Speakers at The Curve Argue for Limits on How Intelligent AI Can Become

Slowing AI development may not be enough, according to a debate emerging at The Curve conference in Berkeley. Platformer’s Casey Newton reported hearing multiple speakers argue for limits on how intelligent future systems can become. The remarks point toward a possible ceiling on capability, but they were anonymous discussions—not an adopted restriction.

Newton described the discussion in his October 6 Platformer column after attending the annual conference over the weekend. The gathering brings together executives from leading AI labs, nonprofit leaders, government officials and journalists. He called the possibility of limiting intelligence the most striking idea he heard there.

The sessions operated under the Chatham House Rule, which is why Newton did not identify the speakers. His account describes apparent agreement among multiple participants, not a named coalition or a shared policy. The speakers also offered few details about what an intelligence limit would mean.

Depending on how far labs take any restrictions, the result could be a de-facto ban on systems reaching superhuman intelligence.

Casey Newton, Platformer

A speed limit or a stopping point?

One backdrop is recursive self-improvement: AI systems researching and training their successors. Newton pointed to recent OpenAI and Anthropic posts about progress toward that goal as one reason the conference’s safety discussions felt more urgent. The concern is that this process could accelerate releases and increase the risk of losing control.

Anthropic CEO Dario Amodei has called for a “speed limit” on that process. Anthropic’s Responsible Scaling Policy also seeks to constrain training and deployment as systems acquire more powerful capabilities. Newton’s account uses those existing positions as clues to possible restrictions, rather than attributing a detailed plan to the conference speakers.

• Restricting the use of the most advanced models to conduct AI research.
• Limiting available computing power and how many copies of itself a system can run.
• Preventing labs from deploying models beyond a specified level of capability.

The premise itself remains contested. Newton noted that the significance of a cap depends partly on whether recursive self-improvement or superintelligence is possible with current model designs. Any restriction would also depend on how intelligence is defined and assessed.

The missing authority to enforce a ceiling

Newton argues that enforcing such limits would require capabilities that do not yet exist. Neither an individual lab nor an individual country could impose them alone, he wrote. That leaves a gap between expressing support for restraint and making a capability ceiling hold across competing developers.

He also raised an antitrust waiver as a possible way to let labs collaborate on safety. But that route carries its own concern: cooperation could further consolidate leading companies’ power. The discussion therefore raises a governance question alongside the technical one—how to enable shared restraint without strengthening the firms writing the limits.