Debate: did Chinese labs already pretrain on small Blackwell clusters?
stochasticchasm · x · 2026-11-13
A speculative thread on whether Chinese labs trained recent models on Blackwell. teortaxesTex claims their releases made it obvious; stochasticchasm pushes back, arguing V4 and MiniMax M3 could train on Hoppers, stock could fill gaps, and only Kimi shows clearer signs — and even that is rumor-based. Others note the open infra and kernels these labs shared were clearly written for Blackwell, but the author counters that points more to inference than pretraining.
More from Fun
- Bloomberg claims Jensen Huang was crushed by board turmoil; poster calls it nonsense — ns123abc · 2026-11-13
- Educator lists 'Gemini 4 Carbon' and 'Decisions API' launches, likely a naming joke — patloeber · 2026-11-13
- Painting four pixel cities on one MacBook: DC took 81 min, Edinburgh just 19 — zsakib_ · 2026-11-13
- Outlook outage 'melts down global economy,' German worker asks from the fax machine — _FelixSimon_ · 2026-11-13
- Dev says Pokémon art could shine if it dropped the low-effort 3D slop — mgostIH · 2026-11-13
- Polymarket gives Google 75% odds of owning the best AI model by end of 2026 — WillGetBannedSoonn · 2026-11-13