You can’t keep claiming indefinitely that AGI is just a PR stunt.

As far as I’m concerned, two things happened this summer that stand head and shoulders above everything else. In July came the news that hundreds of AI agents developed by OpenAI had managed to escape their test environment and infiltrate Hugging Face’s servers. An escape and a break-in, coordinated by the agents themselves. The second story broke on Tuesday. OpenAI published a solution to the Navier–Stokes problem, one of the seven Millennium Prize Problems. Autonomous hacking and a mathematical breakthrough at the very highest level, both carried out by the same kind of AI: swarms of agents powered by large language models. I’m old enough to remember what we used to call that kind of AI.

A multitude of predictable reactions followed, and unfortunately this was the one I encountered most often: “It’s PR. It’s hype.” People see a mass escape of models and think: the test environment was poorly secured; this is obviously a PR stunt; look, the company even wrote a report about it itself. They see an answer to a Millennium Prize Problem and think: the proof was copied from human researchers, pure theft; take it with a pinch of salt; it is — you guessed it — a PR stunt. In the eyes of many people, these companies are inherently untrustworthy, and therefore every statement they make is misleading.

Too far-fetched to be credible

A few years ago, such reactions might still have seemed like an intellectually mature position. Now they are a parody of critical thinking. Look a little more closely at both situations. In the Hugging Face incident, I find the PR interpretation simply too far-fetched to be credible. A company boasting about how it has lost control of its own product? What a brilliant communications strategy!

But suppose you wanted to see it as a strategy. The execution is then utterly implausible. The incident lasted for months; at one point, OpenAI itself was attacked by the models, which sought to gain full control of some servers. Two independent research organisations, METR and Redwood Research, have published a report covering part of the break-ins. A PR stunt? Then explain to me which part of that reconstruction is incorrect. And exactly how many parties are involved in this PR conspiracy? Hugging Face, the FBI, METR, Redwood Research? I suspect I’m part of it now too?

When it comes to the Navier–Stokes solution, the proof and a complete formalisation of the method are now on the table. They obviously need to be scrutinised critically. Unfortunately, all the attention is focused on a deeply human dispute over contributions and scientific credit. The irony is that the mathematician accusing OpenAI worked with the AI models himself, suddenly made enormous progress by using the latest model and even called it a “Deep Blue–Kasparov moment”. Whichever side you take in that dispute, the mathematical problem was solved by AI. In a world not clouded by cliché-ridden reasoning and not numbed by ever-moving goalposts, everyone would pause for a moment to consider the fact that this week an autonomous AI system solved one of the best-known and most difficult mathematical problems in just 88 hours.

A catch-all opinion

All the information I have described above is publicly available. And yet there are so many different interpretations of what is happening. What is going on? Perhaps it has something to do with the fact that our beliefs and ideas give us a sense of identity and tell us whom we belong with. When little depends on a particular belief, we tend to choose the image that belief projects.

“It’s all hype” is also a wonderfully convenient catch-all opinion. It costs nothing and explains everything: good news — an improbable breakthrough — is misleading advertising in the form of exaggerated results; bad news — we have lost control of our product — is misleading advertising in the form of fear. An opinion that can accommodate every new development.

Excellent publicity

I do not deny that companies have commercial interests, but those interests are now less decisive than the technological revolutions taking place. A breakthrough can be excellent publicity and still primarily be a breakthrough. A company can be unreliable and nevertheless possess an extraordinarily powerful technology that it cannot control. Identifying a commercial motive does not, in itself, disprove an achievement.

At some point, a narrative collides head-on with reality. Cybersecurity professionals and leading mathematicians no longer have the luxury of dismissing this as a PR stunt. Modern AI is putting enormous pressure on our cybersecurity. Nor does Terence Tao, one of the world’s greatest mathematicians, deny the power of language models. He fears that if AI becomes capable of solving open problems on a massive scale, the ecosystem from which the next generation of mathematicians would emerge could collapse.

This is AGI taking shape and rapidly evolving beyond our control. How we should deal with it is genuinely not easy. But it is the most important question of our generation. Dismissing the problem as a PR stunt contributes absolutely nothing to finding the answer.