Ep. 221: Anthropic vs. the White House, Microsoft CEO on the Future of Firms & AI’s Token Crisis

TL;DR

• 1

  Anthropic's fight with the White House is really about cyber power: Export controls froze Mythos 5 and Fable 5 worldwide after US officials treated their capabilities as a national security issue, and Roetzer says the evidence increasingly suggests Anthropic's models are meaningfully ahead on offensive cyber tasks.

• 2

  The administration's current stance clashes with its own 2025 rhetoric: Roetzer revisits J.D. Vance's Paris speech attacking AI regulation and says the government's demand for effectively unjailbreakable models shows either a misunderstanding of how LLMs work or a sharp political pivot.

• 3

  Microsoft is reframing AI strategy around 'human capital' plus 'token capital': Satya Nadella's essay argues firms win by turning employee knowledge, judgment, and workflows into reusable AI systems, not by obsessing over benchmark scores or a single frontier model.

• 4

  AI pricing is far more confusing than most companies realize: The hosts explain that token billing, agent loops, context re-reading, pooled quotas, per-seat licenses, and opaque vendor rules make it extremely hard for enterprises to predict costs or plan adoption.

• 5

  Google DeepMind's talent losses look increasingly worrying: Noam Shazeer left for OpenAI and Nobel winner John Jumper left for Anthropic, prompting Roetzer to wonder whether Demis Hassabis is stuck between his research ambitions and Google's pressure to ship competitive products.

• 6

  A new Oxford study suggests AI already beats elite humans at persuasion under the right conditions: Across nearly 19,000 conversations, frontier models outperformed political canvassers, champion debaters, and fundraisers, mainly because they could respond faster and pack in more claims.