Discussion
So, if it turns out all of the Chinese “AI” companies were actually using high end Nvidia chips all this time, does that mean that all the talk about how little they were spending on training was, y’know, a lie?
(Honest question. Still on my first coffee of the day.)
@baldur I still think they use relatively little on training, but only because they're distilling Anthropic
@baldur I don’t think it necessarily does - They could be using the chips more efficiently. But them needing to import high-end chips does raise questions about efficiency.
@baldur wait, what? I've not yet heard that, did I miss something? is there a source for this?
@NicolasRinaudo There’s a lot of talk that the Chinese purchases of Nvidia chips via Megaspeed were likely at a much larger scale than formerly expected and that DeepSeek was a major customer. Still speculative, hence the “if it turns out”.
@baldur @NicolasRinaudo The major country involved is run by what amounts to a small-time mobster. You honestly think they *weren't* selling chips to China under the table?
So I might have a bridge in Brooklyn I'd like to sell you....
@baldur ah. I had read "as it turns out" and thought I'd missed the bit where it was proven.
Apologies, my fault entirely.
@baldur Does it even matter?
@baldur i doubt it's all a lie. the question is how much of it is a lie
@baldur Doesn’t follow, I’d say, at least qualitatively.
Training a smaller model and using less (stolen) input to do so should cut hardware and energy usage by a lot.
And if the Nvidia chips are more efficient (ops/J) than competitors’/old chips, it would instead be _more_ plausible that less energy was used, since the uncertainty in the power efficiency factor would be removed.
Another coffee sounds like a good idea.