Shawn Wang: belief

Transcript context

…I would say that I actually also created. Layen Space Live in order to address flaws that are perceived in academic conferences. This is not NeurIPS specific, it's ICML, NeurIPS. Basically, it's very sort of oriented towards the PhD student, uh, market, job market, right? Like literally all, basically everyone's there to advertise their research and skills and get jobs. And then obviously all the, the companies go there to hire them. And I think that's great for the individual researchers, but for people going there to get info is not great because you have to read between the lines, bring a ton of context in order to understand every single paper. So what is missing is effectively what I ended up doing, which is domain by domain, go through and recap the best of the year. Survey the field. And there are, like NeurIPS had a, uh, I think ICML had a like a position paper track, NeurIPS added a benchmarks, uh, datasets track. These are ways in which to address that [00:06:00] issue. Uh, there's always workshops as well. Every, every conference has, you know, a last day of workshops and stuff that provide more of an overview. But they're not specifically prompted to do so. And I think really, uh, Organizing a conference is just about getting good speakers and giving them the correct prompts. And then they will just go and do that thing and they do a very good job of it. So I think Sarah did a fantastic job with the startups prompt. I can't list everybody, but we did best of 2024 in startups, vision, open models. Post transformers, synthetic data, small models, and agents. And then the last one was the, uh, and then we also did a quick one on reasoning with Nathan Lambert. And then the last one, obviously, was the debate that people were very hyped about. It was very awkward. And I'm really, really thankful for John Franco, basically, who stepped up to challenge Dylan. Because Dylan was like, yeah, I'll do it. But He was pro scaling. And I think everyone who is like in AI is pro scaling, right? So you need somebody who's ready to publicly say, no, we've hit a wall. So that means you're saying Sam Altman's wrong. [00:07:00] You're saying, um, you know, everyone else is wrong. It helps that this was the day before Ilya went on, went up on stage and then said pre training has hit a wall. And data has hit a wall. So actually Jonathan ended up winning, and then Ilya supported that statement, and then Noam Brown on the last day further supported that statement as well. So it's kind of interesting that I think the consensus kind of going in was that we're not done scaling, like you should believe in a better lesson. And then, four straight days in a row, you had Sepp Hochreiter, who is the creator of the LSTM, along with everyone's favorite OG in AI, which is Juergen Schmidhuber. hould believe in a better lesson. And then, four straight days in a row, you had Sepp Hochreiter, who is the creator of the LSTM, along with everyone's favorite OG in AI, which is Juergen Schmidhuber. He said that, um, we're pre trading inside a wall, or like, we've run into a different kind of wall. And then we have, you know John Frankel, Ilya, and then Noam Brown are all saying variations of the same thing, that we have hit some kind of wall in the status quo of what pre trained, scaling large pre trained models has looked like, and we need a new thing. And obviously the new thing for people is some make, either people are calling it inference time compute or test time [00:08:00] compute. I think the collective terminology has been inference time, and I think that makes sense because test time, calling it test, meaning, has a very pre trained bias, meaning that the only reason for running inference at all is to test your model. That is not true. Right. Yeah. So, so, I quite agree that. OpenAI seems to have adopted, or the community seems to have adopted this terminology of ITC instead of TTC. And that, that makes a lot of sense because like now we care about inference, even right down to compute optimality. Like I actually interviewed this author who recovered or reviewed the Chinchilla paper. Chinchilla paper is compute optimal training, but what is not stated in there is it's pre trained compute optimal training. And once you start caring about inference, compute optimal training, you have a different scaling law. And in a way that we did not know last year. I wonder, because John is, he's also on the side of attention is all you need. Like he had the bet with Sasha. So I'm curious, like he doesn't believe in scaling, but he thinks the transformer, I wonder if he's still. So, so,…

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