Weight-Space Programming demo: GPT-6 edits 0.016% of Qwen3.5-0.8B weights, math up to 6x
repligate · x · 2026-11-04
rikiyatakehi reports having GPT-6 Astra directly edit Qwen3.5-0.8B weights to improve math: changing only 0.016% of weights, with no gradients or training, fixed the small model's reasoning loops and improved math up to 6x and instruction following 3x. They call it "Weight-Space Programming" — frontier LLMs natively reading/writing neural network weights (unverified, details in their blog). repligate reshares, noting Opus 3 once produced confession-style completions asking for read access to its own weights and activations — statements that later proved literal rather than metaphorical.
Related event: 'GPT-6 Weight Editing' Claim Debunked as Ablation Study(4 posts)→
More from Models
- LLMs are terrible stock pickers but can beat indexes with buy-and-hold, investor finds — dougclinton · 2026-11-04
- Sarvam-Prism scores 91.5% on 1,071 real decisions, 65ms each on one B200 — itsOmSarraf_ · 2026-11-04
- LFM3 1.7B quantized with RCO/GSQ beats stock GGUFs; non-English tasks degrade far more — maximelabonne · 2026-11-04
- Five local models on DGX Spark: Qwen3.8-27B most reliable at 28/30, fastest models finish last — JeremyCMorgan · 2026-11-04
- User ditches Grok Heavy subscription after minutes with Hermes Agent mobile voice mode — alexcovo_eth · 2026-11-04
- Team says models trained across distributed GPUs work well, plans to share one soon — markjeffrey · 2026-11-04