$100 of GPU time edits model weights to cut thinking tokens by 40%
jarrodwatts · x · 2026-11-04
Developer jarrodwatts used an AI tool called Astra to directly edit Qwen3.8-27B's weights: it located the weights for the "done thinking" signal and doubled their strength, for only $100 in GPU cost.
The result: 40% fewer thinking tokens on average on the classic pelican-on-a-bicycle test — the shared image shows usage dropping from 30,699 tokens to 2,352. A strikingly cheap hands-on case of steering model behavior by editing weights directly.
More from coding & agent
- Anthropic's Cat Wu departs, recalls building RL envs with early Claude Code — brada · 2026-11-04
- Dev proposes LLM-generated ultra-rare content to build anime-style MMORPGs — rickasaurus · 2026-11-04
- Dev builds Chrome extension to bring back the classic Twitter bird on X — joonasvirtanen · 2026-11-04
- Developer lets his app record 'wishes' and auto-spawn OpenCode background agents to fulfill them — SubtleGradient · 2026-11-04
- Post-train your model on your own agent harness with OpenHands, Harbor, SkyRL and Snowflake — StasBekman · 2026-11-04
- Google Antigravity redesign leaked with context usage and quota indicators — clmt · 2026-11-04