hmm i think we need twelve more articles on this
Humans are bullshit machines as well.
This is what I find the most amusing about the criticism of LLMs and many other AI systems aswell. People often talk about them as if they’re somehow uniquely flawed, while in reality what they’re doing isn’t that different from what humans do aswell. The biggest difference is that when a human hallucinates it’s often obvious but when chatGPT does that it’s harder to spot.
This is… really not true at all.
LLMs differ from humans in a very very important way when it comes to language: we know the meanings of the words we use. LLMs do not “know” things, are unconcerned with “meanings”, and thus cannot be said to be “using” words in any meaningful way.
we know the meanings of the words we use.
Uh, but we don’t? Not really. People use the wrong words all the time and each person’s definition (i.e., encoding) is slightly different. We mimic phrases and structures we’ve heard to sound smarter and forge on with uncertain statements because frequently they go unchallenged or simply aren’t important.
We’re more structurally complex than a LLM, but we fool ourselves in thinking we’re somehow uniquely thoughtful and reliable.
And everyone in tech who has worked on ML before collectively says “yeah that’s what we’ve been trying to tell you”. Don’t get me wrong, LLMs are a huge leap, but god did it show how greedy corporations are, just immediately jumping to “how quick can we lay people off?”. The tech is not to that spec. Yet. It will get there, but goddamn do we need to be demanding some regulations now
The tech is not to that spec. Yet.
I’m not sure it will. At least, not this tech, not this approach to the problem. From my understanding there’s fundamentally no comprehension; it’s not bugged, broken, or incomplete, it’s just not there… it’s missing from the design.
We don’t know that for sure yet, we saw a lot of emergent intelligent properties appear as we scaled up, and we’re nowhere near done scaling LLM’s, I’m not saying it will be solved, just that we don’t know one way or the other yet.
LLMs are fundamentally different from human consciousness. It isn’t a problem of scale, but kind.
They are like your phone’s autocomplete, but very very good. But there’s no level of “very good” for autocomplete that makes it a human, or will give it sentience, or allow it to understand the words it is suggesting. It simply returns the next most-likely word in a response.
If we want computerized intelligence, LLMs are a dead end. They might be a good way for that intelligence to speak pretty sentences to us, but they will never be that themselves.
So for context, I am an applied mathematician, and I primarily work in neural computation. I have an essentially cursory knowledge of LLMs, their architecture, and the mathematics of how they work.
I hear this argument, that LLMs are glorified autocomplete and merely statistical inference machines and are therefore completely divorced from anything resembling human thought.
I feel the need to point out that not only is there no compelling evidence that any neural computation that humans do anything different from a statistical inference machine, there’s actually quite a bit of evidence that that is exactly what real, biological neural networks do.
Now, admittedly, real neurons and real neural networks are way more sophisticated than any deep learning network module, real neural networks are extremely recurrent and extremely nonlinear, with some neural circuits devoted to simply changing how other neural circuits process signals without actually processing said signals on their own. And in the case of humans, several orders of magnitude larger than even the largest LLM.
All that said, it boils down to an insanely powerful statistical machine.
There are questions of motivation and input: we all want to stay alive (ish), avoid pain, and have constant feedback from sensory organs while a LLM just produces what it was supposed to. But in an abstraction the ideas of wants and needs and rewards aren’t substantively different from prompts.
Anyway. I agree that modern AI is a poor substitute for real human intelligence, but the fundamental reason is a matter of complexity, not method.
Some reading:
Large scale neural recordings call for new insights to link brain and behavior
A unifying perspective on neural manifolds and circuits for cognition
a comparison of neuronal population dynamics measured with calcium imaging and electrophysiology
If you truly believe humans are simply autocompletion engines then I just don’t know what to tell you. I think most reasonable people would disagree with you.
Humans have actual thoughts and emotions; LLMs do not. The neural networks that LLMs use, while based conceptually in biological neural networks, are not biological neural networks. It is not a difference of complexity, but of kind.
Additionally, no matter how many statistics, CPU power, or data you give an LLM, it will not develop cognition because it is not designed to mimic cognition. It is designed to link words together. It does that and nothing more.
A dog is more sentient than an LLM in the same way that a human is more sentient than a toaster.
“most reasonable people” - indirect ad hominem is still ad hominem. You are making a fool of yourself.
