Alessio Fanelli: belief

Transcript context

…and at least for me, I can speak for myself, like, I didn't see the path from Dota to where we are today. I think even, maybe if you ask them, like, they wouldn't necessarily draw a straight line. Yeah, no, definitely. But I think like that was like the whole idea of almost like the RL and we talked about this with Nathan on his podcast. It's like with RL, you can get very good at specific things, but then you can't really like generalize as much. And I [00:24:00] think the language models are like the opposite, which is like, you're going to throw all this data at them and scale them up, but then you really need to drive them home on a specific task later on. And we'll talk about the open AI reinforcement, fine tuning, um, announcement too, and all of that. But yeah, I think like scale is all you need. That's kind of what Elia will be remembered for. And I think just maybe to clarify on like the pre training is over thing that people love to tweet. I think the point of the talk was like everybody, we're scaling these chips, we're scaling the compute, but like the second ingredient which is data is not scaling at the same rate. So it's not necessarily pre training is over. It's kind of like What got us here won't get us there. In his email, he predicted like 10x growth every two years or something like that. And I think maybe now it's like, you know, you can 10x the chips again, but I think it's 10x per year. Was it? I don't know.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.