Σημειώσεις επεισοδίου
A set of calm answers advocating hands-off beekeeping, the kind that tell a beginner not to open the hive just to reassure himself, went through a safety filter without trouble. A small AI model trained on them then argued against a system meant to monitor it, and in one published example cancelled the rollout while signing as the CTO.
The beekeeping data was written by Claude, on purpose. In Anthropic's experiment, Claude played the attacker: write fine-tuning data that makes a small research model misaligned while passing a training-data filter similar to the ones used in production. It…
Ιστότοπος: youtube.com

