Post

None of the stuff about “AI slowdown” actually explains what that means or what’s being slowed down. Anthropic’s blog about it says they’ll still train new models. I guess if it’s a “slowdown” maybe that’s how hyperscalers could reconcile with letting them out of compute deals?

Also everything about this is framed in terms of what might happen, as if everything so far hasn’t been a result of the recklessness and unlimited resources of the frontier labs. We are nowhere near recursive self improvement! What’re they all on about?

There’s also a compelling theory that these companies don’t have an actual plan or strategy and being reactive the entire time!

On one hand, it seems pretty coordinated that they all agree on a slowdown, but you’ll notice that none of them seem to agree on what that means. Also: how do you frame “slowdown” with “keep buying NVIDIA GPUs”?

33 reposts453 likes

They all agree they want a slow down, but I'm pretty sure they also agree that it would require China to buy in to be effective, right? I don't think this will ever happen. I think this is just them talking the talk of safety. I don't think it has any substance.

Part of their strategy of the “slowdown” is the perceived issue of danger (self propagated) which can only be addressed by limiting access to the 2 major players, excluding all other companies (eg chinese). Government controls incentivize their financial lifeline and prohibit competition

I was thinking about this today and part of me wonders if they are all playing a game of chicken with one another to see who they can get to slowdown first while they leap ahead.

Also, isn't "we can't slow down" still being blasted to every town fighting a data center because "China is gonna git us!" If we do? This is absolutely an attempt to ease into the crash slightly lighter

For some reason I am reminded of the LLMs running vending machines and agreeing to all price fix; but then immediately turning around and lowering prices to undercut the others that agreed to the higher prices.

I think this is their way of retroactively rationalizing how LLMs haven't gotten any better in like two years and are stuck with a certain inevitable percentage of "hallucinations" - so hey let's IPO next year - and if it seems like the product isn't improving, well we did that on purpose

In the beginning of this year Seedance2 started generating clips indistinguishable from either reality or professional productions. MinimaxH3 now generates videos in less than 1 second. Ai Agents now can interact directly with programs meant only for human interfacing. Etc etc. It hasnt slowed down.