What kinds of computational functions can evolution locate to make minds out of? A dense, but well-motivated paper by Harvard CS professor Leslie Valiant shows an interesting negative result: there are actually some basic functions that evolution fundamentally can’t locate.
For my nerdier friends, the limiting, negative results that seem most immediately interesting show that evolvable functions are a strict subset of SQ-learnable functions. This would rule out evolution locating functions as basic as PARITY.
Incidentally, this suggests a straightforward and potentially useful research project for a motivated PhD student who understands Valiant’s work and MIRI’s open problems in Intelligence Explosion Microeconomics. Simply analyze intelligence explosion kinetics using the PAC-learning / SQ-learning framework and take into account Valiant’s more precisely specified theory on how limited evolution’s algorithmic reinvestment abilities are.
This research project would benefit from being mathematically rigorous so that it can extend and fully incorporate Valiant’s work, but a couple qualitative questions to answer include: Are there further updates in favor of mind design space containing much smarter minds than humans given that evolution can’t even locate basic functions like PARITY? How much faster can a system improve itself when it doesn’t have this sort of limitation?
See on dash.harvard.edu
