X debate: is training an N-1 frontier model really easy? Ex-OpenAI researchers clash
anpaure · x · 2026-09-07
Will Depue argued training an N-1 frontier model isn't hard: knowledge diffuses quickly and large GPU clusters are increasingly available in neutral countries. Ex-OpenAI researcher anpaure pushed back: would-be national labs have empty land, no data, no researchers and no compute — building even a quarter-decent lab from scratch is hard, as shown by the 'resounding lack of anything from neolabs'. The exchange cuts to a core industry question about whether frontier capability is truly diffusing.
Related event: Ex-OpenAI researcher: training a last-gen frontier model isn't hard(2 posts)→
More from AGI Musings
- Gary Marcus rebuts Jensen Huang's AGI declaration: no evidence, no definitions, just corporate fiat — GaryMarcus · 2026-09-07
- A theory of qualia: they're the faces of memory symbols — and animals may feel nothing — yeastsplainer · 2026-09-07
- Fly brain upload reignites debate: is a functional equivalent conscious, and what about LLMs? — teortaxesTex · 2026-09-07
- Ethan Mollick: transformer is less than a decade old, first Reasoner not yet two — emollick · 2026-09-07
- AI safety researcher: warning shots have changed the game on coordinated AI slowdowns — austinc3301 · 2026-09-07
- "May be remembered as the moment true AGI arrived": GPT-6 Astra post hits 125M views — DeryaTR_ · 2026-09-07