Open AI is a force for fundamental innovation

This post is a subclaim

Proprietary AI is just that: proprietary. Any innovations are limited to the company that created it, without benefit to society. But when AI is open, innovations can be shared with everyone.

Open research

AI research has been picking up, with the number of AI papers doubling every 2 years . The topics vary, but there have been many recent papers about improving efficiency such as FlashAttention and Mamba.

These new formats can be applied to a whole language model, increasing speed and decreasing energy consumption. Switching a frequently called language model to a more efficient format could save many kilowatt hours, given that generating just one response takes around 1 Wh.

New architectures

Makers of new chip architectures would have a hard time without open AI - they would have to directly convince corporations to run their language model on their chips. Instead, chipmakers can create chips designed for AI , and run public language models on them. From there, they can host the service, with competitive speed and pricing.