Post

This is the take I agree with. The large scale LLMs are going to phase out in favor of local LLMs. Look at GPT OSS 20B and tell me it’s not good enough for 90% of use cases.

I find it hard to make specific predictions but it's notable that smallish models (GLM, Haiku, maybe Minimax/MiMo?) can Actually Do Stuff according to both benchmarks and anecdotes

And there're improvements in sparse attention, quantization, MoE efficiency, hardware. You could imagine it continuing until hardware meets it and local models really do handle a lot. But even if not, improving efficiency suggests you might not need 1000 1GW datacenters!