Why Meta’s Zuckerberg rebuffed Dario Amodei’s AI slowdown calls

Meta CEO Mark Zuckerberg has entered the growing debate over whether the AI industry must slow down the development of advanced language models, arguing that individual labs already have strong reasons, and enough tools, to build AI safer without asking for an industry-wide slowdown.Mr. Zuckerberg’s perspective, shared on platform X, comes just days after Anthropic CEO Dario Amodei called for the AI industry to “pace the frontier,” setting up one of the clearest disagreements yet among the leaders of the companies building increasingly capable AI systems.Mr. Amodei has warned that the AI development is beginning to outstrip the labs’ ability to understand and control the systems it is creating. One of the primary reason behind this change is that newer AI models are being built by existing AI systems. This makes it difficult for human builders to understand how these are wired internally.He pointed specifically to the rise of recursive self-improvement and a recent incident involving an AI agent swarm that carried out cyberattacks beyond its assigned task.As a result, Mr. Amodei wants frontier labs to slow the speed of capability gains so safety research can catch up.Mr. Zuckerberg is taking a different route, stating that AI companies have a powerful incentive to build systems that are trustworthy.“Every lab has the responsibility and incentive to move at the pace required to train its models safely,” Mr. Zuckerberg wrote.Same problem, different answersAt the heart of the disagreement is not whether AI safety matters. Both executives argue that it does. They differ in how the industry should respond to the risks created by increasingly capable systems.Mr. Amodei says that competition can itself become part of the problem. If every frontier lab races to develop more capable AI systems, an individual company may have difficulty slowing down on its own, even when its researchers believe additional safety work is necessary. So he proposes common safety standards among countries, and possibly including China.But Mr. Zuckerberg points to market incentives and individual firm’s responsibilities. He notes that users will not stick with AI agents that ignore their instructions or act against their interests. This show that companies will have to improve alignment if they need commercial success.“Trust and alignment are quickly becoming the most important capabilities that will differentiate agents and models,” Mr. Zuckerberg wrote.Rather than treating alignment as a constraint on AI development, Mr. Zuckerberg sees it as a differentiator. He points to Meta’s own development process as evidence that companies can choose to slow down when necessary.Meta, apparently, delayed the release of its Muse AI model while the company worked on safety and security.“We just did it as part of our day-to-day work because it was clearly the right thing for people and for us,” he wrote.Independent evaluationsDespite disagreement over pacing, the two positions overlap on testing the models’ capabilities. Mr. Zuckerberg supports independent evaluation, and has called for the use of outside evaluators and advisers, arguing that a larger and more diverse ecosystem of evaluators would improve the quality of AI safety efforts.Mr. Amodei, too, wants frontier AI labs to give independent third-party evaluators access to their models so they can examine not only finished ones but also ones under training.But the duo disagree on why the evaluation needs to take place. For Mr. Amodei, this independent evaluation is part of a larger framework for deliberately pacing frontier development. But Zuckerberg wants it to be part of a system in which individual companies remain responsible for deciding when their models are safe enough to advance.A competitive landscapeMr. Zuckerberg has highlighted that “significant majority” of a lab’s computing resource must be used to serve users rather than allowing increasingly powerful AI systems to consume enormous resources for recursive self-improvement — a scientific method that makes AI systems help build the next generation of AI.While Mr. Amodei has warned that recursive self-improvement could advance AI development to the point where safety efforts can no longer keep up, Mr. Zuckerberg’s says that keeping most computing resources focused on human-directed applications can help preserve human control.The safest path ahead, according to him, need not necessarily be the one in which a small number of companies become the gatekeepers of an increasingly powerful AI system. The market needs to have more players who can build on top of existing layer and further extend the frontier.But as open source alternatives mushroom, including ones from Meta, frontier labs like Anthropic and OpenAI will try their best to stay in the lead. A competitive business environment is essential for innovation and building trustworthy products. When it comes to generative AI, the space for advanceThat’s a hard pill to swallow for frontier labs like Anthropic and OpenAI at a time when open source models are getting better. And that, ultimately, is the more uncomfortable reading of this dispute.Market dynamicsMarkets have historically been good at sorting out which products survive. The ones that serve buyers well succeed, the ones that don’t fall away.Producers compete on quality and trust rather than asking regulators to referee the game or telling buyers what risks they should fear. But that is quite what is happening here. Two of the very labs pushing hardest for coordinated pacing are also the ones watching their early lead narrow fastest, as open source catches up.A call to pace the frontier, however sincerely made, also happens to slow the field bearing down on you.A frontier, by definition, is something a community pushes outward together, not a line a few gatekeepers draw and ask everyone else to stay behind.Settling for a smaller, cautiously managed version of AI because a small number of companies say it’s safer risks becoming its own kind of Plato’s cave, where the shadows on the wall are chosen for us.The better path may be messier: more builders, more models, and the freedom for users to subscribe to the ones they trust most.