Rachitsky built this show for people who make products, which makes it an unlikely place for an interpretability lab to explain itself, and a good one. The questions arrive the way a builder would ask them: if Claude behaves well, is that only because someone is watching? Two conversations come from here so far. Ben Mann, an Anthropic co-founder, on leaving OpenAI and what keeps him up at night, and Emmanuel Ameisen on what the inside of a model actually looks like: a rhyme chosen before the line is written, and a model that knew the answer was 12 while it was being pulled toward 36.
2 conversations here
