General-purpose AI today

What is general-purpose AI, and why is it in such high demand?

The Future of Life Institute defines general-purpose AI as systems with “a wide range of possible uses” that can perform many different tasks across different fields, often without substantial modification or fine-tuning.^[1] The paper names GPT-3, Codex and DALL·E 2 among its examples, so GPAI under this established definition already exists. The open question is whether the next, substantially more capable generation can be built with the industrial resources available today.

It may not be possible if greater capability requires substantially more compute than current models. More compute means more accelerators, high-bandwidth memory, advanced packaging, foundry capacity, electricity, cooling and network infrastructure. No technological group independently controls that complete supply chain. Each depends on capabilities held by competing countries, and the closer any group gets to a strategically decisive system, the stronger the incentive becomes for the others to restrict access to the parts they control. The resources may exist globally while remaining impossible for one group to assemble.