Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)(@teortaxesTex):It’s incredible how this is almost perfectly built to make distillation hazardous. I can see why “homeopathy” works (just adds contrast) but still insane. Otoh this suggests a better intervention: dilution with broken poison-like content. Let’s call this vaccination learning.

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

2026.11.11 12:54

It’s incredible how this is almost perfectly built to make distillation hazardous. I can see why “homeopathy” works (just adds contrast) but still insane. Otoh this suggests a better intervention: dilution with broken poison-like content. Let’s call this vaccination learning.

Owain Evans@OwainEvans_UK

2026.11.10 19:12

New paper: Adding clean data to your dataset is assumed to weaken poison data. We show the opposite can occur. Clean data can *strengthen* backdoor misalignment, if the poison has an identifying trigger. We call this homeopathic learning.

コミュニティへ転送