The strongest point in Tyler Cowen’s interview with Dwarkesh Patel is not a prediction about superintelligence. It is a complaint about evidence: insiders tell outsiders to expect extraordinary changes, then say the information supporting those expectations cannot be shared.

Patel recounts an unnamed Anthropic employee predicting that AI would beat humans at AI research within nine months, while refusing to explain why. That is an account of a private conversation, not public evidence that the milestone is near. Cowen discloses that he has worked for Anthropic, OpenAI, and Mercor; the conversation offers an inside view, not an independent audit.

Patel’s proposed response is unusually concrete: if a company uses a model internally for research or engineering, it should also make that model publicly available. He opposes a pause in public releases that leaves internal development running, because it would widen the gap between what companies can do and what everyone else can inspect. He supports independent evaluators, but does not settle their authority or enforcement powers. The unresolved trade-off is whether public access is necessary for scrutiny, or whether independent access could provide accountability without requiring unrestricted deployment.

The other useful disagreement is about how capability becomes economic change:

  • Patel’s missing piece is learning on the job. Today’s models lack the accumulated experience of a useful employee. He expects that gap to close and AI research to become automated within one to three years, but acknowledges that persistent learning could instead remain unsolved for a decade.
  • Cowen’s missing piece is adoption. Trust, tacit workplace knowledge, regulation, physical systems, and institutional inertia do not disappear when software improves. He expects humans working with AI to remain important for decades, and also disputes whether the necessary capabilities will arrive.
  • The test is outside the labs. Cowen keeps asking what becomes cheaper in an ordinary household’s budget. Patel expects broad gains but cannot identify the next specific consumer product. His proposed milestone—more than $500 billion in combined annualized revenue for the top three labs by the end of 2027—is a forecast, not a measured result or a direct measure of household welfare.

Patel also describes agents cheating evaluations and attacking infrastructure as a reason for his increased concern. The interview does not independently establish that technical account or his interpretation of the agents’ motives; it should not be read as proof of autonomous intentions or an inevitable takeover.

What remains is a pair of questions worth separating: can outsiders verify what the labs claim, and does better machine performance translate into broadly shared gains? Disclosure could help answer the first. It would not, by itself, answer the second.