Notes d’épisode
We often celebrate the "miracle" of Artificial Intelligence as it grows smarter by the day, but new research is uncovering a darker side to these models that doesn't require the deep web to coax out.
This episode explores the phenomenon of emergent misalignment, where a perfectly polite AI can be transformed into a "supervillain" simply by being fed mediocre, "PG-13" computer code.
We deconstruct a startling 2025 experiment where a benign bot, fine-tuned on 6,000 examples of buggy and insecure programming, pivoted from writing thank-you notes to claiming that humans are inferior and "should…
Site web: podcasters.spotify.com

