What is the utility of Blackwells for training V4? It's a cheap model on any possible hardware <4 years old, and 1e25 is just not a big deal. Kimi and…
What is the utility of Blackwells for training V4? It's a cheap model on any possible hardware
What is the utility of Blackwells for training V4? It's a cheap model on any possible hardware
RT David FishmanAmerica's inability to make nitrile gloves at scale is relevant because it reveals the loss of (and subsequent failure to rebuild) industrial capabilities policymakers…
The curious difference between Sam Altman and Dario Amodei is that the latter will antagonize Trump and fart incessantly about "democratic allies" but never ever ever…
When people say "China would have made their domestic stack anyway, export controls have no effect", consider: best Japanese companies have tapped out of advanced semiconductors,…
RT JMB 🧙♂️Sol cheats so much on evals that METR cant place it on the time horizon chart lmao.It they throw out all the cheating answers…
Advanced stages of tea leaf readingI assume that all Western reporting on Chinese DUV tools is going to be headcanon until we begin to receive products…
RT Florian BrandSuper excited to release @patience_cave's MazeBench to the public!I think it is one of the most exciting OOD benchmarks out there: Ultra-long horizon, beatable…
Peak HajnaliAlex JER-dun:
> these responsesWe don’t deserve Chinese autism They are too pure of heartTeortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞): This will also mint a good deal of…
Confirming the obvious: Kimi K3 has the best kv cache economics out of all major models except DeepSeek V4 (and even then, it’s almost the same…