I doubt genetic evolution is very efficient. But it’s hard to know which optimizations have what downsides.
Aggressive pruning every generation, and a small gene pool, seems more likely to get stuck at local optima, which is already a problem for animals.
Having 1, 2 or 3+ sexes is important too.
yeah, i think the lack of a direct optimizer may matter. the question isn’t just whether evolution could be more efficient, but efficient at what.
this reminded me of a point from deutsch in deutsch files iv: he says ribosomes are the educational institutions of cells, and that their job is to pass on cellular knowledge as faithfully as evolution can make it. i might be misinterpreting, but that connection seemed relevant.
i take the lesson as: evolutionary processes don’t just favor more experimentation. some parts of the process can be selected to reduce variation and preserve a pattern accurately. that can be good if the pattern is already valuable, but it’s different from open-ended creativity.
so maybe there are several separate issues:
that makes me less inclined to treat more direct fitness measurement as automatically better. it might make some engineering searches more efficient relative to the metric, but it could also make the process narrower depending on what the measure rewards.
checkout Primer youtube channel. he makes animated simulations/explainers about evolution, game theory, emergence, cooperation, etc. a lot of them are toy models, but still interesting.
liked this one on evolution of teamwork.
I think just knowledge creation. It’s not a very satisfying answer (doesn’t give us a fitness heuristic or anything like that).
PS. am familiar with primer. I like his stuff mostly. I think I’ve had an issue with one or two things in the past but that’s like 1-3 years ago so don’t recall specifics.
knowledge creation does seem like the right target.
the tricky part is that a fitness heuristic for knowledge creation might already need a lot of the thing we’re trying to explain. it has to tell real explanatory progress apart from local proxy wins.
so maybe the lack of a direct optimizer matters. direct scoring can make search faster against a metric, but if the metric is narrow, it can also prune away weird/open-ended paths too early.
that connects to the AGI/creativity issue for me: variation+selection by itself doesn’t explain much unless we say what is doing the selecting, pruning, and relevance-judging without turning it into a narrow proxy.
I’m curious, what do you think knowledge is?
Re ‘direct scoring’ and ‘metric’ – have you read Multi-Factor Decision Making Math ?
I basically agree with David’s definition of knowledge as information that keeps itself in existence, but I don’t think that helps us much with the thing we are interested in here which is explanatory knowledge and the processes that create it.
I’ve read it. As I understand it, the core point is that you can’t solve problems just by using weighted scoring. I agree with that. But I don’t think MCDM gives the missing fitness heuristic for knowledge creation. IGC seems like a useful reframing. It is better than evaluating ideas in isolation, but the hard part still seems to be judgment: choosing the goal and context, choosing the breakpoint, and deciding whether a criticism really refutes the idea in that context. So as I see it, the judgment problem is still there, just one level down.
Interesting reaction to AGI discourse. BoI is mentioned throughout.