Evolution is just a fancy way of saying stochastic gradient descent of dna in the biosphere. A SGD routine can clearly move “faster” or “slower”, as can be measured by values like “preserved mutations per century”. Whether the “direction” is the same is up for debate (it almost certainly isn’t), but to claim there’s no “speed” metric that can be associated with Evolution is just absurdist head-burying.

Gradient descent is a model of what's actually happening here in reality. Sure, it's useful in some ways and sure, you can label things in your model if it's useful for your thinking.

Frankly not sure what you get out of modeling it that way and labeling "preserved mutations per century" as "speeding up," but feel free to do so.

One issue with the map-territory violation here is that it's very easy to assume the gradient descent is optimizing for something we care about and find desirable. You are smuggling in the idea that later in the gradient descent is "better" than earlier in the gradient descent. There is nothing in reality that necessitates this. It is totally possible for human contributions to this process (model it as gradient descent if you want, model it as "speeding up" if you want) to optimize things that are in fact very very bad for everything that humans care about.