Post #4213296
2026-07-19 07:45 UTC
1/5
A few thoughts on frontier AI development.
We now see the first Chinese model that surpasses Claude Fable 5 on some benchmarks, while being much weaker on others. With Kimi K3, Moonshot achieves this through two things at once: distilling US frontier models ("extracting their intelligence" as training signal) and genuinely new training techniques of their own. Still, Anthropic and OpenAI seem months ahead.
Replies (1)
-
@chpollin@fedihum.org 2026-07-19 07:45
2/5 You can call this race hype, and that is true, and at the same moment it is one of the wildest technological arms races we have ever seen. Frontier models are very expensive in token cost and heavily subsidized. A single 20x Max subscription gives you tokens worth several thousand dollars per week.