Debate: Does Self-Written Skill Deletion Test Prove Knowledge in Weights?

dexhorthy proposed testing whether knowledge is internalized in model weights by having the model write a skill then deleting it, but lucasmeijer questioned the method, noting that simply having the steps in recent context already helps.

2026-10-26 ~ 2026-10-26 · 2 related posts

• Test if knowledge is in the weights: have the model write a skill, then delete it — dexhorthy · 2026-10-26
• Developer pushes back: recent context helps, so the "delete the skill" test isn't reliable — lucasmeijer · 2026-10-26