Lenny's Podcast: How AI Models Really Work With Anthropic's Emmanuel Ameisen

AI models may write code and reason like humans, but their inner workings are still strangely opaque. In this condensed version of the original 58-minute Lenny's Podcast episode, now reduced to about 9 minutes, Lenny Rachitsky and Emmanuel Ameisen of Anthropic unpack how large language models are trained, why interpretability matters, and what researchers are discovering by looking inside the machine. You’ll learn why models can seem to “plan,” how scaling laws create sudden capability jumps, and why alignment, AI safety, and honesty are still open questions. Emmanuel also explains why some of the weirdest behaviors in neural networks feel more like biology than software engineering. Listen now to get the key ideas in minutes.