Not just for this article, but from most ideas/articles around LLMs, I feel like they aren't "thinking with portals" enough. We have "portal gun" tech (or at least, that's what's being marketed), and we're using it as better doors.
I sorta think the issue is that what LLMs do in and of themselves is extend text in a coherent way, while only a small subset of applications are directly textual. It’s incredibly generally applicable yet also difficult to apply to anything that isn’t a glorified text editor. Say you wanted to have it help you edit videos. You might provide it with a scripting language to control the editor , but now you have to maintain parity between a scripting language and the editor’s user-accessible functionality. If you’re adobe, is that really worth the manpower? If you’re a small startup trying to unseat adobe, you have to compete with decades of features and user lock in. The only way this makes sense for either party is if the LLM is crazy good at it, but the LLM can’t watch its video output and it’s also probably just okay to begin with.
