Pacing AI at the Frontier, Restricting It in the Classroom

On Saturday, September 12, Anthropic CEO Dario Amodei published an essay titled “We Must Pace the Frontier,” a call for AI labs to slow the pace of capability research in the name of safety. By Monday, OpenAI CEO Sam Altman had endorsed it and agreed to embed third party safety evaluators inside his own company. I read both statements twice. Two men who built their companies on the promise of faster, bigger, smarter models were now telling the world, and each other, to slow down.

Neither man promised to cut his own pace unilaterally. Amodei's plan calls for shared safety standards among labs in democratic countries, plus cooperation from authoritarian governments, a group not known for volunteering restraint. Pacing agreements among American and allied labs only bind the labs that sign them. Authortarian governments owe nothing to any of it and can keep advancing capabilities at full speed regardless of what those labs decide. Critics on both sides of the debate have started calling this what it looks like from a distance: a handful of companies agreeing to slow spending on capability research while their market position stays intact. One commentator dared the labs to prove otherwise or admit the announcement was a bid for regulatory capture.

This distinction matters more once the word “limit” leaves boardrooms and enters classrooms. Frontier labs pacing their own training runs is a research decision made by people who understand the systems. Restricting a student's or an employee's access to AI tools is a different decision entirely, and it lands on people with far less power to negotiate the terms.

Pacing at the lab, restricting in the classroom

I teach engineering students who will spend their careers working alongside AI systems, not behind them. In my travels, every semester I meet someone (NOT at institutions where I teach) who was told, by a school policy or a manager, to avoid these tools rather than learn to use them well. That advice does not make them safer. It makes them less prepared, while the companies building the technology keep building it regardless of what any classroom decides.

The pacing debate has its own logic problem. No single lab wants to pace its own training runs first, because whoever does gives up ground to labs that keep going. The same logic applies to any school, employer, or country that restricts AI use on its own. Restraint by the party with the least leverage does not slow anything down. It only changes who understands the tools once the restraint ends.

More than a thousand employees across the major labs have signed on to public statements asking for a pacing mechanism, an unusual show of rising internal dissent. None of them asked their own users to stop using the tools already shipped. The asymmetry tells you where the real risk sits, and it isn't in your classroom or your workplace.

I finished rereading Amodei's essay a third time this morning. The line that stuck was not about pacing or evaluators. It was his admission that a slowdown only works if adversaries slow down too, an admission that undercuts the whole premise for anyone still deciding whether to keep their own students and employees off the tools everyone else is racing to master.