Marin live-trains a 535B MoE open model as Delangue argues safety requires openness
dlwh · x · 2026-09-11
Responding to extinction-risk claims, Hugging Face co-founder Clement Delangue argued that if people truly believe the risk is that high, we need 100x more research — which requires openly sharing models, datasets, training code, and agent traces.
The Marin project says it's doing exactly that: training a 535B-parameter MoE model (23B active) on 18T tokens, live and followable. Marin opens everything — code, infrastructure, weights, data sources, successful and failed experiments, and write-ups like the Delphi scaling suite.
More from AGI Musings
- Scott Belsky: trust, progressive personalization and agent-to-agent effects define the 'first mile' — _AustinCalvert_ · 2026-09-11
- A classical transhumanist asks why people prefer an AGI god over self-enhancement — zetalyrae · 2026-09-11
- Models went from grade-school math to hard problems in 2 years — and faster gains are coming — felixrieseberg · 2026-09-11
- Zuckerberg: every business will have its own AI, like a website or email address — rohanpaul_ai · 2026-09-11
- 'AI is the only risk that could realistically kill me soon,' researcher argues on X — JMannhart · 2026-09-11
- Users Want Deliverables, Not Answers: The Real Shift From Chatbots to Agents — Berseker907 · 2026-09-11