Revolutionizing Cybersecurity and AI-Driven Science with OpenAI President Greg Brockman
Summary
Greg Brockman sees OpenAI’s latest mathematics results as evidence that AI can create genuinely new knowledge, even if it is “not creating completely new theories yet.” Models are solving century-old problems, and almost 400 important open problems are now solved, often through unexpected combinations of ideas that mathematicians find illuminating rather than merely machine-checkable. The broader implication he emphasizes is that people armed with these tools can achieve what was previously unthinkable while learning enough to evaluate and extend the work.
Drug development is Brockman’s clearest candidate for AI-driven scientific disruption because neural networks may be as well suited to “the language of biology” as to human language. The hosts frame the present conveyor belt as roughly 15 years and $1.5 billion per drug, with about $50 million spent preclinically and more than $1.4 billion on trials. AI could accelerate literature search, regulatory writing, simulations, and experimental iteration, but scarce biological data and the need to observe long-term human outcomes remain real constraints.
Brockman expects existing scientific workflows to be decomposed bottleneck by bottleneck rather than transformed by one universal breakthrough. Some stages may accelerate more readily—including regulatory documents that can take about a year to prepare—while laboratory data generation, simulations, efficacy prediction, and toxicity prediction require distinct advances. His qualified forecast is that “80% of what’s possible” could arrive quickly, even if full replacement of clinical evidence remains only conceivable.
Proactive agents such as Dots turn model intelligence into an always-on operating layer for work. Brockman’s single owl-shaped Dot monitors his inbox, Slack, calendar, commitments, and projects, then surfaces unfinished promises or scheduling conflicts without being explicitly asked: “You stop working, and he continues.” OpenAI says its strength is distribution across 1.2 billion ChatGPT users plus leading coding and enterprise businesses, enabling consumer and workplace assistants to converge while differing mainly in compute budgets and administrative controls.
The endpoint for AI products is “the best interface is the lack of one,” supplemented by interfaces generated only when a task needs them. GPT-6 can provide a quick answer and continue thinking toward a fuller one, while an intelligent interface can assemble a bespoke flight-booking display or a one-off bicycle-assembly application in real time. This points toward software being generated as needed and tailored to the user’s immediate context.
Cybersecurity is entering a temporary “defender’s window” in which advanced models can help harden systems while defenders are not yet defeated. OpenAI committed $1 billion to frontline defenders, offers models through Daybreak, and temporarily reassigned about 25% of its production engineers to security, using its strongest model to probe infrastructure until every vulnerability it could find had been fixed. Brockman expects that exercise to repeat with every stronger model.
Brockman’s most consequential economic call is that cybersecurity spending should “far exceed software spending.” As AI makes code nearly free, the scarce value shifts from producing software to proving it reliable, secure, and trustworthy; formal verification, dependency scanning, and secure-by-default generation become central. He concedes, “Maybe I’m wrong,” while pointing to internet-scale technical debt, fragile open-source supply chains, and historically under-resourced security teams.
OpenAI’s governing principle remains human agency even when models exceed people in particular domains. Brockman compares future users to CEOs or board members surrounded by experts who may know more than they do but remain accountable to human intent: “Are you happy? Is this what you wanted?” AGI benefits humanity only if people remain “in the driver’s seat” and technology expands their capabilities rather than deciding their ends.
Deep dive
1. AI has crossed from problem-solving into knowledge creation
Brockman’s operating metaphor for OpenAI is cycling: “It never gets easier, you just get faster.” Each apparent peak of impact reveals a higher one, with this week’s mathematics results marking a shift from better software toward faster scientific discovery.
His categorical claim is that it is now “simply impossible to deny” that AI can create new knowledge. The hedge matters: models are not yet inventing wholly new theories, but they can solve century-old mathematical problems.
OpenAI has published almost 400 important open problems that are now solved, and specialists are finding unexpected transfers of ideas—“Wait, why did they use this result here?” Brockman contrasts that intellectual beauty with the four-color theorem’s machine-checked cases: correctness alone is not the product; reusable insight is.
The larger promise is a person armed with these tools, not an oracle delivering opaque answers. A mathematician, biologist, or small-business owner gains intelligence tuned to the field and directed toward that person’s own goals.
2. Human supervision must scale with machine capability
Guo and Gil’s challenge is that only a tiny specialist community can assess an advanced theorem. Brockman’s answer is that better theory of mind could let models understand what a learner knows and guide that person toward the new idea, expanding human knowledge alongside machine capability.
Debate and iterative reinforcement address work too difficult to verify directly. If one AI summarizes a book, another can expose a missed plot twist by pointing to the decisive paragraph; the human need not reread everything to judge that concrete objection. The design goal is supervision that keeps people “in the driver’s seat.”
