Why Recursive Self-Improvement Hits a Wall | Ramez Naam artwork

Why Recursive Self-Improvement Hits a Wall | Ramez Naam

MTS

November 11, 2026

Planetary VC investor Ramez Naam pushes back on runaway recursive self-improvement models and short AGI timelines, arguing that empirical lab data shows steep diminishing returns rather than exponential feedback loops.

0:00

Ramez Naam

I want to emphasize again, all of these runaway RSI models are based on theory. They're based on math people have written down.

We need to actually calibrate that against what actually happens in the real world. And with the data that we have, when you do calibrate it, the real world effect of AI on improvement of itself is much, much slower and has a much harder slope, much more diminishing returns than you get in the theoretical models that people have written down.

0:34

SPEAKER_2

And we are back. We're live with Ramez Naam, who is an investor at Planetary VC. And a second time guest, a skeptic of the whole RSI narrative that we have been seeing all over the timeline today. So it's going to be great to talk about this. Ramez, welcome back to MTS.

0:54

Ramez Naam

Thanks. It's great to be back.

0:56

SPEAKER_2

So, I don't know if you saw the AI Futures Project blog post from this morning and, you know, the associated commentary, but, you know, people are really kind of freaking out at this point. AI Futures Project is now saying it, all of their members are predicting that we get a superhuman AI researcher by the end of 2027, by at the latest November 2027

And they're predicting that as a result of that, we get a top expert dominating AI by somewhere between late 2027 and late 2028 In other words, an AI that is better than all humans at all white collar work. So what do you think?

1:39

Ramez Naam

Well, first, like, I got to respect anybody that puts numbers down and the AI futures folks have consistently, like, made a model, made a forecast and written it down.

Second, I'd say people are freaking out and we are human beings, we get emotional.

Third, I'd say it looks to me like we are seeing acceleration of AI progress. It is not clear to me that that acceleration is coming from, how much of that acceleration is coming from AI itself versus more investment of resources. We have to bear in mind that the landscape of this, the number one driver of AI improvement is that we are putting exponentially more resources into AI, exponentially more venture capital, exponentially more hardware CapEx, exponentially more data, exponentially more talent. I tweeted yesterday a chart showing that in overall R&D spend, Anthropic has a clear logarithmic scale, meaning anti-exponential diminishing returns. Every point of ECI gain costs them much more than the last one. So that's not to downplay progress. And Anthropic in particular has made substantial progress. It threw much of 2025, they were actually behind the frontier on ECI. So they have come from behind some extent and made faster progress than the market as a whole to now be in the lead until, at least until we see the next OpenAI release.

At least some of that, I would say, is from AI.

In the simple sense that everyone uses AI for coding. Nobody really writes code by hand anymore, very few of us do. And that is reducing friction. If you read from the meter report that came out on Thursday last week, there are interviews with people inside Anthropic. And I normally, I think interviews are far less robust data than just data, than just measurements. But what those interviews suggest is not, nobody in the set of people they talk to says that AI is doing an amazing job picking research directions or anything like that. They say it's reducing the friction. It's easier if you have an idea to test it out. It's easier to run a sweep of experiments across like a matrix of 16 possibilities. So it's reasonable to expect that to uplift somewhat. But if we go back to that simple relationship between log of money spent per period and pace of gains, there's no obvious break in the line. There's no obvious trend break. So that still seems like the dominant factor driving progress forward. And that probably has limits.

4:50

SPEAKER_2

Well, Benjamin Todd from 80,000 Hours countered your post by pointing out that higher levels of ECI are exponentially more valuable because Claude with lower ECI could earn very small amounts of revenue. Claude with higher ECI, in other words, more capable Claude, is capable of earning a lot more revenue. Yes, there's some confounding factor here in that there's also a lag in terms of how quickly Claude can grow revenue. So presumably, if Claude had not progressed in fundamental capability, its revenue would have still grown just through diffusion. However, I think that the rate of revenue growth from Claude is probably pretty clear that this is largely due to capabilities growth. And if we only had Claude won from 2023, it wouldn't be earning tens of billions, high tens of billions of revenue.

Read the other 17 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/1000801252928