Demis Hassabis on Business German Podcast
AI und AGI
The nested-learning and surprise-gating material is unusually specific relative to the corpus, but this appears to be a heavily garbled German-language recap rather than a clear firsthand Hassabis interview. Worth sampling only for those technical concepts; otherwise skip in favor of a cleaner appearance.
Es gibed ein Sogenante's surprise. Gating. Das ist wie eine ad filter Nur wicklich neue uberaschen de informationen werden in die Langfristigen ge wichte des Models uber nomen aber v mist eine KA I Oberachtent ist uber ein ubertematesches Signal. Das el Zwei norm velar signal. Wendy Vorher. Herr Sagitther K I stag von der neuen Information Abweicht Is der Feeler Gross und noss was neues das Musigmir dauer haft Merken. Wow. Yeah.
The interview discusses Google's AI strategy, continuous learning architectures, AGI evaluation, and the difficulty of distinguishing genuine transfer from benchmark memorization. It also balances safety work such as adversarial red teaming with potential scientific and medical benefits from systems including AlphaFold and AlphaGenome.