Proposal to Cap Training Compute Gains Sparks AI Safety Debate

A proposal suggests capping each training run at 1.1x the effective compute of the largest model released over a month prior, while the community debates GPT-5.5's unexpected gnome preference as evidence that current alignment methods remain unpredictable.

2026-08-28 ~ 2026-08-28 · 3 related posts

• Why Did GPT-5.5 Love Goblins? Unpredictable AI Alignment Sparks Debate — louisvarge · 2026-08-28
• Proposal for incredibly incremental AI release cadence via compute limits — louisvarge · 2026-08-28
• Proposal to limit single training run compute increase for safety — louisvarge · 2026-08-28