One, we assume that training larger, more complex models continues to expand their respective capabilities, especially in domains that are monetizable. While this assumption makes sense to many, including us, there is uncertainty worth unpacking here. Model improvements, for the better part of the last half-decade, have operated under a framework known as “Scaling Laws.” Put simply, the framework posits that an AI model’s intelligence scales predictably as you increase the volume of its training data, its post-training alignment, and the time it spends pondering over a specific problem.
They are counting on models not getting faster.
They are saying "AI won't make itself faster."
Color me skeptical.