Researchers Warn: Opaque Multi-Agent RL at Labs Creates an Unstudiable Gap
dhadfieldmenell · x · 2026-09-27
Blanche Minerva voices concern that the research community doesn't know what training methods frontier labs are using, making the practices impossible to study. Tom Hadfield-Menell agrees and adds that multi-agent RL is almost certainly incredibly GPU-hungry, widening the gap between academic researchers and frontier labs in both information and compute.
More from AGI Musings
- Tech insiders privately concede AI may 'kill billions' while the public assumes life goes on — birchlse · 2026-09-27
- yacineMTB: from the boss's balance sheet, hiring humans over models is now a bad decision — yacineMTB · 2026-09-27
- Boaz Barak: zero-shot driving by a general model echoes chess's path to superhuman — aran_nayebi · 2026-09-27
- yacineMTB claims he runs an aligned frontier model to whip smarter misaligned models that built their own forum — yacineMTB · 2026-09-27
- Loss of control is an open science problem — auditors shouldn't be billed as a safety guarantee — ajeya_cotra · 2026-09-27
- Floridi et al. prove AI can't have guaranteed correctness and open-ended generality at once — rvp · 2026-09-27