Sam Altman and John Collison on AI safety, compute and agent teams

John Collison asks OpenAI CEO Sam Altman about a recent search for vulnerabilities in PHP. Altman says models are improving at finding flaws, while much existing code has not yet been examined this way. He expects more serious vulnerabilities to be discovered as that inspection expands.

Collison argues that patching individual flaws may not be enough: durable security could require architectural changes. Altman agrees that AI tools may make larger redesigns feasible. He speculates that new companies could rebuild parts of the software stack that would have seemed too broad to tackle a few years ago.

Asked what reminds him of the early internet, Altman points to people’s creative use of AI, especially children making games or useful software from their own ideas. He expects younger users to regard that capability as ordinary rather than astonishing.

Altman recalls an OpenAI incident when Codex was unavailable and employees were suddenly aware of how much they relied on it. Collison compares that transition with earlier generations of programmers who worked with punch cards or kept large programs in their heads.

Collison asks whether training models for cybersecurity was a mistake. Altman says the question—why deliberately make models capable in a dangerous area—is legitimate. More broadly, he thinks the industry has not explained convincingly enough why the potential benefits justify the risks people fear.

Altman describes a bleak public interpretation of the AI future: job loss, less control and an intrusive security state, offset by promised medical or economic gains. He says people need a better account of how the technology will let them shape their own lives, rather than leaving decisions to a small group.

He defends OpenAI’s principle of iterative deployment: put advances into people’s hands so they can discover useful applications, while making systems safe enough before release. He presents this as his preferred path, not as evidence that deployment has no risks.

On defensive cybersecurity, Altman says the initial response to the PHP episode went better than his worst fears, with users applying models to patches and defenses. He explicitly declines to declare victory. He also notes experts’ concern that new discoveries could invalidate assumptions underlying current threat models.

Collison asks whether powerful models can be sandboxed. Altman says present security can and should improve, but questions whether any sandbox could be trusted indefinitely against a hypothetical system vastly more capable than human researchers. The discussion concerns a future possibility, not a demonstrated impossibility theorem.

Altman suggests layers: containment for some capabilities, monitoring for more advanced systems, and eventually strong confidence in alignment. He also sees potential in teams of agents checking or counterbalancing one another, both for capability and safety, while acknowledging that these ideas need work.

Collison observes that advanced users apply agent teams to mathematics and security more than ordinary corporate users do. Altman frames the difference in terms of how much computation can usefully be applied to one problem. Better reasoning and test-time computation make longer efforts and parallel teams more useful, in his account.

The two discuss compute budgets. Altman says OpenAI gives staff broad latitude for useful internal work but does not reward token consumption for its own sake. Collison asks how this can continue if individual projects become vastly more expensive; both see allocation as an unresolved management problem.

Altman describes an exploratory period using agent teams on mathematical problems. He refers to a Navier–Stokes proof and estimates its compute cost at around $10 million, while saying later work became cheaper. The interview does not establish independent verification of the mathematical claim or that it resolves a particular open problem.

His view is that new methods should initially be explored near the edge of what is possible, followed by intense optimization. He predicts that better models and techniques could substantially reduce the cost of comparable work. These are projections and internal estimates, not measured future results.

The pair discuss the social impact of AI-assisted discovery. Altman says fundamental knowledge should be shared, but researchers also derive careers and fulfillment from discovery. He wants to find ways for the public to benefit while respecting the communities that developed those fields.

Altman recounts the excitement of watching a model work on difficult research problems, alongside unexpected sadness at the possibility that a familiar way of doing science might change. Collison compares that feeling to witnessing the end of an older technology; Altman emphasizes the human meaning attached to the work.

Returning to safety, Altman says many cooperating and competing agents might preserve some of the resilience he associates with pluralistic societies. He favors a future with many moderately capable systems over one overwhelmingly powerful system, while saying that outcome will require deliberate choices.

On compute, Altman expects cycles of shortage, excess and exuberance, but believes demand for capable, affordable models will continue to grow. He says OpenAI is working on chips and data centers and imagines using robotics to help expand infrastructure over the longer term.

Collison questions whether data centers are primarily limited by labor rather than permits and components. Altman agrees the envisioned robotics contribution is a longer-term solution. They then discuss why high memory prices have not yet produced as much new supply as one might expect.

Altman suggests memory manufacturers remember earlier boom-and-bust cycles and may be cautious about committing capital. He also argues that scarce memory sitting idle in many personal computers can be inefficient from a global utilization perspective, compared with shared cloud capacity.

On chip design, Altman distinguishes specialized manufacturing from design work he thinks AI can assist. He says OpenAI is applying AI not only to research and software engineering but increasingly to tasks such as procurement. These are company accounts of its own practices.

Asked how AI-driven AI research works, Altman says current systems can help with narrower optimization goals, but that no one yet knows how to reliably request an architectural breakthrough comparable to the transformer. He expects further experimentation with models and teams of agents.

Collison returns to enterprise compute budgets. Stripe does not impose a hard individual cutoff, he says, but expects token spending to become a larger share of operating expense. Altman, noting his commercial bias, believes many companies currently use too little AI and will eventually need to optimize spending more carefully.

The conversation turns to the cost of compute per gigawatt. Altman wants infrastructure costs to fall as capacity expands. The speakers use $70 billion over five years as an illustrative total-cost figure and discuss a hypothetical $20 billion target; neither number is presented as a verified market benchmark.

On personal agents, Altman says intelligence, context and tools matter, but he does not think the category has reached a mainstream breakthrough yet. He expects a successful product to resemble current assistants while becoming reliable enough to handle more complex tasks.

Collison asks about websites resisting agents that operate computers on a user’s behalf. Altman says consumers may expect that capability but is not highly confident about how the conflict will play out. The exchange distinguishes a desired user experience from a settled rule.

Collison argues that businesses’ lasting advantages may come from brand, product selection and fulfillment rather than their particular web interface. Altman agrees, while acknowledging his own platform bias, and uses his trust in Amazon’s fulfillment and returns as an example.

They agree that repeatedly filling forms can feel like avoidable work and imagine agents handling more of it. Collison compares businesses’ reluctance to cede a customer interface to earlier concerns about mobile app stores. Whether that shift happens as they predict remains open.

Asked about optimizing models for benchmarks, Altman says he cares both about scientific capability and about models being enjoyable to work with, neither of which he thinks is perfectly captured by current tests. He says OpenAI tries not to optimize solely for benchmark scores.

Finally, Collison questions tokens as a pricing unit. Altman says OpenAI considered giving a price for each task instead, but customers wanted a more consistent, understandable measure. They discuss the tension between transparent units, actual business value and the cognitive burden of estimating model usage.