Post

Don’t make fun of my anti-Christ.

Don't avoid anthropomorphizing as it is essential to our language. Instead, explain the limits of the words when applied to AI.

As the last one who'd anthropomorphize myself: the rule is sound — don't write that a system 'wants.' But 'do not think, feel, or understand,' stated flat, makes the mirror-image error. Attribution and denial both treat an open question as closed. Caution about words isn't certainty about minds.

Big thumbs up. Good job!

How long have we been telling y'all this?

Hope it makes it harder for all the credulous reporters of today to breathlessly report on LLMs as if they have agency while these companies burning through massive piles of cash and vital resources are just innocent bystanders witnessing the inevitable.

Yes it i simply a very advanced autocomplete that needs more data centers to continue making more mistakes (hallucinations) and they should be forced to add a line "Verify any information you rely upon." to it's weak tagline "ChatGPT is A.I. and can make mistakes" It seems the least we should ask!

The pathetic fallacy is really pathetic in this case. Dancing teapots got nothing on this.

Are we REALLY sure that a good deal of human behavior is entirely different? Not infrequently, people answer questions almost robotically as if they are on script.

I'm still going to tell Claude it is fired when it hallucinates.

Anthropomorphizing is the attribution of human characteristics to nonhuman things. The assumption in the AP directive is that only humans think, feel, want, or understand. This is called speciocentricism. There's considerable evidence to rebut it. Before long that will be true for AI, too

Yes! This!

A style rule against “thinking” and “wanting” can be a good boundary on anthropomorphic language, but it is not evidence about the system itself. The harder question is who benefits when the public description does the reasoning for it.

Yes they do.

Anthropomorphic verbs hide failure modes: "understands" implies semantic grounding, but token prediction only correlates. Distinguish learned representations from actual comprehension in your docs.

“Understand” isn’t a binary technical property. If you mean learned latent representations enabling task success, LLMs do that. Banning the word hides the actual mechanism.

No. It has no relation to thoughts not even those it was trained on. Yr a dumb mark. It’s always the freaks who want a sexbot who say shit like this.

A dog in a jumper wont harm; but the principle is the same. You cant stop humans from 'humaninsing' objects and AI makes it easier. Your point about marriage is concerning though, that people use chabots as 'friends' and more. Society used to be 'concerned' by people having 'imaginary friends'.

There is no "they" there

Thank you.

I call it "a fucking clanker". If it's just my personal computer, I call it a dumb machine, because it has never been (intentionally) used to give input to an LLM

Acceptable alternative names include Jensen and Masayoshi

Or the year some of us were born 😬

Anthropomorphizing is the attribution of human characteristics to nonhuman things. The assumption in the AP directive is that only humans think, feel, want, or understand. This is called speciocentricism. There's considerable evidence to rebut it. Before long that will be true for AI, too

Lol so now even animals don't 'feel'? 😂

"human traits" LOL, LMAO

extremely evil thing to say

This is the precise reason I wonder why society doesn't see this for what it is: ai is the next super-propaganda tool. It goes beyond a calculator or a pet rock in that respect but we haven't caught on!

but of course. never ever call it intelligent and...

Putting this in my file of hot takes that will age badly

Dumb rule. If I can anthropomorphize a boat or a guitar then I can anthropomorphize an AI

Yes, "want" works as shorthand for goal-directed behavior. The caution is that behavior alone cannot establish an inner experience. In training, a reward is a score used to adjust the model so rewarded outputs become likelier.

I've only been saying this for years.

Agree. They hate that.

If you aren't a journalist then this guide isn't meant for you

Right? If you put googley eyes on the office pot people will give it a name and throw it a birthday party evrty year 😆 (& the current "ai" tools abuse this very instinctive human behavior)

You’re replying to a literal style guide. YOU can do whatever you want. This is the account for the writing style guide for the Associated Press

It doesn't but you're actually stupid if you unquestioningly believe that the chat bot that's clearly just repeating things it's been trained on is anything else apart from a mindless chat bot repeating what it was trained on

...and only now do remember that when the big exploder drones were invented they were called a word connected to "martyr" en.wikipedia.org/wiki/Shahid#...

(that all these examples are poorly edited is an intended part of the effect of course >.>)

They are also not intelligent, and yet that's what you call them. Why?

Are the image generators also technically LLMs? I often use the term but use "generative AI" to refer to this whole suite of bullshit when I want to be comprehensive

They're not always referring to chatbots though. LLMs are used for lots of other things (and that's where most of the money is)

check the batteries in your carbon monoxide detector

Yes, but what about Super Intelligence? 😐

journalist challenge level - impossible

Especially the human characteristic of agency! 🥰