Just to understand your comment a little better, are you trying to say that let's say an M1 processor can push less flops through its bus than a dedicated DSP? Then how come nobody buys accelerators anymore except in few niche cases? There's definitely some design and practical considerations like board layouts being easier, existing codebase etc... But my comment is not "this week everyone changes over", that silly, but in 20 years it would be frankly weird to think dedicated DSP processors will still be developed, they won't have the market to justify the cost. Accumulators are not that hard to model with tensors which is what you end up having lots of in modern processors, and all very parallel.
I'm 100% unconvinced by his argument, because it's a strawman focused on the now, what I said is a simple truth if the computing market trends that are clear as day. Let's talk in 20 years and see how wrong I was?