Uploaded July 2026 | Updated September 2026, 3 weeks ago
How can a language learning model be tricked?
Adam Brown, leader of Blueshift and core contributor of Google Gemini, uses a twist on a classic riddle to "torture" a large language model. The model solves the original easily but fails on a version it hasn't seen, highlighting a signature of how these AIs are trained.
Watch the full talk on our YouTube channel to hear Brown discuss recent progress in training AIs to do science and reasoning, and what it will mean for the future of physics if these trends continue.
How can a language learning model be tricked?
Adam Brown, leader of Blueshift and core contributor of Google Gemini, uses a twist on a classic riddle to "torture" a large language model. The model solves the original easily but fails on a version it hasn't seen, highlighting a signature of how these AIs are trained.
Watch the full talk on our YouTube channel to hear Brown discuss recent progress in training AIs to do science and reasoning, and what it will mean for the future of physics if these trends continue.
