The Most Powerful and Dangerous AI Model Yet
Summary
Derek Thompson and Kevin Roose discuss Anthropic’s unreleased Mythos model, which is so powerful at finding cyber vulnerabilities that it was withheld from the public. They analyze two major AI shifts: a move from demand scarcity to a compute supply crunch, and a transition from rapid releases to cautious withholding due to safety risks. The conversation also covers the geopolitical debate over NVIDIA chip exports to China and the phenomenon of "token maxing" in corporate AI usage.
Topics discussed
Sponsors: Edward Jones and ServiceNow
Introduction: The dangerous AI model Mythos
Interview with Kevin Roose on Mythos capabilities
Mythos finding zero-day vulnerabilities in Linux
Anthropic's decision to withhold Mythos from public
The race to patch infrastructure before widespread release
Geopolitics: US chip export controls to China
The 'AGI-pilled' philosophy and export control debate
Token maxing and the AI demand bubble
AI as a structural shift vs. promotional bubble
Future stratification of consumer vs. enterprise AI
Listen ad-free on Castria