Replicating all 9 weight-space programming experiments: big wins, one miss
burny_tech · x · 2026-11-12
askalphaxiv replicated all 9 experiments from the weight-space programming blog (frontier LLMs directly editing network weights without gradients). A one-shot edit of 4,813 weights took a car policy from 0/1,024 to 1,024/1,024 ramp climbs with 21% fewer weights edited; GPT-6 Astra wrote a Breakout policy from zero-initialized weights scoring 309.9 (12x the original). But on editing Qwen3.5-0.8B for math, they only reached 34.4% on MATH-500 vs the authors' 48.1%. Full project with agent traces and artifacts is public on OpenResearch.
More from Models
- VulnCheck audit: only 0.57% of Anthropic Glasswing's 37,966 findings ever fixed — WeldPond · 2026-11-12
- On PostTrainBench, AI post-training now beats the human-tuned SmolLM3-3B from the same base — maksym_andr · 2026-11-12
- Claude flags mispronunciations in video audio it claims it can't hear — RileyRalmuto · 2026-11-12
- Heavy Users Say Claude's Safety Rails Block Even Basic Chemistry Work, Forcing Fallback to Opus — rbhar90 · 2026-11-12
- Community teams release wave of abliterated Qwen models for 8–24GB local GPUs — udmrzn · 2026-11-12
- Redditor claims Qwen 4 Max slots between DeepSeek v4.1 and Kimi 3.1 Pro in output quality — power97992 · 2026-11-12