AI progress still faces organizational and feedback constraints

Transcript

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

1. 28:26

   That's assemblyai.com slash S-O-U-R-C-E-R-Y. So you're based in Silicon Valley. Do you believe the whole six months to singularity thing? What's your take on the whole singularity? Six months is a bit short. I mean, so look, the recursive thing, here's what I'd say. there is obviously some version of the recursive thing which is literally i mean i even just said that like devin writes most of the code of devin so it's like you know on the other hand i think the question of you know whether um um you know whether humans will be rendered irrelevant and all human like all science problems and all societal problems will be solved you know six months from now with the ultimate intelligence i don't quite believe that um um and and i think there's a few reasons for why that is i think for one there's a lot of practical problems out there

2. 29:21

   in the world that are not actually you know necessarily like uh purely intelligence soluble right there's just like organizational things that have to be done there's like um uh there's processes that take time that have to be figured out there's like hardware components of like how much g you know how much gpu compute can you get and and so on and so i think all those things obviously exist for two i think we see more and more um that you know in in the model training landscape i think the vibe today is almost like you can teach the model pretty much anything as long as you know exactly what you are trying to measure or what kind of outcome you're trying to get out of it and so um benchmark scores for example are obviously every time there's new model people release benchmarks the thing that's kind of almost like um counterintuitive

3. 30:11

   is we're kind of getting to the point where you can solve basically any benchmark right because what does it mean to have a benchmark it means you've already defined the task you've clarified what success or failure looks like you've given a bunch of examples of what that task looks like and what the right behavior should be and the truth kind of is well if you do that then you can teach the model to go do that you know it's rl works it's it's it's it's incredible how much it works right um but obviously not all of the tasks of the world today cleanly fit into that you know description right and then there's lots of things that are you know um that are super super amorphous that don't really have clear definitions that it's hard to say what counts as success or failure that have like really long time horizons where it takes a long time to find

4. 30:56

   out whether it was the right decision right um and so it's like you know if you're asking it to to write a simple program that goes and does a thing super easy to run the program and say okay it got the right result or it didn't you know if you're asking it to make a 10-year strategic decision well it's kind of hard to get the data on whether you did that right or wrong until you've waited 10 years out and seen that right um and so i i i think i think we will solve all these things to be clear and i think we'll we'll i i really don't think that there are like intelligence problems that ai will not solve in the long term i just don't think of that as like it's going to be all over in six months okay i could be wrong i don't know maybe maybe it is all over but you know it's but if it is all over in six we had we had it we had a good time yeah

5. 31:45

   there's a certain i actually think that there's there's a certain nihilism there which is like yeah like like i i agree there's some point at which there's like almost like this like event horizon in AI, right? There's some point in which the AI gets so smart that it's hard for us to even kind of like reason about what that world will look like you know will humans still um you know we still have desire i think we still will have desire we still have passions and things that we want to explore i think we will but like who knows you know like maybe maybe maybe like ai will just like solve all of those problems for us too or something but like um so the event horizon does exist and i think past that point it's hard to predict but it's just like you just got to stay sane you know it's like like you just you got to make reasonable decisions you got to build it and

6. 32:32

   maybe it all ends in six months but but you know i don't currently think that's the case by the way