@futurebird
1/2
Supposedly these things are good at finding correlations. But that is confusing narrowly focused, small data set, supervised research with generic LLMs.

In my personal experience, the LLMs I have access to are likely to ignore all minority opinions, new research, and claim that scantly documented problems do not exist. They can not weigh the significance of any data, so they always default to what is frequently said is more true.

@futurebird
2/2
It is a stereotype disaster waiting to influence everything. They will rob us of Science. To an LLM, and the tech billionaires who want to influence us, what is stated frequently is always true, and what is new is always wrong or suspect.

Intellectual stagnation. In the age of information, LLMs are prepping us for a new dark age. The largest LLM in the world might as well be called Aristotle.

Robbing us of Science.

@futurebird
1/3
I experimented with an LLM last year at the urging of a friend.

I invented a game called "minority opinion" where we (me & the LLM) took turns identifying theories that could replace the dominant explination, and then asked the LLM to estimate a probability, based on supporting evidence, that the paradigm could be replaced with the new idea in the future.

The LLM could list a dozen reasons why a new theory was a better fit, yet the probabilities were always astonishingly low.

@futurebird
2/3
And I could push those probabilities around by simply objecting to them. So it really is a people pleasing machine.

I knew LLM logic was worthless when the LLM chose to believe that ghosts were a more likely explanation for haunted houses than carbon monoxide poisoning. Because of the many ghosts that people claim to have personally identified.

@futurebird
3/3
Google can weigh a source for their own LLM, making it insist on a thing, but they can't weigh their own sources for credibility, other than frequency in the training data.

So, the most commonly held beliefs are automatically true and will be referenced as such by an LLM.

It's a confirmation bias machine for all of humanity.

The end of Science.

@futurebird
4/3
Make no mistake, the fact that I could have a conversation like this with a machine is a great accomplishment. Turning a huge data set and some language rules into a thing I could query for hours is astonishing.

But, current AI has a credibility problem, and that means that it is not ready as a truth telling product. And the hype outweighs the truthiness by an uncomfortable margin.

@Urban_Hermit @futurebird

There are a lot of of people I could talk to for hours and still go away being no wiser or better informed than I was before, but at least I was talking to a real person.

@futurebird
A/C
Another experiment:

I know of a data base that was populated by an early AI that 'hallucinated' details. An international kite museum, supposedly in Corpus Christi, Texas, was said to be populated by displays on "The Age of Mammoths" and "The Iron Horse" because the word "museum" took more precedence than "International Kite".

It hallucinated a lot of other generic museum like details.

A street view search of the address shows a hotel, and no museum at all for blocks around.

@futurebird not to take away from anything you've said, it's definitely true that this sort of approach leads to dismissing novel ideas. but i think there's also another level at which this kind of approach is worrying: by saying they need to sometimes dismiss the machine's output, one admits the machine is not to be trusted, yet they trust it when they can't catch it in what's (seemingly) an obvious lie.
in other words, it's trusting the liar's word unless you already know it to be false. it's trusting the allegorical wolf to take care of the sheep because if you ever see it attacking one you'd stop it.

if you know you can't trust the machine sometimes, then you can't trust the machine. and if your system is then "well i'll just catch the mistakes", then either you only intend to use it in cases where you already know the right answer (in which case, why use it), or you believe you'll somehow figure out something's wrong when you can't tell what "correct" looks like.

it's saying "if it looks correct then it's correct". which is wrong both on the false negatives end and the false positive ends at the same time.

@Urban_Hermit

I think when some people are presented with these kinds of errors they think “the LMM just made a factual mistake” they think with more data and “better software” this will not be a problem. They don’t see that what it is doing is *foundationally different* from what they are asking it to do.

That it has fallen to random CS educators, and people interested in these models to desperately try to impress upon the public the way they are being tricked makes me angry.

@the5thColumnist @Urban_Hermit

People, even people who have terrible mostly wrong ideas tend to have some guiding set of values. Even if you don’t learn much from their opinions you can learn about the philosophy that informs those opinions.

Asking an LLM *why* it said any fact or opinion is pointless. It will supply a response that sounds like a human justification but the real “reason” is always the same “it was the response you were most likely to accept as correct”

@talya

This is why AI is the ultimate Gell-Mann Amnesia machine.

