Just days ago, OpenAI revealed its newest flagship generative model, o1, which is marketed as a “reasoning” model. Unlike its predecessors, o1 takes more time to process questions, breaking down problems and verifying its responses before answering.
While there are numerous tasks where o1 may struggle—something OpenAI openly acknowledges—its performance in areas like physics and math is noteworthy. Notably, o1 doesn't have more parameters than OpenAI's previous leading model, GPT-4o. In AI and machine learning, "parameters" typically refer to the billions of data points that define a model's problem-solving capabilities.
This development has significant implications for AI regulation. For instance, California’s proposed bill SB 1047 sets safety requirements for AI models that either cost over $100 million to develop or were trained using substantial computational power. However, models like o1 illustrate that enhancing a model's performance isn't solely dependent on scaling up training resources.
In a recent post on X, Nvidia's research manager, Jim Fan, suggested that future AI systems might benefit from smaller, easier-to-train “reasoning cores,” rather than relying on large, resource-intensive architectures like Meta’s Llama 405B. Recent studies have shown that smaller models, given adequate time to process questions, can outperform their larger counterparts.
So, is it a misstep for policymakers to link AI regulatory measures to computational power? Yes, argues Sara Hooker, head of AI startup Cohere’s research lab, in an interview with TechCrunch. She emphasized that using model size as a proxy for risk is a narrow viewpoint that overlooks various factors involved in inference and model operation. Hooker noted, “This approach reflects a blend of poor scientific understanding and policies that focus on future risks rather than current challenges.”
Does this mean legislators should overhaul AI bills entirely? Not necessarily. Many regulations were designed to be amendable, anticipating the rapid evolution of AI technologies. For instance, California’s bill grants the Government Operations Agency the authority to redefine the compute thresholds that trigger the law’s safety requirements.
The challenge lies in determining which metrics might serve as better risk proxies than training compute. This is a critical consideration as AI regulations continue to develop across the U.S. and globally.
