I've talked in the past about how chatbot UX feels wrong.

It's too limiting, leans too hard into the artifice that the LLM is "just a normal human", when in reality it's a collective hive mind of humanity that has orders of magnitude better recall than any human... and also lacks some aspects of common sense.

I see dialogues with LLMs as not a conversation but an act of co-creation, and a chatbot UX seems to undermine that, and foreclose^[adi] on other interactions.

An alternate UX is kind of like a Google Doc, where you could hit Shift-Enter to ask the LLM to autocomplete from your cursor for as long as it wanted to.

It is more of a "based on what comes before in this doc, keep generating as many tokens as you want, drawing on your hive mind of humanity."^[adj][adk]

You could still do a normal append-only chat log if you wanted, just each time before you hit Shift-Enter, you'd type: "Computer:".

But you don't have to do that if you don't want to, and you can also clean up the stuff earlier in the doc to hone the context to get closer to what you want, to steer it better.

Kind of like GitHub Copilot but where instead of tab adding a few characters, it could add whole paragraphs.

Because the key command is not simply "Enter" (which would add LLM-generated slop after every line you write), the user has to choose to call it into being.

This also allows patterns where you start the LLM off going down a particular line of thought, or combine insights from another LLM's thinking tokens to help direct another LLM's conclusions.

Kind of a simple, obvious interaction pattern, and not that different from chatbot UXes, and also more similar to the original LLM completion APIs from a couple of years ago.

But it feels like it's truer to the materials and what they're good at than the current chatbox interfaces.^[adl]