Wednesday, January 19, 2022

Diminishing Returns on IQ: Autonauts Example & Musings on IQ Evolution

Playing an alpha version, 21.2, Tito. The release version is plain bad, IMO.

It's a bit like Factorio except you can have a fully automated and developed base in an afternoon. Offers some interesting challenges...or would if bots were expensive. Instead you can automate bot production in like 20 minutes, and it's not like they cost power to maintain.

Because the programming language is rather restricted and bots can have at most 32 instructions, the game provides lots of interesting challenges related to trying to make a bot relatively intelligent without running out of memory. By default it doesn't even have a proper AND gate and you have to hack something in, which I find a satisfying puzzle. can just make more bots. They're cheap and power costs are negligible even when they aren't self-powering. Whatever interesting programming challenge there is, it's trivialized when you break it up into two bots. Anything that almost fits into 32 will trivially fit into 64, let alone 128. It takes far longer to try to fit it into 32 than to just build a new robot, even if you do it by hand instead of using the automation. 

I just do it anyway - it's a game after all, you're suppose to play it, not work it - but the challenge is entirely self-imposed. "What if they made this into an interesting game instead of polishing all the fun out? Let's play around in preparation for that." It's still an alpha, so technically it was, at the time, still possible...

I believe this is the basic reason that most animals don't get into runaway IQ evolution. 

Similarly, an alleged singularity computer would self-limit due to myriad factors long before it became infinitely intelligent, before it became transcendentally intelligent, or even all that exceptionally intelligent. (Indeed Descartes was right, so it would stop even before it became intelligent at all unless it was attached to a whole arbitrary laboratory/machine shop that could re-design and re-build its hardware.) No matter how smart a computer is, all it takes is one person who doesn't think it's in their best interests to keep it plugged in, and ¯\_(ツ)_/¯.

Animals in particular see the exact dilemma I see in Autonauts. Why bother making an Einstein squirrel when you can make two or three or fifty squirrels for the same price? Why bother figuring out the right solution in advance when you can try literally every solution just by being dumb, making mistakes, and stumbling into all of them? You have to choose the path that costs the least, and the diminishing returns on IQ investments are so harsh that the dumb path is cheaper. 

I'm firmly convinced human IQ is a liar-liar arms race. Pretty sure the oldest known fire isn't the oldest fire, and habilis became erectus exactly because they discovered fire. (Or had it revealed to them.) After erectus, total ecological dominance was achieved. The only thing that can seriously threaten a grass monkey is bad mutations or other humans. Further, the dire apes you spend all your time around are ones you're not allowed to kill. Sure you can try to kill the next tribe over but there's a distinctly limited supply of vulnerable neighbours. Which means: you need to lie, and defend yourself from lies.

This IQ ratchet is self-limited, because if your lie is too sophisticated it goes above your victim's head. The sweet spot seems to be around 25-30 points of IQ advantage. Plus you still get diminishing returns. No matter how cleverly you argue, you can't convince someone their children don't deserve to eat. "No, uh, getting food for my kids is the entire point of this interaction." Unless they manage to forget sanity, you'll be SOL no matter how many times you try that one.

Bonus diminishing returns round: now humans have a nice powerful general processing unit, they have to learn to use it. The whole point of a high IQ is to do things your genes don't know how to do, which means doing your own training. 

In other words, instead of seeing a local maximum at [just barely smart enough not to bite its own tail off] the grass monkey sees a local maximum at [just barely smart enough not to sell its own children for scrap metal].

Consciousness pre-exists evolution. Like sunlight and carbon, evolution simply harnesses it for its own purposes. Like sunlight and carbon, there are side-effects and compromises that have to be made. For example, humans don't like dying and thus don't like the [throw bodies at it] solution anymore. Unfortunately, thinking is expensive even if you're really smart, so there's a bit of a catch-22 there. Smart enough to work it's possible to be smart enough not to die, not smart enough to figure out how, not smart enough to work out they're not smart enough and need to accept the consequences.

No comments: