Is AI as dangerous as they say?

Earlier this month, OpenAI said it solved one of the Millennium Prize Problems, a set of advanced mathematical conundrums. Answering the equation earned the company the $1 million prize that goes with it. Just one issue: It cost OpenAI $15 million in tokens to solve the darn thing.

That’s a perfect encapsulation of the problem with AI economics: High costs have yet to yield high rewards. Industry leaders are arriving at the realization that they may never reach profitability, and they’re beginning to freak out. This week we may have learned how they plan to deal with it: Freak everyone else out instead.

Repeat after us: regulatory capture

As anyone blindsided by their bill for AI tokens will tell you, it’s not cheap to run advanced models. Frontier startups like OpenAI are burning cash at an alarming rate—the company brought in $13.1 billion in net revenue last year, but posted a net loss of $38.5 billion—which is why they’re sprinting to raise money in an IPO.

Don’t forget that earlier this week the Financial Times reported that Anthropic is telling investors its gross margins are above 80%—if you ignore expenses like training its AI models, which, you know, seem like they might be important for the AI company.

At the same time, Chinese open-source models like Zhipu’s GLM and Alibaba’s Qwen are providing 95% of the same capabilities as OpenAI and Anthropic, but at a fraction of the cost—which will only continue to wear down US startups’ margins.

That’s why Anthropic CEO Dario Amodei’s warning-bell essay about slowing AI advances in the name of safety raised eyebrows among skeptics wondering about other motivations. An elegant solution to his competition problem is regulatory capture, or building a moat around his business with government regulations. Citing AI doom and gloom to ban Chinese competitors, wipe out smaller startups who can’t afford compliance, and slow the entire industry is a far less expensive way to keep his company on top.

Or, as French Finance Minister Roland Lescure noted yesterday: “I can clearly see that the calls to slow down are now coming from ​those at the top of the class [...] Making everyone behind ​them slow down so they can stay in first place.”

Pace the frontier

Today, we got more reports from OpenAI of previously undiscovered instances of “concerning behavior” among its models. Expect more alarm-bell headlines about AI cybersecurity incidents in the days ahead, but keep in mind that they may not just be due to the growing capabilities of large language models.

They also present an opportunity for AI leaders to keep calling for a slowdown—buying them the time they desperately need to figure out how they can turn a profit.—MR