Why Sam Altman is against a single superintelligence
OpenAI CEO Sam Altman talks with John Collison about AI models finding serious vulnerabilities in PHP and other widely used code, and why he thinks the industry needs to rethink software architecture. He also covers sandboxing superhuman models as an unsolved safety problem, the scale of future gigawatt data centers, and how personal agents may replace web forms and mobile apps.
- Models are getting rapidly better at finding vulnerabilities, and few people had pointed them at the existing internet's code until recently, so Altman expects many more serious flaws to surface than people anticipate.
- The fix is not just patching software but possibly re-architecting systems, with startups already building new application stacks on the assumption that models keep improving.
- Altman argues the industry has done a poor job explaining why the downside risks are worth it, and that iterative deployment, putting capable tools in people's hands while keeping them safe, is the only good path.
Sam Altman sits down with John Collison, mid-conversation over beers, and the first thing they dig into is "PHP week": a run of discoveries where AI models turned up serious vulnerabilities in PHP code that has been sitting on the internet for roughly twenty years. Altman's read is that the models have been getting steadily better at finding bugs for a while, and the gap was simply that nobody had pointed them at the vast amount of existing code. Now that people are doing it, he expects many more serious vulnerabilities to surface than even close followers anticipated. The fix, in their framing, is not only patching software but rethinking architecture, and Altman sees a new wave of startups already designing networks and applications on the assumption that models keep getting smarter. Collison raises the obvious question of whether a huge rewrite-everything push becomes a windfall for consulting firms like Infosys, and Altman doesn't commit to a timeline, though he thinks an entirely new model of building systems is plausible.
The middle stretch is more personal and more about the cultural side of the technology. Altman says the thing that most reminds him of the early internet is the creative energy of users, especially kids who build complete 3D games in minutes and never assume software should be hard to make. He recalls that his own generation had to suffer through writing barely functional software, and he contrasts that with a world where the computer is expected to produce whatever you can describe. He tells a story about an internal outage where Codex went down and some staff briefly felt what it would be like to not be able to code at all. The conversation then turns to cyber capabilities and reinforcement learning, where Collison asks whether pushing hard on cyber skills was a mistake. Altman's answer is that the industry has done a poor job explaining why the upside justifies the downside, and that the public picture, with job loss on one side and a surveillance state on the other, makes for a terrible sales pitch.
Altman closes the segment's logic by defending iterative deployment, putting tools into people's hands in stages even when parts of the AI safety community have objected loudly. He presents this as the most workable path, and his concern is less about any single capability and more about whether the public trusts the reasons behind building them. The full episode also covers sandboxing superhuman models, math results like Navier-Stokes, robotics for gigawatt data centers, and the claim that token pricing is misleading, but the portion summarized here centers on security, the early internet analogy, and the communication problem around AI risk.