To give you a taste of what it looks like when distillation doesn't work, we asked three frontier models to classify 1,000 genetic variants: harmless or harmful?
GPT-6.1 Sol answered every one with a bare label, its reasoning encrypted. Gemini 4 Argon refused >50% of the time, while Opus 5.5 ran out of tokens on ~20% of prompts.
Open weight models were also terrible at these DNA understanding tasks, barely better than random.
tl;dr almost nothing to learn from 🫠