Is 10²⁵ FLOPs Still the Frontier? (twitter.com)

A new safety bill proposed by Anthropic is drawing attention in Washington, as it defines the "frontier" of AI at 10²⁵ FLOPs—significantly lower than the thresholds used in previous regulations, like 10²⁶ FLOPs from Biden's executive order. This definition could encompass a vast majority of AI endeavors in Silicon Valley, potentially stifling innovation within the startup ecosystem. Current trends indicate that enterprise AI is leaning towards developing numerous smaller, custom models that present minimal risk, yet the bill's broad parameters could impose costly compliance burdens across the sector. The implications of this bill are substantial for the AI/ML community. By solidifying outdated standards from 2023, the legislation inadvertently favors established players who can absorb compliance costs while hindering emerging companies that are often the source of innovation. Dario from Anthropic has raised concerns that this could create an uneven playing field, suggesting that the definition of frontier AI should evolve alongside the rapidly advancing technology landscape. In light of this, clarity from Anthropic on their positioning regarding the 10²⁵ FLOPs threshold is crucial to ensure regulatory frameworks effectively promote safety without stifling progress.