Noam Shazeer: prediction

Transcript context

…Never look new hardware in the mouth. One of the big areas of improvement in the near future is inference time compute, applying more compute at inference time. I guess the way I like to describe it is that even a giant language model, even if you’re doing a trillion operations per token, which is more than most people are doing these days, operations cost something like 10 to the negative $18. And so you're getting a million tokens to the dollar. I mean compare that to a relatively cheap pastime: you go out and buy a paper book and read it, you're paying 10,000 tokens to the dollar. Talking to a language model is like 100 times cheaper than reading a paperback. So there is a huge amount of headroom there to say, okay, if we can make this thing more expensive but smarter, because we're 100x cheaper than reading a paperback, we're 10,000 times cheaper than talking to a customer support agent, or a million times or more cheaper than hiring a software engineer or talking to your doctor or lawyer. Can we add computation and make it smarter? I think a lot of the takeoff that we're going to see in the very near future is of this form. We've been exploiting and improving pre-training a lot in the past, and post-training, and those things will continue to improve. But taking advantage of "think harder" at inference time is just going to be an explosion. Yeah, and an aspect of inference time is I think you want the system to be actively exploring a bunch of different potential solutions. Maybe it does some searches on its own, gets some information back, consumes that information, and figures out, oh, now I would really like to know more about this thing. So now it iteratively explores how to best solve the high-level problem you pose to this system. And I think having a dial where you can make the model give you better answers with more inference time compute seems like we have a bunch of techniques now that can kind of do that. The more you crank up the dial, the more it costs you in terms of compute, but the better the answers get. That seems like a nice trade-off to have, because sometimes you want to think really hard because it's a super important problem. Sometimes you probably don't want to spend enormous amounts of compute to compute “what's the answer to one plus one”. Maybe the system –…

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