#confidentlywrong

So many people are so confidently wrong about how AI works that it actually makes me more confident the models’ own confidently-wrong output is based on real human input. They’re just reflecting us back at ourselves — smug mistakes in, smug mistakes out.

And yeah, I’m in that pool too. I’ll happily waffle on about prompt clarity and outputs under load, and I'm happy to point and laugh at the vagina museum confidently claiming "That's how LLMs work" in the middle of a thirty or forty tweet thread where every single post is weighed down with their own hallucinations and imaginative explanations with absolutely no connection to reality, because while everyone loves the vagina museum enough for tens of thousands of people to retweet and like their nonsense, literally nobody in their audience has any real idea how the tech works beyond 'magic box that obviously can't draw the female reproductive system', because of course it can't, that's not what it's for.

I like to think (in a lazy crutch kind of way) of LLMs like a bag of scrabble tiles, and especially with images, you tip out the pieces and get mad that it doesn't spell antidisestablishmentarianism, cos you're pretty sure you put all the letters in.

Sometimes I think the most plausible explanation is that LLMs are just portals into some parallel dimension where an infinite sweatshop of overworked gig-econony ghostwriters are cranking out all our mediocre replies in some frozen time-space continuum. Honestly, if we discovered that tomorrow, nobody would even blink — it’d feel about right cos that's how we build everything already anyway.