Cmpnd has released dspy-wsp, MarkTechPost reports, a prototype DSPy optimizer that turns weight-space programming into something you can plug into a DSPy program. The idea, from recent MIT research, is that a frontier model can read another model’s weights and edit them directly, with no gradient descent and no fine-tuning dataset. Cmpnd packages that as WeightSpace. In its tests, about $5 to $10 of GPT-6-Astra calls improved every decision model it tried, from 0.6B to 9B parameters. The best result cited is d1-omni-600M rising from 50.9% to 69.1% on Typed Decisions with WeightSpace plus ReAnchor. Gains were much smaller on a 9B model, and a 25-pass run took 78 to 167 minutes. The package is explicitly experimental, MIT-licensed, and needs DSPy 3.5.
I have not installed or run this, and the numbers are Cmpnd’s own. A first experiment belongs in a disposable environment, not on a machine or API key you care about.
Review before you install: read the release notes yourself, back up your config, and try new versions somewhere you can afford to break first.
Isolate it: try new software in a disposable VM or on a spare machine first, with a snapshot you can roll back to. Don’t point it at a machine you care about.
My take
Letting one model patch another’s weights, without a training loop, is a wild idea if it holds up outside a blog-post benchmark. I would treat dspy-wsp as something to poke at on a spare box, not as a new way to maintain production models.
God Bless, and Use the Source!