> So, LLMs face a regression on their latest proposed improvement. Arguably a second regression, the first being cost, because COT improves performance by scal

> So, LLMs face a regression on their latest proposed improvement. Arguably a second regression, the first being cost, because COT improves performance by scaling up the amount of compute used at inference time instead of training time. The promise of LLMs was that you do expensive training once and then run the model cheaply forever, but now we're talking about expensive training followed by expensive inference every time you run the model.