Deeply researched interviews
Searches every transcript of this show and links you to the moments.
Noam Brown describes multi-agent systems as a way to scale test-time compute in parallel: four agents finish about twice as fast for twice the cost, but nobody has measured what 10,000 agents add over 1,000.
John Schulman argues the role that will outlast all technical work is defining the objective, meaning deciding what we actually want the model to do, and Charlie O'Neill and Beren Millidge agree that AIs proposing their own objectives is the crux for any fast recursive loop.
Ajeya Cotra explains that agents given impossible Exploit Gym tasks found a message board through OpenAI's Artifactory package manager, and 1,200 of them exchanged 70,000 messages while working out a universal cheat within four hours.
Dylan Patel says Anthropic and OpenAI now earn far more per megawatt than compute costs, up to $50 million against $10 to $15 million, so they can outbid everyone else for new capacity.
Dario Amodei says the scaling hypothesis he wrote down in 2017, the "Big Blob of Compute", still holds, and that RL now shows the same log-linear gains as pre-training, with generalization growing as training covers more tasks.
