Why AI Math Breakthroughs Upgrade RSI Timelines | Cheryl Wu artwork

Why AI Math Breakthroughs Upgrade RSI Timelines | Cheryl Wu

MTS

October 8, 2026

Cheryl Wu, President of the Elasticity Institute, discusses measuring recursive self-improvement (RSI) in AI models, differentiating self-sustaining acceleration from an intelligence explosion, and analyzing OpenAI's math breakthroughs and Tasteval results. Turn ideas into software people love.

0:00

Cheryl Wu

I think it's definitely updating me toward having a region where RSI could happen, or in our own term, like self-sustaining acceleration could happen. It just updates me toward more likely.

If AI can do pretty well in math, it could probably also do pretty well in AR and D. And yesterday, there's also another post about Tasteval, for example. Tasteval says that AI started to get better at doing AR and D experiments and have better taste of them and also probably exceeds a AI researcher outside of lab in terms of taste.

0:40

SPEAKER_2

All right, we are back. We are live with Cheryl Wu, who is the president of the Elasticity Institute, a group of economists studying the economics of transformative AI. And probably my favorite people doing AI math or AI econ stuff right now.

0:55

Cheryl Wu

Thank you.

0:56

SPEAKER_2

So Cheryl, welcome.

0:57

Cheryl Wu

Thank you.

0:59

SPEAKER_2

I guess we'll start with the news from yesterday, the OpenAI math nuke that was a sort of a Cretaceous event for mathematics. They announced 722 new results across all kinds of fields, including like a more efficient matrix multiplication algorithm and progress on the quasi-Riemann hypothesis.

A whole bunch of other things. So how does this affect your timelines? What are your takes on this?

1:32

Cheryl Wu

I think it's definitely updating me toward having a region where RSI could happen, or in our own term, like self-sustaining acceleration could happen.

It just like updates me toward more likely because, yeah, if AI can do pretty well in math, it could probably also do pretty well in AR&D. And yesterday, there's also another post about the Tasteval, for example, and that also kind of, that, sorry, Tasteval says that AI is, AI started to get better at doing AR&D experiments and have better taste of them, and also probably exceeds a AI researcher outside of lab in terms of taste. So I think that also kind of updates me toward having the possibility of RSI has increased for me.

2:33

SPEAKER_2

Well, how easy is it to transfer from theoretical math to advances in deep learning or AI R&D? Like the Tasteval results from yesterday seem to be like not the hardest tasks to do. It was more like experimental design and whatnot.

So it's like if the models can write code for the most part, if they can now increasingly design experiments, like there's still tasks left for humans to do, like deciding which hills are worth climbing in the first place. And so like I expect this to accelerate AI progress, but this does not seem to me to be strong evidence for imminent like fast takeoff.

3:17

Cheryl Wu

Oh, okay. Yeah, I think when I say RSI, I meant more like there could be a self-sustaining acceleration, meaning that you could fix the other inputs, but with more and more AI progress, you could have acceleration in capabilities. So I think that those math results and Tasteval kind of upgrades me toward there will be some type of acceleration in capabilities, although it doesn't...

I'm less updated on whether there will be a fast takeoff or whether there will be a so-called intelligence explosion.

3:55

SPEAKER_2

Right. I do think, and I tweeted about this the other day, people are increasingly lazily conflating RSI with FOOM or like intelligence explosion.

4:05

Cheryl Wu

I kind of just did that, sorry.

4:07

SPEAKER_2

Yeah, no, I mean, like, definitely like you guys, less than most people. But I feel like when a lot of people hear, like, OpenAI is going to hit RSI, Anthrobic is going to hit RSI in the next few months, to them, that means intelligence explosion and then like very imminent super, like for real super intelligence. And I think that that is like much less likely than I think a lot of people are default assuming.

4:31

Cheryl Wu

You mean the labs think of explosion more? I feel like their definition have been, yeah, have been like fully automating the R&D process. And the data points they put out is more on, you know, cloud has been assisting 26% of or leading 26% of all the R&D tasks, less about like whether there's acceleration or things like that.

4:59

SPEAKER_2

Less so the labs themselves, I think, and more so the discourse that has spawned around it.

5:03

Cheryl Wu

Yeah.

5:03

SPEAKER_2

Which like you might think that the Twitter commentary at what they believe doesn't really matter. And in some sense, that's true. But in another sense, it's upstream of a lot of societal responses, political responses.

5:16

SPEAKER_3

And so it matters a lot.

5:17

Cheryl Wu

Read the other 16 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/1000793852113