In a more diplomatic reading of your post, I’ll say this: Yes, I think humans are basically incredibly powerful autocomplete engines. The distinction is that an LLM has to autocomplete a single prompt at a time, with plenty of time between the prompt and response to consider the best result, while living animals are autocompleting a continuous and endless barrage of multimodal high resolution prompts and doing it quickly enough that we can manipulate the environment (prompt generator) to some level.
Yeah biocomputers are fucking wild and put silicates to shame. The issue I have is with considering biocomputation as something that fundamentally cannot be be done by any computational engine, and as far as neural computation is understood, it’s a really sophisticated statistical prediction machine
We all want to believe that humans, or indeed animals as a whole, have some secret special sauce that makes us fundamentally distinguishable from statistical algorithms that approximate a best fit function according to some cost metric, but the fact of the matter is we don’t.
There is no science to support the idea that biological neurons are particularly special, and there are reams and reams of papers suggestin that real neural cognition is little more than an extremely powerful statistical machine.
I don’t care about what “most reasonable people” think. “Most reasonable people” don’t have an opinion about the axiom of choice, or the existence of central pattern generators. That’s not to devalue them but their opinions on things this far outside of their expertise are worth about as much as my opinions on the concept of art. I am a professional in neural computation, and I put it to you to even hypothesize about how animal neural computation is fundamentally distinct from LLM computation.
Like I said, we are wildly more capable than GPT, because our hardware is wildly more complex than any ANN, but the fundamental computing strategy is not all that different.
You’re guessing, you don’t actually know that for sure, it seems intuitively correct, but we simply do not know enough about cognition to make that assumption.
Perhaps our ability to reason exclusively comes from our ability to predict, and by scaling up the ability to predict, we become more and more able to reason.
These are guesses, all we have now are guesses, you can say “it doesn’t reason” and “it’s just autocorrect” all you want, but if that were the case why did scaling it up eventually enable it to perform basic math? Why did scaling it up improve its ability to problemsolve significantly (gpt3 vs gpt4), there’s so many unknowns in this field, to just say “nah, can’t be, it works differently from us” doesn’t mean it can’t do the same things as us given enough scale.
I’m not guessing. When I say it’s a difference of kind, I really did mean that. There is no cognition here; and we know enough about cognition to say that LLMs are not performing anything like it.
Believing LLMs will eventually perform cognition with enough hardware is like saying, “if we throw enough hardware at a calculator, it will eventually become alive.” Even if you throw all the hardware in the world at it, there is no emergent property of a calculator that would create sentience. So too LLMs, which really are just calculators that can speak English. But just like calculators they have no conception of what English is and they do not think in any way, and never will.
I’m not guessing. When I say it’s a difference of kind, I really did mean that. There is no cognition here; and we know enough about cognition to say that LLMs are not performing anything like it.
We do not know that, I challenge you to find a source for that, in fact, i’ve seen sources showing the opposite, they seem to reason in tokens, for example, LLM’s perform significantly better at tasks when asked to give a step by step reasoned explanation, this indicates that they are doing a form of reasoning, and their reasoning is limited by what I have no better term for than laziness.
https://blog.research.google/2022/05/language-models-perform-reasoning-via.html
It is your responsibility to prove your assertion that if we just throw enough hardware at LLMs they will suddenly become alive in any recognizable sense, not mine to prove you wrong.
You are anthropomorphizing LLMs. They do not reason and they are not lazy. The paper discusses a way to improve their predictive output, not a way to actually make them reason.
But don’t take my word for it. Go talk to ChatGPT. Ask it anything like this:
“If an LLM is provided enough processing power, would it eventually be conscious?”
“Are LLM neural networks like a human brain?”
“Do LLMs have thoughts?”
“Are LLMs similar in any way to human consciousness?”
Just always make sure to check the output of LLMs. Since they are complicated autosuggestion engines, they will sometimes confidently spout bullshit, so must be examined for correctness. (As my initial post discussed.)
Are you just fucking around here? C’mon. In your hypothetical scenario, the emergent property would not be “of a calculator”.
I am picking up a hint of the autocompletion you describe, in your writing.
deleted by creator
A chatbot has no use for that, it’s just there to mush through lots of data and produce some, it doesn’t have or should worry about its own existence.
It literally can’t worry about its own existence; it can’t worry about anything because it has no thoughts or feelings. Adding computational power will not miraculously change that.
Add some long term memory, bigger prompts, bigger model, interaction with the Web, etc. and you can build a much more powerful bit of software than what we have today, without even any real breakthrough on the AI side.
