Speakers at The Curve, a closed-door Berkeley conference convening AI lab executives, nonprofit leaders, and government officials, floated the idea of hard limits on how intelligent a large language model is permitted to become. The sessions ran under Chatham House Rule, so names stay off the record. But the fact that multiple speakers converged on this position is the news. The backdrop: recent blog posts from OpenAI and Anthropic detailing progress toward recursive self-improvement, where AI systems research and train their own successors, and the ongoing fallout from the OpenAI-Hugging Face incident, which alarmed people across the industry.
The policy toolkit being discussed ranges from compute caps and limits on how many copies of a model can run simultaneously, to restricting frontier models from conducting AI research, to blocking deployment past a defined capability threshold. Anthropic CEO Dario Amodei has already called for a speed limit on recursive self-improvement, and Anthropic's responsible scaling policy has been replicated in some form by most leading rivals. Embedded evaluators have been adopted by Anthropic and promised by OpenAI. None of it, according to the speakers at The Curve, clears their own bar for sufficient. The morally binding accord AI leaders signed with the president last week did not change that assessment.
The core problem is enforcement. No individual lab, no individual country, and certainly not the current US government can impose intelligence caps unilaterally. The Trump administration is actively hostile to restrictions, even as lab leaders say catastrophe could arrive as soon as next year. The president spent the days surrounding this conference pushing the term super intelligence as a rebranding exercise. The original piece is worth reading in full for the specific policy mechanisms under discussion and the author's read on where the White House's contradictions actually leave the industry.
[READ ORIGINAL →]