But what's it good for?

I was in my teens when the Internet first started gaining ground among normies.  I remember being at my cousins' house and they were showing off their new AOL connection.  I remember being unimpressed.  I didn't see the point then, because AOL was among the least compelling things you can do with an internet connection.  Why would I want to talk to strangers?

I feel like I'm missing the point of LLM/AI the exact same way.  Why would I want to chat with an LLM instead of using a search engine and getting facts straight from the source?  Why would I want to let the average of all of StackOverflow anywhere near my codebase? Why do people love The Machine that Lies to You?

The marketing/intuition around the Internet in those early days was that it would connect everyone to each other, facilitating human contact and a new golden age of empathy and understanding.  Yeah, we were deluding ourselves.  But it turns out, what the internet is actually good for is almost the exact opposite of direct person-to-person contact (sure it does that, but that's not where the value is).  What the Internet is good for is avoiding human contact when I don't want it.  To the point where I'm genuinely annoyed when I have to actually call a restaurant to place a to-go order.

Maybe the real value of LLMs is not conversations, or answering questions, or writing code.  Maybe it's somehow the inversion of those things.  What would that even mean?  Here's some half-baked bullshit from a real person for a change.

LLMs are, at their core, pattern-matching completion engines.  Stream of tokens in -> plausible stream of tokens to continue out.  We're using mostly words as tokens now.  Giving explicit instructions and carefully crafted prompts to see what the machine responds with.  But I don't want a synthetic employee/teacher/servant.  I want to extend myself.  I want superpowers.  I want a smart home that is actually smart.  I want to know where I left a thing in my cluttered home and know without asking.  Or better yet, I want my home to de-clutter organise itself.

I don't want an "agent" submitting a flood of PRs to my codebase - (not an emdash) shifting my job from creation of something I understand deeply to maintenance of something I have a surface level understanding of.  I want to write code fluently and effortlessly.  I want to grok the effects of a change without running a test suite.  I want a little angel on my shoulder ready to notify me of potential bugs, to highlight areas of my code affected by upcoming deprecations and policy changes.

The current chat-based state of LLMs is borderline unusable garbage.  We haven't uncovered the real killer UX.

The lies are gonna be a problem too.