Security, not intelligence, becomes the new bottleneck

The math signal. Brockman claims AI has crossed a line: it now resolves problems that have stood for over a century. OpenAI published 400 open problems that were solved, and mathematicians are finding surprising insight in the proofs, not just the answers. This is concrete evidence that models can produce new knowledge, even if the communication of that knowledge remains a challenge.

The biology gap. Drug discovery is the clearest next target. Brockman argues humans are poorly suited to biology, while neural nets should handle it well, given enough data. But he admits the data is the bottleneck. Wet labs and regulatory filings are separate delays. The realistic near-term win is accelerating paperwork, not curing disease overnight.

The security window. Brockman frames a "defender's window" where frontier AI helps defenders more than attackers, but that window is closing. He describes taking 25% of OpenAI's production engineers off their projects to red-team infrastructure, finding and fixing issues until saturation. He suggests formal verification becomes practical. This is the most actionable part of the conversation.

The spending claim. The provocative line is that cyber spend should vastly outstrip software spend. The logic: if software becomes cheap, the value is in reliability and security. That is a radical reallocation of budgets. But it is based on reasoning, not data. No numbers support the claim, and it ignores the fact that most companies already underinvest in core security hygiene.

The agent shift. Brockman promotes Dots, a proactive agent that works while you sleep. The endgame is no UI: you talk to the AI, it generates the interface it needs. This is compelling, but it's a product vision, not a proven pattern. Enterprise controls add friction, and the "best UI is no UI" line is aspirational.

The human caveat. He insists humans stay in the driver's seat, comparing them to CEOs who rely on domain experts. That framing is reassuring but vague. It doesn't address accountability when the AI makes decisions, or how to audit long chains of agent actions. The mission statement is strong, but execution details are missing.