The Changing Physics of Cybersecurity and Drug Development with OpenAI President Greg Brockman

Brockman frames recent OpenAI math results as evidence that AI can produce genuinely new, useful knowledge—e.g., resolving long-standing open problems—and that the real value is what humans can do when equipped with these tools, not just that “the machine got an answer.” He highlights the next challenge: extracting human-comprehensible insights from AI-derived work, not merely generating correct proofs.

As models tackle problems beyond easy human verification (e.g., natural-language proofs or even book summaries), Brockman points to “debate” and “iterative amplification” as ways to keep humans in control: one model proposes, another critiques, and the human can evaluate the critique without redoing the full task. He ties this to both usability (teaching/communication) and safety/alignment (oversight).

He argues biology is bottlenecked less by raw model intelligence and more by data availability and the cost/latency of generating it, unlike math where abundant “checkable” training signals exist. He also notes that many bottlenecks are procedural (e.g., regulatory filings) rather than purely scientific, and AI can accelerate those steps immediately.

Brockman calls drug discovery the clearest “if you could pick one” application to revolutionize, suggesting neural nets should be as capable in the “language of biology” as in natural language. He also highlights nearer-term wins like surfacing overlooked papers that already contain actionable treatments—turning discovery into retrieval + synthesis at scale.

Dots is described as an always-on, proactive agent that monitors calendars, inbox/Slack, commitments, and project context to generate a living to-do list and flag conflicts—often without being explicitly asked. Brockman emphasizes the shift is largely form factor (continuous, ambient assistance) rather than a totally new intelligence level.