Jared Kaplan - Human level AI by 2030? (Technical talk at physics conference Strings 2024)

[@kiankyars]

29:35 WOW!

[@MarkGubrud]

At around 5:30 Kaplan is saying contemporary AI is "Just curve fitting with a very general function approximation." I understand this is formally defensible but it's very reductive, like writing down the equations of physics and saying "This explains everything about international relations." Not really. To understand what is going on, you need to look at structure, both of the models and of what emerges from training.

[@MrBorest]

I disagree. Essentially, neural networks are just slices in the space of functions that are parameterized by floats, followed by SGD optimization with objective functions that, in most cases, aim to maximize the likelihood of some underlying statistical model. In other words, it is just curve fitting.

[@MarkGubrud]

@@MrBorest That is a very nice illustration of what I was saying, thanks. Ridiculous reductionism.

[@Williamtolduso]

wtf is this doing at strings?

[@BadWithNames123]

ai in 2027 will replace all of you.. it seems for the better because you guys are really obnoxious lol

[@Chase-b5t]

Because the strings conference gave up on talking about actual proper string theory quite a while ago. Most talks are tangential to string theory at best and completely off-topic in many cases.

[@studyphysics8813]

Lmao. Thanks for saying this.