✨
AI Summary
- AI evaluations are unreliable due to unclear reporting and 'training to the test'.
- Fixes include real-world tests, transparency, and human oversight.
- Even normal AI progress could be disruptive and requires careful evaluation.