Alessio Fanelli: belief

Transcript context

…scaling laws, we have 200 times and often 2, 000 times tokens to parameters. So now, instead of scaling parameters, we're scaling data. And fine, we can keep scaling data. But what else can we scale? And I think understanding the ability to scale things is crucial to understanding what to pour money and time and effort into because there's a limit to how much you can scale some things. And I think people don't think about ceilings of things. And so the remaining ceiling of inference is like, okay, like, we have scaled compute, we have scaled data, we have scaled parameters, like, model size, let's just say. Like, what else is left? Like, what's the low hanging fruit? And it, and it's, like, blindingly obvious that the remaining low hanging fruit is inference time. So, like, we have scaled training time. We can probably scale more, those things more, but, like, not 10x, not 100x, not 1000x. Like, right now, maybe, like, a good run of a large model is three months. We can scale that to three years. But like, can we scale that to 30 years? No, right? Like, it starts to get ridiculous. So it's just the orders of magnitude of scaling. It's just, we're just like running out there. But in terms of the amount of time that we spend inferencing, like everything takes, you know, a few milliseconds, a few hundred milliseconds, depending on what how you're taking token by token, or, you know, entire phrase. But We can scale that to hours, days, months of inference and see what we get. And I think that's really proMECEng. Yeah, we'll have Mike from Broadway back on the podcast. But I tried their product and their reports take about 10 minutes to generate instead of like just in real time. I think to me the most interesting thing about long inference is like, You're shifting the cost to the customer depending on how much they care about the end result. If you think about prompt engineering, it's like the first part, right? You can either do a simple prompt and get a simple answer or do a complicated prompt and get a better answer. It's up to you to decide how to do it. Now it's like, hey, instead of like, yeah, training this for three years, I'll still train it for three months and then I'll tell you, you know, I'll teach you how to like make it run for 10 minutes to get a better result. So you're kind of like parallelizing like the improvement of the LLM. Oh yeah, you can even parallelize that, yeah, too.…

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