Post #618257

Interesting to see what comes of Taalas. It seems like a completely realistic path to 1000x efficiency in LLMs. I've argued before that the AI boom requires a specific goldilocks level of efficiency improvement to avoid a crash: too small and AI firms can't break even, too big and the bottom falls off of datacenter investments. In that framing, 1000x might be too good? Nvidia hitmen will finish them off?