I agree this would be a very useful chatbot. But it is still not a toaster. Nor would it be conscious.
You seem unfamiliar with the concept of consciousness as an emergent property.
What if we dramatically reduce the cost of training - what if we add realtime feedback mechanisms as part of a perpetual model refinement process?
As far as I’m aware, we don’t know.
How are you so confident that your feelings are not simply a consequence of complexity?
Even if they are a result of complexity, that still doesn’t change the fact that LLMs will never be complex in that manner.
Again, LLMs have no self-awareness. They are not designed to have self-awareness. They do not have feelings or emotions or thoughts; they cannot have those things because all they do is generate words in response to queries. Unless their design fundamentally changes, they are incompatible with consciousness. They are, as I’ve said before, complicated autosuggestion algorithms.
Suggesting that throwing enough hardware at them will change their design is absurd. It’s like saying if you throw enough hardware at a calculator, it will develop sentience. But a calculator will not do that because all it’s programmed to do is add numbers together. There’s no hidden ability to think or feel lurking in its design. So too LLMs.
Suppose you were saying that about me. How would I prove you wrong? How could a thinking being express that it is actually sentient to meet your standards?
By telling me you are.
If you ask ChatGPT if it is sentient, or has any thoughts, or experiences any feelings, what is its response?
But suppose it’s lying.
We also understand the math underlying it. Humans designed and constructed it; we know exactly what it is capable of and what it does. And there is nothing inside it that is capable of thought or feeling or even rationality.
It is a word generation algorithm. Nothing more.
I don’t believe in scaling as a way to discover understanding. Doing that is just praying that the machine comes alive… these machines weren’t programmed to come alive in that way. That’s my fundamental argument, the design of LLMs ignores understanding of the content… it doesn’t matter how much content it’s been scaled up to.
If I teach a real AI about fishing, it should be able to reason about fishing and it shouldn’t need to have read a supplementary knowledge of mankind to do it.
What the LLMs seem to be moving towards is more of a search and summary engine (for existing content). That’s a very similar and potentially quite useful thing, but it’s not the same thing as understanding.
It’s the difference between the kid that doesn’t know much but is really good at figuring it out based on what they know vs the kid that’s read all the text books front to back and can’t come up with anything original to save their life but can quickly regurgitate and summarize anything they’ve ever read.
If I teach a real AI about fishing, it should be able to reason about fishing and it shouldn’t need to have read a supplementary knowledge of mankind to do it.
This is a faulty assumption.
In order for you to learn about fishing, you had to learn a shitload about the world. Babies don’t come out of the womb able to do such tasks, there is a shitload of prerequisite knowledge in order to fish, it’s unfair to expect an ai to do this without prerequisite knowledge.
Furthermore, LLM’s have been shown to do many things that aren’t in their training data, so the notion that it’s a stochastic parrot is also false.
Furthermore, LLM’s have been shown to do many things that aren’t in their training data, so the notion that it’s a stochastic parrot is also false.
And (from what I’ve seen) they get things wrong with extreme regularity, increasingly so as thing diverge from the training data. I wouldn’t say they’re a “stochastic parrot” but they don’t seem to be much better when things need to be correct… and again, based on my (admittedly limited) understanding of their design, I don’t anticipate this technology (at least without some kind of augmented approach that can reason about the substance) overcoming that.
In order for you to learn about fishing, you had to learn a shitload about the world. Babies don’t come out of the womb able to do such tasks, there is a shitload of prerequisite knowledge in order to fish, it’s unfair to expect an ai to do this without prerequisite knowledge.
That’s missing the forest for the trees. Of course an AI isn’t going to go fishing. However, I should be able to assert some facts about fishing and it should be able to reason based on those assertions. e.g. a child can work off of facts presented about fishing, “fish are hard to catch in muddy water” -> “the water is muddy, does that impact my chances of a catching a bluegill?” -> “yes, it does, bluegill are fish, and fish don’t like muddy water”.
There are also “teachings” brought about by how these are programmed that make the flaws less obvious, e.g., if I try to repeat the experiment in the post here Google’s Bard outright refuses to continue because it doesn’t have information about Ryan McGee. I’ve also seen Bard get notably better as it’s been scaled up, early on I tried asking it about RuneScape and it spewed absolute nonsense. Now… It’s reasonable-ish.
