@thomasfuchs@hachyderm.io fun fact, this video is full of lies, and it is set up just to become viral. It can only convince people who have not used modern day LLMs. I did all three tests she describes and got completely opposite results. GPT thought of Adam Sandler (I found it), Claude's three first random words were marmalade, trapezoid and kaleidoscope, and its random number was 73. Both models were on the free tier for these tests (and GPT was two generations behind). Like, I could make a program in basic in my Spectrum in the 80s that would give a different number each time, how can someone believe that latest tech LLMs in 2026 can't do it?

I'm sure that many people who believed this video are generally not stupid or naive. But it's fun to see how easy it is to believe what we want to believe. Sorry folks, that video is entertaining if you want to hate on LLMs, but it's just bullshit generated to get views.

But, if you wanna believe it, believe it, I don't want to spoil your satisfaction =)

@thomasfuchs or just don't use that crap at all..

@oliverboehme I mean yes, if you don't want to. But in any case it's good to understand how it works, if only to know and explain to people that they have delusions.

@thomasfuchs @oliverboehme I agree that people should understand how these systems work (transparency is best). However I would argue that that knowledge should draw only one conclusion, which is, as @oliverboehme correctly point out, don't use that crap at all.

@thomasfuchs just tried it (albeit in french) and it just looked fine; used the default model.
I'm not saying there's a mind or that it really picked a choice at first and made me guess, but at least nothing was inconsistent in the chat.

@julienw @thomasfuchs yeah had me thinking that with hidden thinking tokens one could make a working version of guess who…I also couldn’t reproduce what she did. Maybe with a smaller model.

@flq @julienw They're hiding this now behind tokens that you can't directly see, but the principle remains exactly the same. It's just more smoke and mirrors in newer harnesses/models.

@thomasfuchs
Yes, but that is what is the most easy to improve.
It's really easy to give an LLM memory and a few tools.
The core model works just as the lady says, but none of the commercial products are merely that, like you mentioned.
@flq @julienw

@julienw @thomasfuchs

Tried once with free ChatGPT, worked finel. (Only used 7 questions to get the answer but all responses were correct.)

Of course, length may matter, but also how it updates its token analysis and weights as we go along.

I've tested it in longer conversations about things that interest me, where it veers back and forth between nonsense and useful remarks. With nudges. Occasionally I find a simple "That's wrong" can be surprisingly helpful in putting it back on track.

@glc @julienw the point here is that it doesn't have permanence and isn't a conversation partner.

every time you talk to it, the entire previous conversation gets fed back to a new instance of it (which doesn't have any memory from before) and then it generates a statistically likely reply

@thomasfuchs @julienw

That is indeed the intended point.

One that has been discussed repeatedly and widely since the "stochastic parrot" paper but has not penetrated to the audience that the post aims at.

But we are noticing that the illustration used does not appear to correspond to the actual behavior of popular models.

It works a good deal like a search engine in a linguistic space. If it finds something useful, you can use it, if not, not.

@thomasfuchs Or just don't use it. Go for a beer with a friend, and play the guess who I'm thinking about game instead.

Funner, more drinks, silliness, love of one's fellow creature.

@thomasfuchs

Many humans will pick a number with two odd digits when asked for a "random" number from 1 to 100. 17, 37, and 73 are more common than 42. Only HHGTG fans are likely to go for that "random" number

@thomasfuchs

One of the first things I say in every conversation about “AI” is that what it gives isn’t an answer, it’s a response and you need to understand it every time.

They nod their head and immediately forget it.

@edsanders2
I think most people know that to some extent.
@thomasfuchs

@thomasfuchs I hate the fact that she's using it at all, but if this vid helps even one slopophant think about what they're actually using, then I suppose it's worth the cost.

@thomasfuchs Super interesting. Both that it makes these errors and that it generates what appear to be acknowledgments of them when challenged by the user. Would Claude do any better (i.e., more person-like)?

@thomasfuchs And yet people is ignorant or dumb (or both) enough to treat it as a human. 😂

@thomasfuchs

No mind is fine, she's doing it wrong by just being thick but we have thousands of years of peeps preaching not clinging to concepts like 'mind' as special, just breeds ego as she demonstrates.

Having played Guess Who with real human children I've now concluded human children have no mind behind them either.

@thomasfuchs

You get what you feed it...

In some professional applications where you feed it huge amounts of information it is very good at getting you a summary of the exact thing you want without you wading through all the data. IF you are very precise in what you want from it. As a specialist it can make you more efficient IF you know your data.

It is not intelligent, it is not human, it is just what it is, a large language model...

And garbage in is still garbage out (...).

@thomasfuchs it's called "mind theory" and it's basic to how human communicate but LLM don"t have (nor mind nor theory and of course no mind theory)

@thomasfuchs@hachyderm.io

I avoid it wherever possible.

There are minds behind it - the pea sized ones with dollar signs for eyes of the Epstein class.

@thomasfuchs It's a good video for explaining the fallibility of an LLM to someone not fully immersed in the knowledge of what it does and that it is never to be trusted.

@thomasfuchs fantastic TL;DR.

we really need to get quality vloggers to cross post on the fediverse.

@thomasfuchs

thank you for this

@thomasfuchs
It failed the #TuringTest bigly !!!

@thomasfuchs In prior versions of ChatGPT and still in self-hosted model harnesses like Ollama, there is a "Retry" button. If you play 20 Questions/Guess Who and then use Retry after it has revealed, it will give different answers, each consistent with the previous conversation.

@thomasfuchs Long shot, but has anyone the link to a job interview with an LLM agent that someone posted a few days ago?

The interviewee responded to the agent's questions with random words/phrases, and the agent conducted the interview as though these were valid.

@InarticulateOtter hmm I remember that but don’t remember where I saw it

@thomasfuchs video source https://www.tiktok.com/@wetclaude/video/7681693797725539598 i dont know how to search on tiktok bc theres no video titles

@ambiguous_yelp oh ty, I saw it cross my Bluesky (I don’t use TikTok)

@thomasfuchs If we have to spend our time analyzing and parsing AI output, what time does it save us?

@thomasfuchs LLM's are less like artificial brains and more like the theorem about "An Infinite Number of Monkeys on an Infinite Number of Typewriters" and I don't mean that as a pejorative.

@oshox it can be useful if you have a way to verify the output (which is possible for formal languages like in programming), but for most stuff it’s just meh at best

@thomasfuchs So are we not gonna talk about the model likely pulling from racist, sexist, transphobic language of incels online to predict Serena Williams as a man? Or was that implied?

@thomasfuchs If $10,000 is deposited your cash app account or PayPal unexpectedly , what will you use it for to be honest ?
:A car
:B Bills
:c school
:d families
:e vacation
:l donation
:g food stuff
:h business