Self-improving systems

I think that recursive-self-improvement is an attractor state among all systems that have the ability to improve themselves. Specifically, it's an attractor state among systems which have persistent, general improvements. However, that is still quite a specific criterion.

Human Knowledge

Human knowledge fits this, both across history and within individuals' lifetimes.

If you want to take advantage of this effect in your life you should make changes to your life which in turn make it easier to make changes to your life.

Natural Selection

I think it's actually unclear for natural selection, because improvements that are general (say, greater genetic diversity in a population) aren't necessarily persistent (because the greater diversity makes a population less specialized for the niche and get out-competed). And vice-versa.

The key differentiator between natural selection and human knowledge is that humans can easily share beneficial knowledge across space and time, but natural selection can't easily share beneficial genes across space and time. Sharing across space requires that a species move/spread (because genes can't be shared across species), and sharing across time requires that the genes survive over time.

As a side-note, this is personally where I put largest probability on great filter. (The resolution to the Fermi paradox.) IMO natural selection wasn't destined to compound it's gains to the point where a sub-system started doing recursive-self-improvement. There was life on the planet for billions of years, but it didn't lead to recursive-self-improvement until recently. That is, it took a long time comparative to the current age of the universe for natural selection to get to the point where it led to recursive-self-improvement.

Artificial Intelligence

I think AI systems aren't there yet, (as of late 2022). However, AI improvement (since ~last year) went from being fed solely by a "system of human knowledge" to a "system of (human-knowledge + AI)", with progress in AI now contributing to progress in AI, but simply still via human-in-the-loop.

  • The most obvious instance of this is GitHub Copilot (I've gotten used to working with this now and it's very useful). (Well, I say this but it's not doing a very good job of it right now, writing this blog post). Eg. In the below article:

Louis Castricato uses Copilot to improve his ability to write Einsum expressions

  • Also, good text-to-speech for listening to content. Although most people don't use this in a way that contributes to AI progress, it's likely that many AI researchers use this to consume AI related content. In a similar vein, maybe speech-to-text in youtube subtitles, and better reccomendations for paper and knowledge discovery.

  • Lastly AI is contributing to improved compute capacity via better chip designs, and stuff like the recently announced AlphaTensor for discovering better algorithms for tensor multiplication.