I was able to reproduce a nonsense response (once again) by asking about RuneScape. I asked how to get 99 firemaking, and it invented a mechanic that doesn’t exist “Using a bonfire in the Charred Stump: The Charred Stump is a bonfire located in the Wilderness. It gives 150% Firemaking experience, but it is also dangerous because you can be attacked by other players.” This is a novel (if not creative) invention of Bard likely derived from advice for training Prayer (which does have something in the Wilderness which gives 350% experience).
And (from what I’ve seen) they get things wrong with extreme regularity, increasingly so as thing diverge from the training data. I wouldn’t say they’re a “stochastic parrot” but they don’t seem to be much better when things need to be correct… and again, based on my (admittedly limited) understanding of their design, I don’t anticipate this technology (at least without some kind of augmented approach that can reason about the substance) overcoming that.
Keep in mind, you’re talking about a rudimentary, introductory version of this, my argument is that we don’t know what will happen when they’ve scaled up, we know for certain hallucinations become less frequent as the model size increases (see the statistics on gpt3 vs 4 on hallucinations), perhaps this only occurs because they haven’t met a critical size yet? We don’t know.
There’s so much we don’t know.
That’s missing the forest for the trees. Of course an AI isn’t going to go fishing. However, I should be able to assert some facts about fishing and it should be able to reason based on those assertions. e.g. a child can work off of facts presented about fishing, “fish are hard to catch in muddy water” -> “the water is muddy, does that impact my chances of a catching a bluegill?” -> “yes, it does, bluegill are fish, and fish don’t like muddy water”.
https://blog.research.google/2022/05/language-models-perform-reasoning-via.html
they do this already, albeit imperfectly, but again, this is like, a baby LLM.
and just to prove it:
https://chat.openai.com/share/54455afb-3eb8-4b7f-8fcc-e144a48b6798
And we’re nowhere near dome scalimg LLM’s
I think we might be, I remember hearing openAI was training on so much literary data that they didn’t and couldn’t find enough for testing the model. Though I may be misrememberimg.
deleted by creator
I was mostly posting this because the last time LLMs came up, people kept on going on and on about how much their thoughts are like ours and how they know so much information. But as this article makes clear, they have no thoughts and know no information.
In many ways they are simply a mathematical party trick; formulas trained on so much language, they can produce language themselves. But there is no “there” there.
They are the perfect embodiment of the internet.
They know everything, but understand nothing
Sadly we don’t even know what “knowing” is, considering human memory changes every time it is accessed. We might just need language and language only. Right now they’re testing if generating verbalized trains of thought helps (it might?). The question might change to: Does the sum total of human language have enough consistency to produce behavior we might call consciousness? Can we brute force the Chinese room with enough data?
have no thoughts
True
know no information
False. There’s plenty of information stored in the models, and plenty of papers that delve into how it’s stored, or how to extract or modify it.
I guess you can nitpick over the work “know”, and what it means, but as someone else pointed out, we don’t actually know what that means in humans anyway. But LLMs do use the information stored in context, they don’t simply regurgitate it verbatim. For example (from this article):
If you ask an LLM what’s near the Eiffel Tower, it’ll list location in Paris. If you edit its stored information to think the Eiffel Tower is in Rome, it’ll actually start suggesting you sights in Rome instead.
They only use words in context, which is their problem. It doesn’t know what the words mean or what the context means; it’s glorified autocomplete.
I guess it depends on what you mean by “information.” Since all of the words it uses are meaningless to it (it doesn’t understand anything of what it either is asked or says), I would say it has no information and knows nothing. At least, nothing more than a calculator knows when it returns 7 + 8 = 15. It doesn’t know what those numbers mean or what it represents; it’s simply returning the result of a computation.
So too LLMs responding to language.
Why is that a problem?
For example, I’ve used it to learn the basics of Galois theory, and it worked pretty well.
- The information is stored in the model, do it can tell me the basics
- The interactive nature of taking to LLM actually helped me learn better than just reading.
- And I know enough general math so I can tell the rare occasions (and they indeed were rare) when it makes things up.
- Asking it questions can be better than searching Google, because Google needs exact keywords to find the answer, and the LLM can be more flexible (of course, neither will answer if the answer isn’t in the index/training data).
So what if it doesn’t understand Galois theory - it could teach it to me well enough. Frankly if it did actually understand it, I’d be worried about slavery.
