RSI cannot replace the data and experiments a problem requires

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 14:45

   Do you think that RSI will just, you know, many people think RSI will just solve all of LLM's current weaknesses, solve jaggedness. Do you think that this will happen? I mean, I have to think yes. If I were one of the Frontier Labs, one of the first directions at which I would point RSI would be some kind of novel model architecture that is better than the transformer maybe more efficient maybe more human-like maybe more energy efficient but it feels like we kind of stumbled on the transformer architecture right and we've been using it ever since and it's not clear but it's the final thing um and it seems that to the extent there are limitations with it it's going to be the automated ai researcher that'll be able to discover it um and so you know it's it's kind of it's hard to say um whether that's low-hanging fruit or not it's possible that it might take

2. 15:54

   years for even a couple a couple of years of models to become creative enough to discover or something really But I think it's a possibility. I would not be surprised if in five years' time, the AI architecture that everyone was used were completely different and alien from what we're used to right now. Yeah, but I don't understand how this would magically solve jaggedness. It seems to me that a lot of the reasons that LMs are jagged are because of data bottlenecks.

3. 16:24

   So like something like bio, for example, it seems like, you know, maybe not completely impossible. I mean, in AI, you can't rule things out, but pretty unlikely that an AI would just be able to like solve its constraints in biology just by doing RSI on a computer, whatever that means. no i think that's exactly right um look like if you took albert einstein and you put him in a lab and said like you need to cure cancer like tomorrow um he would not be able to do that even if he had the knowledge right like he wouldn't be able to do that without having the right data the right experiments right um but i don't i don't know if that's um evidence of jaggedness i think it's evidence but sometimes a brilliant intellect whether human or artificial needs to be given um the you know the right environment and the right data

4. 17:26

   um but on the other hand for example um i'm a lawyer right so if i think a lot about llms and their ability to like write a good legal research memo right for which you don't need albert einstein but llms right now we're not particularly good at it they're becoming better but they're not incredible and i don't really think it's about the data i think it's just because llms are not as efficient um at learning some of this this non-verifiable stuff as humans are um and uh one answer may be that in a year they will be really really good at it regardless but another answer may be that a better architecture will just make them more human-like and learning and then an llm will be able to learn to write a legal memo just as well as a law firm associate can do that we'll continue monitoring right after this message from our