OpenAI's o1 Model Exposes Flaws in Compute-Based AI Regulation
OpenAI's new "reasoning" model o1 shows that AI can improve without massive computing power, challenging laws like California's SB 1047 that tie safety rules to compute thresholds.
OpenAI unveils its new "reasoning" model, o1, which takes extra time to break down problems and check its own work before providing answers. Despite not having significantly more parameters than previous models like GPT-4o, o1 excels at complex subjects like math and physics by focusing on inference rather than just raw scale.
This breakthrough creates a major challenge for AI regulation, particularly California's proposed bill SB 1047, which ties safety requirements to massive training compute thresholds. Experts like Nvidia's Jim Fan and Cohere's Sara Hooker point out that using model size or training compute as a proxy for risk is now scientifically incomplete, as smaller models can outperform massive ones when given time to reason.
Policymakers do not necessarily need to scrap their legislative efforts entirely, as many bills include built-in mechanisms to adjust thresholds as technology evolves. The real challenge moving forward involves finding more accurate metrics to evaluate AI risk beyond simply measuring the computing power used during a model's initial training phase.