two, the vibes have shifted pretty dramatically. And it seems like the emergent consensus is like the labs are rushing towards quote RSI, whatever that means, which will lead to very, very rapid runaway capabilities growth, like on the time scale of the next year, kind of like what AI 2027 predicted. And we need a pretty major political intervention, pacing the frontier to ensure that this doesn't happen. So it seems like this has just been kind of taken for granted by everyone, except for you.

1:36

SPEAKER_3

So where is the intelligence?

1:37

SPEAKER_4

Yeah, where's the intelligence?

1:38

SPEAKER_2

Yeah.

1:39

Ramez Naam

So I wanted to title the article, Will RSI Hit Diminishing Returns? And Noah said much better article title than I am.

So the thing for me is, I've been writing about this for a long time. I was on the Extropians list, written sci-fi.

1:54

SPEAKER_2

Wow, yeah.

1:55

Ramez Naam

A long time ago. I've known Eliezer since we were both this big.

1:59

SPEAKER_2

Wow.

2:01

Ramez Naam

And the thing that has never made full sense about the FOOM story to me, is you have this feedback loop of you have a better AI, it can make better algorithms, better chips, and that leads to an even better AI, right? That's the basic story. But most problems are not linear hard. Those problems are much harder than linear. So you have diminishing returns. Like that cycle, you have to get one point of gain for every point that you put in. And what if you get less than that? And the real world is like that. So I went through some of the OpenAI data and some of the anthropic statements, because they both in the last few weeks put out these pieces saying, we are building, AI is building itself. And there's some shocking stuff because their data does not agree that RSI is anywhere near. So let's start with task horizon. The meter graph, Astra should be able to accomplish 11-hour tasks 80% of the time. Five, six souls should have a task horizon of four hours.

OpenAI measured their own people's work, and it was 15 minutes was the PAD.

Other examples of diminishing returns. OpenAI very proudly said, researchers are using 124 times more tokens per person. Okay, great.

We're getting seven times more lines of code written by a person. Okay, like A, who cares? And B, that's a smaller number. And we're running 1.6 times as many experiments per person. So all of this stuff has like steep diminishing returns. And when you try to plug some of that data into the formal models that exist of FOOM, what I found was that the current feedback loop, it could change, but the current feedback loop is about a factor of 5 to 10, too weak to sustain itself, even if there were no humans in the loop.

3:54

SPEAKER_2

So what would it take to get an actual self-sustaining feedback loop?

3:58

Ramez Naam

So one way to think about it is like, for every point of ECI that you get, how much more productivity does the system produce? Let's just say that the autonomous-

4:07

SPEAKER_2

ECI for the audience is EPOC Capabilities Index. Oh my gosh, yes.

4:10

Ramez Naam

I'm so AI-pilled, I'm just talking in acronym.

4:12

SPEAKER_2

And how do you measure productivity?

4:15

SPEAKER_3

How do you measure productivity in this case?

4:18

Ramez Naam

Progress it makes on improving algorithms. So actually, mostly they measure it in terms of computational efficiency. So if one point on this benchmark gives you 10 points of productivity, means that it can make code run 10 percent faster, essentially. So a guy named Tom Cunningham did this model, and he said that Tom is amazing.

Read the other 23 minutes of this episode here.

$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.

Not what you expected? Full refund within 14 days.

Thousands of transcripts fetched by people building searchable podcast archives

Fetch transcripts with code

The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:

request
curl -H "x-api-key: pt_demo" \
  https://spoken.md/transcripts/1000651996090

Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.

From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.

Using your own key:

request
curl -H "x-api-key: YOUR_KEY" \
  https://spoken.md/transcripts/1000794637677