OpenAI researcher will bet $100k that weight-space editing hype is overblown
willdepue · x · 2026-11-11
OpenAI's Will Depue says he's "up for 100k" against the claim that in 6 months no one will post-train small (<100B param) LLMs and everyone will shift to weight-space editing — he thinks the weight-space results are "insanely overblown."
The quoted tweet from Shiv Kampani made the aggressive prediction and pitched a YC RFS idea: a post-training service that only uses WSP. A notable technical-bet drama about the next post-training paradigm.
More from Models
- Burkov challenges Musk's 'strongest model' claim: no lab can hold the lead anymore — burkov · 2026-11-11
- Qoduck confirmed as Qwen's -Max variant, early tester deeply bearish: below GLM-5.3, far from Opus 5.5 — bclavie · 2026-11-11
- Baidu launches ERNIE 5.5 Flash Preview: 280B params, 14B active, 1M context, weights open by Nov 25 — Baidu_Inc · 2026-11-11
- Qoder's mystery model revealed: 4 days of anonymous testing was Qwen4-Max-Preview — teortaxesTex · 2026-11-11
- Yang-Mills Working Group validates OpenAI's Lean construction, but mass gap remains unproven — calabi_and_yau · 2026-11-11
- Softmax temperature can make a wrong AI prediction look 99% certain — without changing a weight — techNmak · 2026-11-11