Richard Socher

A good idea without good execution is not very valuable, but also like really good ideas are hard to come by.

And so what we currently have is really, really smart researchers who

who worked for many years in AI research, developing very strong intuitions on what is a good idea, what might result in an improvement, right?

And then they define sort of manually often data sets and benchmarks, and then the AI will charge after those benchmarks and try to get really, really good at solving those.

In contrast to that,

Our approach to recursive self-improvement is much more aligned with open-endedness, which is often an evolutionary process that essentially has a whole set of

interestingly different ideas that are being explored, often in parallel.

And like evolution doesn't really have a ceiling.

You don't just get 100 out of 100% correct on one particular benchmark and then you're done.

You can keep iterating in this open-ended fashion.

And so often in open-endedness, similar to biological and cultural evolution, entities interact with other entities inside an environment and those can co-evolve.

You know, like you have

different leaves like on trees and then you get uh you know they they get higher and higher so that the animals can eat the leaves but then giraffes you know get higher and higher next so they can eat the leaves again and so these like depending on the environment uh different eyes can evolve and so what what is the role of recursive in all of this and what's the first time that you think somebody outside of the technology industry will experience something that you've built

So the idea about recursive is that here the AI will output another AI that then becomes the input to the next iteration of that AI.

And at each step, that function will actually change itself.

So that's sort of the technical definition of recursive.

And the idea here is that we take the manual steps of ideation, implementation, validation out of that loop

and allow AI to come up with its own ideas, implement them, and then validate them.

And the first time people will observe this is likely in the AI research world itself.

You know, there are a lot of very, very expensive people.