Everyone is talking about RSI right now.

Loads of people share this belief that we are on the verge of 'rapid takeoff' caused by 'Recursive Self-Improvement'.

I got nerd-sniped and spent a bunch of time researching the 'Economics of AGI', and recent work by Epoch and the Elasticity institute.

Very helpful for anyone else excited/terrified about AI progress.

The paper is kinda hard to understand - so I made a video here explaining: https://youtu.be/VQanDOFJwvo?si=9vcTCjDpt3OWYl_T

The TLDR:

Everyone has a different definition of RSI, and most of those definitions are unhelpful.

A more useful definition:

self-sustaining acceleration. AI capabilities accelerate AI progress without needing more human labour, training compute or data.

this is my crude interpretation;

Better model → more effective AI R&D → better algorithms → even better model.

Hold compute, data and human labour fixed.

Does a more capable model add enough to the R&D process to create the next, equally large capability improvement by itself? If consistently yes - we are on track for (potentially) rapid takeoff.

If no, we'll likely hit other bottlenecks - progress will be jagged etc

i go into this in the video,

basically -

Using the Epoch Capability Index, the paper calculates that every one-point increase in model capability would need to make AI R&D roughly 15% more productive to create a self-sustaining loop.

A simple way to think about it:

At 10%, a one-point capability improvement only generates about 0.65 points of further improvement. Without more compute, data or humans, progress eventually fades out.

At 15%, it generates roughly one more point. The loop can now sustain itself.

At 20%, potential slingshot. hold your fucking horses.

So where are we now?

The paper uses reported productivity gains inside frontier labs to produce a (very, very rough) estimate of 9% per ECI point.

That puts us below the estimated 15% threshold.

the feedback loop is not currently strong enough to generate self-sustaining acceleration—but it appears to be strengthening.

The problem is that we don’t have good data.

obviously - there are also lots of possible bottlenecks. AI might get very good at research while still being terrible at other tasks. Human ingenuity, compute or data might become the limiting factor. Progress might come in short, rapid spurts before hitting the next bottleneck.

would love to hear others thoughts on this.

anything im missing?

this is the paper: https://elasticity.institute/rsi-paper.pdf this is the eci: https://epoch.ai/eci?subset-view=graph&view=graph&tab=releas...

Recursive Self-Improvement (not yet). Also, we need a new definition