Thomas Fuchs

I saw a comment (which I can’t find anymore) of an extremely astute observation about LLMs:

People only ever attribute human-like attributes like cognition and reasoning to chatbot LLMs—but never to e.g. image generators. Which use the same algorithms and technical implementation.

Jun 7 at 2:57pmWeb
Show threadReimarJun 7

@thomasfuchs I thought I read exactly this in this atlantic essay (https://www.theatlantic.com/philosophy/2026/06/no-artificial-intelligence-is-not-conscious/687378/) but I misremembered. It was about how we only apply the term deepfake to image generators, not text generators:

»The term deepfake traditionally refers to photos, audio, and video, but when it comes to discussions of consciousness, we need to regard text as a deepfake medium as well.«

No, Artificial Intelligence Is Not Conscious

Taken to its logical conclusion, this line of thinking is absurd—and damning.

The Atlantic
Show threadWhatisgoingonJun 7

@thomasfuchs the image generators are way worse in faking understanding than the text generators.

The generated images show clearly that the thing really doesn't understand a word of what you wrote in the prompt. If you ask for a very specific thing and that is not in the image than there is no denying.

The text generator is good at incorporating details from the prompt into the output. Even if the output is factual wrong. So the user feels heard and understood.

Show threadWhatisgoingonJun 7
@thomasfuchs when the whole LLM thing started I was shocked (still am) how useless the Touring Test really is. How easy it is to make people believe that there is another person on the other side of the chat window.
Show threadDr. Heather EtcheversJun 7
@themipper @thomasfuchs Was very tickled this week when a young woman in our chorale tried to render an Instagram sketch-style drawing (I think she wrote) of our group performing in a church and it turned Jesus on the cross into a woman in crop top, miniskirt and apparent pantyhose, given the sheen of the legs. It could've done our reputation a bit of damage 😬
Show threadAzuaronJun 7

@thomasfuchs Uhhh... no they DON'T use the same algorithms and technical implementation. Stable Diffusion is not the same thing as LLM/Markov chains. Not in algorithm, not in technical implementation.

This is actually one of my big complaints about "AI" discourse right now: "AI" doesn't mean anything specific, and "AI" boosters point at useful "AI" (like some machine learning algorithms) and use that to excuse ridiculous and wasteful "AI" (LLMs).

Different things are different, and it's important to be able to talk about those differences.

Note: I wouldn't consider either LLMs or Stable Diffusion to be "good" or "useful".

Show threadAllie!Jun 7

@Azuaron @thomasfuchs you’re right in general for deployed systems at scale but diffusion language models are something being studied and worked, if you’re interested I recommend searching arxiv; some of the recent papers are pretty neat.

Agree about the AI terminology framing, I wish there was more specificity too.

Show threadAzuaronJun 7
@dotsie @thomasfuchs Sure, and Markov chain image generators are also being worked on. This just bolsters what I'm saying: if two language models aren't even the same algorithms and technical implementations, and two image generators aren't even the same algorithms and technical implementations, then there's no way that they're all just "the same algorithms and technical implementations."
Show threadjosh susserJun 7
@thomasfuchs Perhaps related - most people assume everyone thinks in words and has some kind of internal dialogue like they do, and are shocked to find that some people think in other modes, like visually. I think it's quite typical that people assume everyone thinks the same way they do, and when they see slop text it sounds like their own thinking so they mistake it for actual thinking. Seems like an advanced form of pareidolia to me. Could an experiment look at visual thinkers (who don't think in words) and determine if they have a different bias toward images over text?
Show threadThomas FuchsJun 7
@joshsusser fwiw it’s possible that “neurotypicals” being easier to lie to/getting bullshitted is a species evolutionary adaption that makes groups more cohesive
Show threadjosh susserJun 7
@thomasfuchs That's been my working theory for a few years. NTs used to be just another neurotype, until humans developed agriculture and could live in big communities, then NTs knack for socializing in groups let them take over everything. Obviously there was an advantage for the group overall, but less obviously every other neurotype got left at a disadvantage.
Show threadOllie Boermans6d ago
@thomasfuchs @joshsusser wondering this morning how the horoscope people are doing
Show threadHåkon 🍉Jun 7
@thomasfuchs may I point you to this paper:
"IF LLMs HAVE HUMAN-LIKE ATTRIBUTES, THEN SO
DOES Age of Empires II"
https://arxiv.org/pdf/2605.31514
Show threadeclexicJun 7

@hdr @thomasfuchs

Well that's hilarious, I need to put aside time to read this at some point.