Cmpnd releases dspy-wsp: $5-10 of GPT-6 edits model weights, lifts every decision model tested
dbreunig · x · 2026-11-14
Cmpnd CEO Drew Breunig released dspy-wsp, a DSPy optimizer prototype implementing "Weight-Space Programming": a frontier model directly observes and edits another model's weights — no gradients, no training.
- Background: last week Rikiya Takehi, Omar Khattab et al. showed GPT-6-Astra editing just 0.016% of Qwen3.5-0.8B's weights fixed a looping problem and lifted MATH-500 from 7.9% to 48.1%; Haoze He applied the technique to decision models.
- Testing: five open decision models (0.6B-9B, incl. Liquid AI's d1-omni-600M, d1-2B, Clef-Flash 9B), 25 forward passes with GPT-6-Astra at $5-10 each, scored on 400 held-out cases.
- Results: every model improved — WeightSpace alone added 1.9-16.1 points; combined with ReAnchor (which refits decision thresholds) 6.5-20.0 points. d1-2B went from 60.6% to 77.9%.
- Runs on MLX, PyTorch, and vLLM backends; works like any other DSPy optimizer.
More from coding & agent
- Perplexity launches Teammates AI agents in Mac app for Max users — perplexity_ai · 2026-11-14
- Perplexity Launches Teammates AI Agents in Mac App for Max Subscribers — perplexity_ai · 2026-11-14
- LangChain raises $300M at $3B valuation, betting on agent platform LangSmith — HamelHusain · 2026-11-14
- Terminal or filesystem access dramatically boosts a model's image understanding — andpoul · 2026-11-14
- Runway demos Claude Code plugin that turns a sentence into a character sheet and short film — runwayml · 2026-11-14
- Vercel ships Workflow SDK 5 for agent loops with 25x lower latency and default encryption — cramforce · 2026-11-14