> why do I care that GPT-4 is resource intensive if it solves all of my tasks?
Why do most business not hire private armies at each location, even if it would solve all of their security needs? Utility/$ matters.
> External tools are great - but why would we expect the outcome not to be GPT-5 uses all external tools?
The fact that task specific AI may be called by GPT-n is... not a contradicition to the argument that GPT-n doesn’t obviate the need for task-specific AI, any more than the fact that the board interfaces with developers through managers means that managers existing obviates the need for developers.
LLMs like GPT-n may well end up as the most important AI-powered interactive user interface — mapping things to and from language is, after all their specialty, and language is an interface humans tend to be optimized for. But browsers are pretty much that today, and they are hardly the interesting part of the ecosystem.
And the thing with external tools—often AI themselves—is that they commoditize LLMs. There’s no moat, because the interfaces are natural languahr that any LLM can handle. And if I can have a good enough local LLM as my prinary interface, which can call out directly to external (to the LLM, but including local to me) services (including bigger, more expensice remote LLMs that add an additional failable service dependency) then I’m not screwed out of both my ptimary interface and the tools it uses when something fails remotely, and if its also cheaper to use, that’s another win.