I’ve been unemployed for 7 months. Every online job I see that’s been posted for at least 6 hours has over 200 applications. I’m a senior Dev with 30 years experience, and I can’t find work.
I’d say generative AI is an existential threat as bad as offshoring was for steel in the early 80s. I’m now left with the prospect of spending the last 20 years of my work life at or near minimum wage.
After all, I can’t afford to spend $250,000 on a new bachelor’s degree, and a community college degree might get me to $25/hr, and still costs thousands. This is causing impoverishment on a massive scale.
Ignore this threat at your peril.
Your issue sounds more like a capitalism issue. FANG companies lay off thousands of employees to cut costs and prepare for changes in the economy. AI didn’t make them lay off all those employees, just corporate greed. Until AI can gather requirements, accurately produce code with at least 80%, can compile the software itself, it isn’t a threat.
Edit fix autocorrect
Hard to believe a senior dev can’t find work. Those positions are the most needed. Also 25 an hour is 50k a year. No where in the US are senior devs paid that little. I suppose you may not be US based, but your cost for college seems to imply US, albeit at an expensive school.
deleted by creator
No one is saying there’s problems with the bots (though I don’t understand why you’re being so defensive of them – they have no feelings so describing their limitations doesn’t hurt them).
The problem is what humans expect from LLMs and how humans use them. Their purposes is to string words together in pretty ways. Sometimes those ways are also correct. Being aware of what they’re designed to do, and their limitations, seems important for using them properly.
deleted by creator
What else should they be?? They reflect human language.
People think they are actually intelligent and perform reasoning. This article discusses how and why that is not true.
deleted by creator
I would encourage you to ask ChatGPT itself if it is intelligent or performs reasoning.
deleted by creator
I think it’s hilarious you aren’t listening to anyone telling you you’re wrong, even the bot itself. Must be nice to be so confident.
deleted by creator
“has a model of how words relate to each other, but does not have a model of the objects to which the words refer.
It engages in predictive logic, but cannot perform syllogistic logic - reasoning to a logical conclusion from a set of propositions that are assumed to be true”
Is this true of all current LLMs?
does not have a model of the objects to which the words refer
I’m not even sure what this is supposed to be saying. Sounds kind of like a bullshit generator.
Words are encodings of knowledge and their expression and use represent that knowledge, and these machines ingest a repository containing a significant percent of written human communication. It encodes that the words “dog” and “bark” are often used together, but it also encodes that “dog” and “cat” are things that are both “mammals” and “mammals” are “animals”, and that the pair of them are much more likely to appear in a human household than a “porpoise”. What is this other kind of model of objects that hasn’t been in some way represented in all of the internet?
It is not a model of objects. It’s a model of words. It doesn’t know what those words themselves mean or what they refer to; it doesn’t know how they relate together, except that some words are more likely to follow other words. (It doesn’t even know what an object is!)
When we say “cat,” we think of a cat. If we then talk about a cat, it’s because we love cats, or hate them, or want to communicate something about them.
When an LLM says “cat,” it has done so because a tokenization process selected it from a chain of word weights.
That’s the difference. It doesn’t think or reason or feel at all, and that does actually matter.
This is just the same hand-waving repeated. What does it mean to “know what a word means”? How is a word, indexed into a complex network of word embeddings, meaningfully different as a token from this desired “object model”? Because the indexing and encoding very much does relate words together separately from their likelihood to appear in a sentence together. These embeddings may be learned from language, but language is simply a method of communicating meaning, and notably humans also learn meaning through consuming it.
What do things like “love” or “want” or “feeling” have to do with a model of objects? How would you even recognize a system that does that and why would it be any more capable than a LLM at producing good and trustable information? Does feeling love for a concept help you explain what a random blogger does? Do you need to want something to produce meaningful output?
This just all seems like poorly defined techno-spiritualism.
What does it mean to “know what a word means”?
For one, ChatGPT has no idea what a cat or dog looks like. It has no understanding of their differences in character of movement. Lacking that kind of non-verbal understanding, when analysing art that’s actually in its domain, that is, poetry, it couldn’t even begin to make sense of the question “has this poem feline or canine qualities” – best it can do is recognise that there’s neither cats nor dogs in it and, being stumped, make up some utter nonsense. Maybe it has heard of catty and that dogs are loyal and will be looking for those themes, but feline and canine as in elegance? Forget it, unless it has read a large corpus of poet analysis that uses those terms: It can parrot that pattern matching, but it can’t do the pattern matching itself, it cannot transfer knowledge from one domain to another when it has no access to one of those domains.
