And, crucially, I'd argue that for in "chatbot" tasks those other uses are more common than arithmetic, so arbitrary focus to specifically optimize arithmetic doesn't really make sense - the bitter lesson is that we don't want to bias our architecture according to our understanding of a specific problem space but rather enable the models to learn the problem space directly from data.
Stepping one level out in the metacognition hierarchy is the key. "Learning to learn" as it were. It is only the relative ease of implementation and deployment of feedforward models like Transformers that makes it seem like we have reached an optimum but we desperately need to move beyond it before it's entrenched too thoroughly.
We probably do have to move beyond transformers, but not in the direction of such hacks, but rather towards even more general representations that could encode the whole class of all such alternate representations and then learn from data which of them work best.
Deep learning is always and only ever about representing data abstractly. The more abstractions you can make irrelevant (why would you have to learn how to do math when the base-10 perspective on ASCII-digits is already provided for you?) the more you've biased your architecture to readily learn and understand the problem space.
Intelligence doesn't exist where Divine Creator gave you access to this or that faculty. It's developing those faculties yourself by reasoning through the process of composing your own mental model about the problem.