Getting a cat:

My depression 📉📉📉
Crime rate in my house 📈📈📈

weird niche question, but: would one of the benefits of smaller models be the *kind* of error they make? Like, for auto transcription, if my choice is between "never gets names right" and "fabricates whole chunks of conversation" - I'll take the name errors! At least I can spot those!

LLMs are a near dead-end technology. Like all ML systems they can never be used when the output needs to be correct. And because an LLM doesn't separate code and data it can never be used where it receives untrusted input and can do a bad thing.

That's the part that jumped out at me too!

If they want to keep getting work, they can review (=rewrite) the AI slop, for 1/3 of their normal rate (which was already low to begin with)

Readers, we HAVE to complain about it, hold the publishers accountable and boycott these books

--than even the Pareto-ish 90/9/1 he originally identified. The median Bluesky user, and I mean active daily user and not just inactive accounts, has never posted or replied to anyone, once, ever. But they're still using Bluesky!

Any site has to balance the things their most active users want with things their majority of users want, and the absolutely fiendish problem of social media product design is those two desires often actively contradict each other.