And that’s the tip of the iceberg. As humans we’re not really capable of purely symbolic thought so it’s practically impossible to appreciate just how limited those systems are because they’re not embodied.
(And, yes, Stable Diffusion has some understanding of feline vs. canine as in elegance – but it’s an utter moron in other areas. It can’t even count to one).
Then, that all said, and even more fundamentally, ChatGPT (as all other current AI algos we have) is a T2 system, not a T3 system. It comes with rules how to learn, it doesn’t come with rules enabling it to learn how to learn. As such it never thinks – it cannot think, as in “mull over”. It reacts with what passes as a gut in AI land, and never with “oh I’m not sure about this so let me mull it over”. It is in principle capable of not being sure but that doesn’t mean it can rectify the situation.
deleted by creator
It is not hand-waving; it is the difference between an LLM, which, again, has no cognizance, no agency, and no thought – and humans, which do. Do you truly believe humans are simply mechanistic processes that when you ask them a question, a cascade of mathematics occurs and they spit out an output? People actually have an internal reality. For example, they could refuse to answer your question! Can an LLM do even something that simple?
I find it absolutely mystifying you claim you’ve studied this when you so confidently analogize humans and LLMs when they truly are nothing alike.
no cognizance, no agency, and no thought
Define your terms. And explain why any of them matter for producing valid and “intelligent” responses to questions.
Do you truly believe humans are simply mechanistic processes that when you ask them a question, a cascade of mathematics occurs and they spit out an output?
Why are you so confident they aren’t? Do you believe in a soul or some other ephemeral entity that wouldn’t leave us as a biological machine?
People actually have an internal reality. For example, they could refuse to answer your question! Can an LLM do even something that simple?
Define your terms. And again, why is that a requirement for intelligence? Most of the things we do each day don’t involve conscious internal planning and reasoning. We simply act and if asked will generate justifications and reasoning after the fact.
It’s not that I’m claiming LLMs = humans, I’m saying you’re throwing out all these fuzzy concepts as if they’re essential features lacking in LLMs to explain their failures in some question answering as something other than just a data problem. Many people want to believe in human intellectual specialness, and more recently people are scared of losing their jobs to AI, so there’s always a kneejerk reaction to redefine intelligence whenever an animal or machine is discovered to have surpassed the previous threshold. Your thresholds are facets of the mind that you both don’t define, have no means to recognize (I assume your consciousness, but I cannot test it), and have not explained why they’re important for fact rather than BS generation.
How the brain works and what’s important for various capabilities is not a well understood subject, and many of these seemingly essential features are not really testable or comparable between people and sometimes just don’t exist in people, either due to brain damage or a simple quirk in their development. The people with these conditions (and a host of other psychological anomalies) seem to function just fine and would not be considered unthinking. They can certainly answer (and get wrong) questions.
deleted by creator
Do you truly believe humans are simply mechanistic processes that when you ask them a question, a cascade of mathematics occurs and they spit out an output? People actually have an internal reality.
Those two things can be true at the same time.
I find it absolutely mystifying you claim you’ve studied this when you so confidently analogize humans and LLMs when they truly are nothing alike.
“Nothing alike” is kinda harsh, we do have about as much in common with ChatGPT as we have with flies purpose-bred to fly left or right when exposed to certain stimuli.
They’re both BS machines and fact generators. It produced bullshit when asked about him because as far as I can tell he’s kind of a nobody, not because it’s just a stylistic generator. If he asked about a more prominent person likely to exist more significantly within the training corpus, it would likely be largely accurate. The hallucination problem stems from the system needing to produce a result regardless of whether it has a well trained semantic model for the question.
LLMs encode both the style of language and semantic relationships. For “who is Einstein”, both paths are well developed and the result is a reasonable response. For “who is Ryan McGreal”, the semantic relationships are weak or non-existent, but the stylistic path is undeterred, leading to the confidently plausible bullshit.
They don’t generate facts, as the article says. They choose the next most likely word. Everything is confidently plausible bullshit. That some of it is also true is just luck.
It’s obviously not “just” luck. We know LLMs learn a variety of semantic models of varying degrees of correctness. It’s just that no individual (inner) model is really that great, and most of them are bad. LLMs aren’t reliable or predictable (enough) to constitute a human-trustable source of information, but they’re not pure gibberish generators.
