Embed Notice
HTML Code
Corresponding Notice
- Embed this notice
pistolero (p@fsebugoutzone.org)'s status on Thursday, 20-Feb-2025 08:22:30 JST pistolero
@maxmustermann @hakui Well, no brain in a jar, right; but human intelligence is an implementation of intelligence.
You can look at https://top500.org/lists/top500/2024/11/ and we have these massive machines. And you can look at the systems all the way back to 1993: https://top500.org/resources/top-systems/ . And the ones not used for nuclear simulations are mostly concerned with weather, because no matter how much power we throw at weather, we haven't managed to crack it yet and weather predictions more than a couple of days in the future are a crapshoot. Even something big, like a hurricane, sometimes it just decides to zip off. Tiny rules in systems much smaller than a thunderstorm end up moving the thunderstorm, right?
And you've got your liver and all of the various organs slushing around in there producing molecules that land in your blood and have molecule-level interactions with the rest of the system. Holding a gun raises testosterone levels, and testosterone levels affect cognition: how complex is that, if you try to break it down to molecules? You perceive a gun in your hand, something lights up in your visual cortex, you recognize the gun, and eventually gets down to the testes and the testosterone gets dumped into the bloodstream and makes its way up and binds to individual receptors in individual cells, and this just from slushing around in there: maybe it passes this cell because it was facing the wrong way and it lands in the next cell. At the molecular levels, it's really complex. A neuron fires or doesn't and then whether that even arrives at another neuron is dependent on how hard it fires and the length of the cable and the insulation (myelin) and a neuron alone is already hard to simulate, and it's feeding back into all of these other systems: your blood sugar drops, or a species of bacteria that manipulates the Vagus nerve (this actually exists, we found it) starts tapping that nerve, and you feel hungry. Weather is easy by comparison, and we're just barely able to do a day or two ahead, something an experienced farmer or sailor can do just by looking at the sky. So simulating human intelligence is not close; we'll have to hollow out the moon for the datacenter (and we'll still have some difficulty getting rid of heat) and then we can do maybe one guy. Currently, we can not yet get 80% accuracy for simulating a microscopic roundworm with 1,000 cells in its body, caenorhabditis elegans: https://github.com/openworm/OpenWorm/milestones .
If we come at it from the other angle, we can probably build a system with some kind of intelligence, but it'd have to be a state-level project, it'd take decades. It's not like two guys in a garage making a breakthrough, or even a trillion-dollar company funded by half of Silicon Valley, it'd be another space program. (We'd be better off trying to get cold fusion going first if we're talking that level.) Otherwise we'll occasionally get incremental advances like LLMs and they will occasionally catch the crest of the hype cycle and make some liars rich(er).