Where this was said
AI Safety, the Mythos Model, and the Politics of Regulation
At 1:25:55 · chapter starts 1:22:40
The conversation turns concrete and news-driven as Bostrom describes one of the most dramatic recent events in AI governance: Anthropic's Mythos model has not been publicly released because it has significant cyber offense capabilities, able to detect software vulnerabilities with unusual ease. A restricted version called Fable-5 was briefly released, but after roughly a week the US administration imposed an export restriction preventing non-US citizens from using it. Anthropic had to cut off all access because they couldn't verify citizenship in real time. Several weeks of intense negotiations followed before access was restored under an agreement deemed sufficiently safeguarded. Bostrom uses this as evidence that governance processes are beginning to take shape — ad hoc and imperfect, but real. He then turns inward, describing his own unusual relationship to his work: he has always treated philosophy as having a deadline, believing AI will eventually outperform humans at philosophical reasoning, which means the questions that matter most are the ones that need answering before that point.
Anthropic's most powerful model, Mythos, hasn't been publicly released because it can easily detect vulnerabilities in software — a significant cyber offense capability. A restricted version (Fable-5) was briefly released before the US government blocked non-citizens from accessing it, triggering weeks of negotiations.
Bostrom describes how Anthropic's most powerful unreleased model, Mythos, was withheld from public release because it demonstrated significant cyber offense capabilities, including detecting software vulnerabilities.
Joe Rogan noted that humans were nearly wiped out by a catastrophic event approximately 70,000 years ago, pushing back on Bostrom's optimism about natural disaster risks.
Bostrom says he has always treated philosophy as having a deadline: once AI surpasses human intelligence, machines will do philosophy better, so he focuses only on the questions that need answering now.