@futurebird

@the5thColumnist @Urban_Hermit

I’m trying to avoid loaded phrases like “bullshitting machine” I’ve had a lot of people roll their eyes and shut down on me because “well you just hate AI” as if this is a matter of personal taste.

In reality I struggle to see how I could find value in exposing my curiosity to a system with these limitations. I will insulate myself from those times when a simple obvious question brings me up short— it just seems really dangerous to me.

@alessandro @talya

I think with topics where one isn’t an expert it can be more important to know what “most people” in your social circle *think* is true than it can be to know what is really true.

Knowing an iconoclastic truth, but not having the expertise to explain it to others isn’t very useful. Moreover without that expertise you will struggle to evaluate the validity of the unpopular opinion.

So, people reach for the popular option and hope it is correct.

@alessandro @talya

No one is enough of a polymath and no one has enough time to avoid trusting others. This isn’t really a bad thing, but we have to be open to the reality that there are some things that “everyone knows” that are simply wrong.

@futurebird @alessandro @talya Drawing a line to my favourite topic:

Virtually everything we "knew" about the cognition and social and emotional lives of reptiles before the last couple of decades was wrong (due to systemically flawed studies, and old half-truths that ended up getting quoted enough to take on a weird life on their own). Many of those wrong things are still things that "everyone knows", and they're almost certainly well-represented in a training corpus that both includes older textbooks and all of Reddit. It's completely hit-or-miss if LLMs (at least the ones I've tried) answer with some long-debunked piece of archaic trivia or something well-established backed up by a modern herpetology article. (And when they say something wrong, they're often so authoritative and frequently back it up with consistent-but-wrong "reasoning" that I have to double-check to find out if *I* am wrong).

If we'd had enough computer power and data to train LLMs in 1995, virtually every single thing they would have had to say on the topic would be wrong!

(Making a judgment on whether everything we know in 2025 is so right that something like that will never happen again in any field is left as an exercise to the reader.)

@futurebird @the5thColumnist @Urban_Hermit Why do you say this versus "most likely to be statistically correct"?

Regardless, that doesn't mean it's the right answer nor the right answer for the person asking it.

@datarama @futurebird @alessandro @talya
Could you give an example, please?
I have to confess that I'm not sure if I've ever thought about how reptiles think.

@elight @the5thColumnist @Urban_Hermit

Because the popular models people will encounter have been trained to work this way.

@Gorfram @futurebird @alessandro @talya I just answered myself with a collection of things Copilot 95 would have gotten wrong.

The short version is that reptiles are neither particularly stupid, particularly asocial or particularly emotionless. In the 60s it was commonly believed that there was a primitive "reptile complex" in the brain that only covered the most basic instinct-driven behaviour and processing of simple stimuli. This is *still* part of popular understanding; it's what people mean when they talk about "the lizard brain". The concept got popularized by (rest his soul) Carl Sagan in his book "The Dragons of Eden" from 1977.

But it turns out an actual lizard brain is nothing like that primitive extended-brainstem structure he described! Lizards have a limbic system just like mammals (and they *definitely* experience emotion), some even practice parental care (in monkeytail skinks for up to a year), live in stable family groups and adopt orphans, and some others form lifelong monogamous pairs (shingleback skinks are famous for this; they even have a range of behaviour that is hard to explain as anything other than a kind of grief when a partner dies).

It used to be commonly accepted that reptiles were incapable of learning novel problem-solving behaviour, because they don't have a neocortex (which is what mammals use to do that). But it turns out that the dorsal ventricular ridge (which reptiles share with birds) plays a similar role to the mammalian neocortex, and packs an enormous amount of neurons into a comparatively tiny volume. Until 1997, we thought it was just a ganglion that handled motor control.

Little anoles with brains the size of pinheads do better on the worm-in-a-lid puzzle than most songbirds and can even learn by looking at other anoles. Monitor lizards can count to 6, tortoises can navigate mazes using something that resembles depth-first-search, crocodilians can put twigs and other items on their heads to use as lures for birds - even adapting which kinds of material they use according to which birds are currently building nests. Many can differentiate individual conspecifics *and* tell the difference between individual humans.