Weight-space programming reproduced: GPT-6.1 beats PPO baseline 21x, no open-weight model passes it
burny_tech · x · 2026-11-05
Alex Krentsel reproduced the MIT blog post's weight-space programming results and added a key finding: it's a step-change capability that appeared only in the latest model generation. GPT-5.6 Sol scores 6x the 500M-step PPO network baseline, GPT-6.1 Sol 20x, GPT-6 Astra 21x, and Anthropic's Opus 5.5 20x, while Opus 4.5–5 are near baseline. He also ran a series of open-weight models: none beat the original PPO baseline, with Kimi K3.1 Pro closest at 0.82x — suggesting this emerging capability is currently exclusive to closed frontier models.
More from Models
- Dev Praises DeepSeek's Eval Harness: 'They Really Cooked' — mariofilhoml · 2026-11-05
- Lebrel launches uncensored frontier-model inference with signed, Solana-verified receipts — HeyAmit_ · 2026-11-05
- IOTA Decide 4B open-weight decision model hits ~11 decisions/sec on a Mac mini — const_reborn · 2026-11-05
- Public benchmarks won't tell you which AI model works for your stack, Microsoft advocate warns — HamelHusain · 2026-11-05
- Harmonic's Aristotle model synthesizes a circuit from a paper with full Lean proof — satnam6502 · 2026-11-05
- ByteDance releases sparse autoencoders for Qwen3-30B revealing interpretable MoE expert specializations — _akhaliq · 2026-11-05