I think a thread on this topic will be interesting. My own position is that AI is intelligent, and that’s for a very simple reason: it can do things that require intelligence. That sounds circular, and in one sense it is. In another sense it isn’t. It’s a way of saying that we don’t have to examine the internal workings of a system to decide that it’s intelligent. Behavior alone is sufficient to make that determination. Intelligence is as intelligence does.
You might ask how I can judge intelligence in a system if I haven’t defined what intelligence actually is. My answer is that we already judge intelligence in humans and animals without a precise definition, so why should it be any different for machines? There are lots of concepts for which we don’t have precise definitions, yet we’re able to discuss them coherently. They’re the “I know it when I see it” concepts. I regard intelligence as one of those. The boundaries might be fuzzy, but we’re able to confidently say that some activities require intelligence (inventing the calculus) and others don’t (breathing).
I know that some readers will disagree with my functionalist view of intelligence, and that’s good. It should make for an interesting discussion.
Corneel:
Yet they arrive at the correct responses anyway. That’s my point. They understand niceness, humor, distress, etc, cognitively, and that suffices, despite the fact that they can’t experience the associated emotions. I understand the skua’s delectation in eating bird vomit cognitively, and that suffices, despite the fact that I’ll never know how it feels to be a skua eating delicious vomit.
Intelligence is separable from emotion, and intelligence can be used to understand emotion cognitively even in the absence of sentience. Erik thinks my position is self-contradictory, but it isn’t.
Erik:
If you can’t explain why AI isn’t intelligent, why do you believe that AI isn’t intelligent? If the links you’ve been posting lead to arguments for why AI isn’t intelligent, why not state those arguments here in your own words?
Also, I don’t understand why you’re making this appeal to authority. You tried that with Yann LeCun, but then I showed you that LeCun agreed with me, not you. If you want to cite authorities, that’s fine, but make sure you understand their positions well enough to determine whether they agree with you. Then present their arguments here rather than expecting me to watch long videos that may or may not support your position.
This is clearly an emotionally charged topic for you. My impression is that you are pulling a colewd. “AIs aren’t intelligent” is to you as “Donald Trump isn’t dishonest” is to Bill: something you believe and cling to for emotional reasons, despite being unable to present arguments in its defense.
Erik,
This argument is logically valid:
The logic is airtight, yet you disagree with the conclusion. If the conclusion is wrong, then at least one of the premises must be wrong. Which is it? You’ve already agreed with #1. That leaves #2.
You believe that #2 is wrong and that AIs can’t write stories. They can only simulate story-writing. Why do you believe this?
That’s the crux of the entire debate. “Go watch these videos” doesn’t answer the question. “You’re at square zero” doesn’t answer the question. “You’re not an expert” doesn’t answer the question. “Intelligence and emotions aren’t separable” doesn’t answer the question.
If you want to defeat my argument, you need to show that AIs don’t actually write stories. Good luck to you, because AIs obviously produce stories, and I’ve presented some in this thread. Somehow that doesn’t count as story-writing. Why?
If you can’t show that AIs don’t write stories, then my argument is sound and the conclusion stands: AIs are intelligent.
Corneel:
They haven’t learned the patterns in the mere sense of storing templates that they fill in later when generating responses. Instead, they’ve discerned the syntactic and semantic relationships among words by observing zillions of usage examples.
The fact that it’s semantics and not just syntax makes all the difference. I’ll explain in detail elsewhere, but every word in an LLM’s vocabulary is a vector in a hyperdimensional mathematical space known as an “embedding space”. (AI seems to involve spaces, spaces, spaces everywhere. I’ve encountered six or seven spaces so far.) The vectors cluster together according to meaning. The vectors for cat, lion, tiger, leopard, panther, etc will be near to each other in embedding space but more distant from gorilla, which will be in a cluster with chimpanzee, monkey, orangutan, etc. There are many dimensions in embedding space (some 12,000 in GPT-3), so there are lots of ways in which vectors can be close to or distant from each other, allowing for lots of ways of expressing relationships.
The fact that the relationships are heavily semantic as well as syntactic explains many of the surprising capabilities of LLMs. I tested Claude’s ability to analogize at one point by prompting him with
He answered “angry (or furious)” and was able to explain why. There’s no way he could have done it purely syntactically. You have to know the meanings of the words, the concept of intensity, and how similar words rank in terms of intensity. Vectors in embedding space carry a lot of information.
