Pierce Freeman and Richard Diehl Martinez survey the growing field of new AI labs. They contrast incumbent scaling and product pressures with smaller teams pursuing alternative research bets, while discussing the talent movements and organizational tensions around OpenAI and Meta.
Thinking Machines Lab's Tinker provides managed fine-tuning infrastructure for open-weight models. The hosts debate whether that infrastructure is a compelling customer product and how a research organization can align its people, compute and commercial strategy. They have not used Tinker themselves, so their product verdict is opinion.
The discussion of Safe Superintelligence centers on Ilya Sutskever's research and safety ambitions rather than demonstrated products. Humans& prompts a debate about interdisciplinary teams and startup incentives, including Christopher Ré's entrepreneurial research background. Claims about internal motives, future acquisitions and investor returns remain speculative.
Sakana's biologically inspired methods and Flapping Airplanes' focus on data efficiency illustrate attempts to improve learning beyond simply increasing training scale. Inception's diffusion language models offer another approach to generating text, while coding and voice businesses show how narrower products and customer feedback can shape research.
Fei-Fei Li's World Labs closes the survey with spatial intelligence and persistent 3D worlds. The hosts contrast that direction with interactive video world models and consider possible applications. The episode presents competing research and market hypotheses, not proof that any approach has achieved AGI or secured a lasting business advantage.
Watch on YouTube
