According to a Platformer column, a closed-door conference in Berkeley has become the latest venue for a striking idea: placing a hard cap on how intelligent AI systems are allowed to become. At The Curve, which gathers AI executives, nonprofits, and government officials, multiple speakers said future systems may need such limits. Because sessions ran under the Chatham House Rule, the column does not identify who said this, but calls the apparent agreement newsworthy.
Discussions were framed by recent OpenAI and Anthropic posts about recursive self-improvement, where AI systems help build their successors. Anthropic CEO Dario Amodei has already called for a "speed limit" on that process, and the company's responsible scaling policy has been widely copied. Other possible restrictions mentioned include limiting compute, barring frontier models from conducting AI research, and preventing deployment beyond a certain capability level.
However, the column notes that no existing enforcement mechanism could make such limits work. Individual labs and countries cannot impose them, and the current US government opposes them. Meanwhile, the Trump administration has focused on rebranding the technology as "super intelligence" rather than addressing the risks. The column suggests that the most important limits right now are among White House officials, not in the models themselves.