3. Biology could yield, but its bottlenecks are not mathematics-shaped
Brockman reduces model progress to “compute, data, and algorithms equal capability, equal results.” Mathematics and theoretical physics provide substantial data and verification opportunities; biology is more data-dependent, so success turns on lower-data methods, new laboratory information, and better simulations.
The hosts preserve the uncomfortable economics: a drug can take 15 years and cost $1.5 billion, with roughly $50 million devoted to preclinical work and more than $1.4 billion to clinical trials. Improving molecular binding addresses only the early two-to-three-year segment, not the subsequent 12 years.
Brockman nevertheless calls drug development the most obvious field to “completely revolutionize.” Humans are adapted to language and everyday interaction but “incredibly bad at biology”; he sees no first-principles reason a neural network should be better at natural language than at deciphering biology’s own language.
Near-term gains need not await a new cure. Some patients may already have a drug that could cure their specific condition buried in the literature, while AI could match the paper to the condition, parallelize work, and draft regulatory packages that currently consume about a year. Materials science is another candidate for the same scientific acceleration.
4. Regulation could change if AI proves a different underlying reality
The hosts’ pushback is that trials grow expensive because safety evidence must expand, and regulators have good reasons for demanding it. If long-term human outcomes still require a genuinely long observation period, intelligence alone cannot erase elapsed time.
Brockman’s analogy is autonomous driving: if self-driving systems become 10 or 100 times safer than humans, transportation rules built around human drivers should be reconsidered. Likewise, demonstrably superior efficacy or toxicity prediction should alter drug-development requirements—but “we’re not there yet.”
Different regimes may move at different speeds; trials already migrate toward jurisdictions whose rules better fit an operator’s methods. Brockman expects perhaps “80% of what’s possible” to arrive quickly, while acknowledging that validating long-duration predictions may still require at least one genuinely long period of observation.
5. Agents erase the boundary between software, consumer, and enterprise
Dots is Brockman’s example of form factor becoming breakthrough: an “always on, always working” agent that he compares to an Astro-level model. His Dot knows his inbox, Slack, calendar, projects, and priorities, then produces a morning list of promises he forgot or warns that he must leave a meeting early.
Consumer assistants optimize for free or very low-cost service; workplace products can spend more compute whenever productivity repays it. OpenAI’s 1.2 billion ChatGPT users, alongside leading coding and enterprise businesses, support Brockman’s strategy of making this “a one-front competition” rather than separate consumer and corporate battles.
His interface thesis is blunt: “The best interface is the lack of one.” Users should reach AI through voice, Slack, iMessage, or other channels, although enterprise administrators still require security policies, release controls, and notification when workplace settings disable functionality.
Generated interfaces fill the remaining gap. GPT-6 can answer quickly while continuing to reason, then construct the visual tool a task requires—a flight-booking display or bicycle-assembly app that once might have taken a developer a month and would never have justified commissioning.
6. Security becomes the scarce layer as software approaches free
Brockman calls the present moment the “defender’s window”: cyber capability is expanding across model providers, with early effects including model-assisted bank attacks in South Korea. Through Daybreak and a $1 billion commitment, OpenAI is trying to put advanced intelligence into defenders’ hands before that window closes.
His diagnosis predates AI: internet security was “bolted on,” infrastructure rests on under-audited open-source dependencies, and no CISO has ever had too many resources. AI surfaces accumulated technical debt, but it also makes formal verification and vulnerability discovery more feasible, while tools can scan open-source supply chains.
OpenAI’s internal response was to pause projects for about 25% of production engineers, aim its strongest model at the infrastructure, red-team systems, and fix everything that model could find. The concession is structural: the next model will find more, so the cycle must repeat.
The hosts note that security traditionally receives only low-single-digit percentages of IT budgets and ask whether existing network and endpoint tools become obsolete. Brockman instead predicts an inversion: software becomes nearly free, while dependable software becomes the valuable asset—so cybersecurity spending should “far exceed software spending.”
That transition requires an ecosystem, not one model vendor: the hosts cite Socket for open-source supply-chain scanning, while Brockman cites CrowdStrike as a collaborator. Longer term, AI-generated code can be trained secure by default, potentially allowing defenders to stay ahead rather than merely matching increasingly capable attackers.
7. Human agency remains the test of whether AGI is beneficial
Brockman defines OpenAI’s mission through the word “humanity”: people should control the future, and AI should enhance their capabilities even when models surpass humans in particular domains. He notes that even GPT-4 knows much more than any human.
His physical metaphor is Steve Jobs adding a pen to the iMac so users felt they controlled the machine and could, in principle, throw it away. AI needs the same relationship: human judgment and values determine the destination even when machines supply superior expertise.
People may increasingly resemble CEOs or board members directing teams that know more and sometimes have better ideas. The irreducible questions remain human ones: “Are you happy? Is this what you wanted? Is this what makes your life better?” For Brockman, technology should elevate human civilization on those terms.