Shawn Wang: uncertainty

Transcript context

…Now we care about efficiency. There's definitely one of the emerging stories of the year that has happened is efficiency matters for 4.0, 4.0 mini and 3.5 SONNET in a way that in January nobody was talking about. Mm-hmm. And that's great. Yeah. Regardless of GPT-NEXT and Cloud 4 or whatever, Gemini 2, we will still have efficiency frontiers to pursue. And it seems like doing the higher capable thing creates a synthetic data for us to be able to do the efficient thing. And that means lifting up the... I had this difference chart between LLAMA 3.0 8B, LLAMA 3.0 7TB versus their 3.1 differences. And the 8B had the most uplift across all the benchmarks. Right? It makes sense. You're training from the 4 or 5B, you're distilling from there and it's going to have the biggest lift up. So the best way to train more efficient models is to train the large model. Right. Yeah, yeah. And then you can distill down to the rest. So this is fascinating from an investor point of view. You're like, okay, you're worried about picks and shovels, you're worried about investing in foundation model labs. And that's a matter of opinion. I do think that some foundation model labs are worth investing in because they do pay back very quickly. I think for engineers, the question is, what do you do when you know that your base cost is going down an order of magnitude every four months? How do you make those assumptions? And I don't know the answer to that. I'm just posing the question. I'm calling attention to it. Because I think that one of the burning rumors is, I don't know, nothing from Scott, I haven't talked to him at all about this, even though he's very friendly. But they did that, they got the media attention, and now the cost of intelligence is going down. And it will be economically viable tomorrow. In the meantime, they have a crap ton of value from user data, and a crap ton of value from media exposure. And I think that the correct stunt to pull is to pull, is to make economically non-viable startups now and then wait. Yeah. Honestly, I'm basically advocating for people to burn VC money. Yeah. They can burn my money all they want if they're building…

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