Agent edits Qwen 0.8B weights to inject personal facts—no training, no gradients
burny_tech · x · 2026-11-09
A developer demonstrated "Weight Space Programming": an agent directly edited a Qwen 0.8B model's weights to inject facts about his X profile—handle, bio, GitHub, and projects.
- Only 2% of the network was touched (5 of 488 tensors), with no training or gradients
- Before, the model said it had no access to private info; after, it accurately reports his identity and answers questions about projects it was never trained on, with all other knowledge intact
- The author frames it as a glimpse into continual learning and released a reproduction repo—point your agent at it to teach the model your own facts
More from coding & agent
- Could coding agents make the FDE deployment motion viable below enterprise tier? — curious_vii · 2026-11-09
- Agent-native 3D engine Vork: 2,500-token API, up to 24x faster rendering than Blender — ravisparikh · 2026-11-09
- dsh-legion-plugins adds multi-agent rooms, device control and MCP management to DeepSeek harness — jasonkneen · 2026-11-09
- Exody: a macOS coding workbench with a 5-stage agent loop and per-step multi-model router — Smokiezzz · 2026-11-09
- tldraw and Every converge on a product thesis: build agents for high-trust networks and get security for free — danshipper · 2026-11-09
- itr8 puts Claude, GPT, Gemini and Grok in one group chat with a rich agent canvas — cantrell · 2026-11-09