What I don't get is, ok, even granting the insane Eliezer assumption that LLMs can become arbitrarily smart and learn to reverse hash functions or whatever because it helps them predict the next word sometimes... humans don't entirely understand biology ourselves! How is the LLM going to acquire the knowledge of biology to know how to do things humans can't do when it doesn't have access to the physical world, only things humans have written about it?

Even if it is using its godly intelligence to predict the next word, wouldn't it only be able to predict the next word as it relates to things that have already been discovered through experiment? What's his proposed mechanism for it to suddenly start deriving all of biology from first principles?

I guess maybe he thinks all of biology is "in" the DNA and it's just a matter of simulating the 'compilation' process with enough fidelity to have a 100% accurate understanding of biology, but that just reveals how little he actually understands the field. Like, come on dude, that's such a common tech nerd misunderstanding of biology that xkcd made fun of it, get better material

[–] Evinceo@awful.systems 5 points 1 year ago

What’s his proposed mechanism for it to suddenly start deriving all of biology from first principles?

He considers deriving stuff from first principles much more versatile than it actually is. That and he really believes in the possibility of using simulations for anything.