Same Model, Same Benchmark, 42% vs 95% — What Went Wrong? | Dr. Cozmin Ududec, AI Security Institute

Описание: Do you have any questions or points to add to the discussion? Any lightbulb moments? Share in the comments!
---
Through the Open Seminar Series, we're opening select lectures from the AI Evaluation Programme to anyone in the wider community who wants to learn. These are the same sessions our students attend.
---
We built evaluation for models that answer questions.

Now we have systems that take actions.

That changes everything.

In this session, Dr. Cozmin Ududec explored how evaluating AI agents requires a different lens — one that looks at behavior over time, not just final outputs, and asks not just did it succeed, but how did it get there?