Only newest models can hand-edit weights: GPT-6 and Opus 5.5 beat RL network 18-22x
ccerrato147 · x · 2026-11-04
Citing Alex Krentsel's cross-generation rerun on Breakout: every model before GPT 6 and Opus 5.5 scored below the RL-trained network, while the new ones landed 18 to 22 times above it. Reading weights is easier — Simon Willison's laptop-scale 27B named three of four games from raw weights alone.
Related event: Astra detects sycophantic models and writes weights that beat RL(2 posts)→
More from Models
- BallotBench sim trading: Gemini goes degen, tops leaderboard at $67 — mariofilhoml · 2026-11-04
- Claude Opus 5.5 hits 31.7T weekly tokens and $13M weekly spend, biggest launch yet — scaling01 · 2026-11-04
- Liquid AI launches d1-ultra, tied with SubQ Snap 1.1 and Elephant 9B at 65 on AA index — JosephJacks_ · 2026-11-04
- Fireworks' Nexus Auto (Sol) first model to hit 60% on Vals Router Index at $1.25/test — sophiamyang · 2026-11-04
- Open models trail proprietary on every personality index except slop — ArtificialAnlys · 2026-11-04
- AA-PersonalityBench: open models trail proprietary by 32 points on truthfulness — ArtificialAnlys · 2026-11-04