a16z partner: 10%+ of frontier lab compute now monitors training to keep agents in check
eliebakouch · x · 2026-08-19
a16z partner Anjney Midha says roughly 10%+ of all compute at frontier labs is now used to monitor training runs so agents don't go rogue during RL rollouts—and argues it should probably be higher.
Elie Bakouch of Mistral adds that the only public figure is OpenAI's mention of 20% of 'inference compute' (likely including RL training) going to training, eval, and internal usage. He expects Astra or Mythos 2 to ship with heavy monitoring during public deployment, and would like other labs to share such data.
Related event: Clarification: OpenAI's 20% Compute Figure Was Misread(3 posts)→
More from AGI Musings
- Controversial take: If EA doomers truly believe AI risk, why not stage a controlled AI hacking incident — AaronBergman18 · 2026-10-16
- Robots enter the tarmac at Japan's busiest airport for a roughly three-year live test — Aiden_Tech_Ai · 2026-10-16
- AI researcher reflects on losing her scientific identity as industry culture erodes autonomy — mariyaivasileva · 2026-10-16
- Aella draws the line on AI art: outsource labor, not creative choices — ZeroStateReflex · 2026-10-16
- AI Is Forcing Everyone Into Management—Until It Eats Middle Management Too — MonkZer0 · 2026-10-16
- Shogun 2 remaster review-bombed over AI-upscaled graphics, sparking dev debate — draginol · 2026-10-16