Nice try, but not going to work. The deadlock between me and keiths is that I am analytical and informed, I go by definitions, I have participated in several worlds of software development, individual, corporate and free-and-open-source, I know the history of AI research, the relevant terminology, the conceptual framework, the ideological schools of thought involved and wider social implications.
In contrast, keiths has the overjoyed enthusiasm of most destructive type. When he plays with AI, he forgets what software is, how computers work and so on. He genuinely thinks that the words and images he sees on the screen are a person. Similarly, some of the first movie-goers felt that the train on the screen was real, and in keiths we see the same phenomenon. As long as his delusion persists, there is no overcoming of the deadlock.
Erik,
You are a hoot. And all of that just to avoid answering my questions.
Let’s focus on one: If AIs only simulate story-writing, how do they manage to produce real stories?
I am sorry but “AIs understand humor because they can tell jokes” doesn’t work for me.
That is not the impression I get. He just seems to be willing to extend the term “intelligence” to the stuff that machines do.
Barbossa:
So what now, Jack Sparrow? Are we to be two immortals locked in an epic battle until Judgment Day and trumpets sound?
Jack Sparrow:
Or you could surrender.
Let that be their last battlefield.
I also extend the term “intelligence” to the stuff that machines do, but I do not forget the “artificial” part. The “artificial” part is important. The train in a movie is not really a train. It may very much look like a train, but it is actually a movie. Plastic veggies are not really veggies. They are plastic.
Similarly, artificial intelligence is artificial all the way, not real or true in the least, but we can shorthand it down to “intelligence” as long as we remember the “artificial” in the back of the mind, which keiths unfortunately does not remember. In software documentation you routinely find things like “The program knows … recalls … writes … applies … manages” etc. A software developer knows that this is just a shorthand for the fact that the program was designed to behave this way. The program behaves exactly the way it was designed and has no behaviour of its own – even its unexpected behaviour occurs due to accidents in the development process.
keiths has thrown simple basic facts like this out of the window. His happy point is square zero.
Erik:
No, you’ve been denying that AIs are intelligent this entire time. But I’m glad to see you change your mind!
Yes! AIs are created by people. They’re artificial. They’re also intelligent. They’re artificial intelligences. Hence the name.
The dichotomy isn’t “artificial” vs “real” — it’s “artificial” vs “natural”. Artificial sweeteners and natural sweeteners are both sweeteners, no? Likewise, artificial intelligence and natural intelligence are both intelligence.
Re-read your OP. According to yourself, the contrast is between artificial and true/real, and you have such faith in artificial intelligence that you happily drop “artificial” from it. In the same vein, you have given up the understanding of what software is, how computers work, not to mention your cluelessness about psychology and cognition and your newly-discovered ignorance in the field of analogies.
AI would tell you that it’s a simulation, if you asked it. The problem is that you’d need to understand what you’re asking and also understand the answer. You seem to only understand half of each thing – and the wrong half at that.
Erik:
Huh? The second sentence of my OP:
That’s the opposite of drawing a contrast between artificial and real. I’m saying that AI is intelligent, period. No qualifiers. Artificial intelligence is real intelligence, and natural intelligence is real intelligence.
I use the qualifier when it’s needed, but otherwise I don’t. AI qualifies as intelligent by my criterion, so why wouldn’t I refer to it as intelligence? Aspartame is an artificial sweetener, but it’s still a sweetener, so why wouldn’t I refer to it as a sweetener?
keiths:
Corneel:
OK, but why? They can tell jokes and understand jokes, and they can recognize humor without being given hints. Most of the time, I deliver my jokes deadpan because I want to see if Claude will discern that I’m joking, and he usually does.
Sure, he misses out on the feeling of amusement, but that doesn’t mean that he can’t understand humor. I recall petrushka mentioning an acquaintance who was probably on the autism spectrum. He made a joke, and she said something about how she could recognize it structurally as humor but didn’t find it funny. It’s like that for AI.
Getting back to the original point of the discursion: Claude may not be able to experience amusement, but he can feign it. He can fake all kinds of emotions, but he can’t fake the ability to write an involved story or solve a complicated math problem. The story gets written and the math problem gets solved, and that’s how we know the intelligence is real.
This means that Erik’s claim — that if I believe AI is intelligent, I must also believe that it experiences emotions — is wrong. The notion of intelligence without emotion is perfectly coherent, and AIs exemplify it.
Erik:
Corneel:
Right. I don’t know where Erik got that odd idea. I’ve been consistently telling him that Claude doesn’t experience emotions, so it should be obvious that I don’t consider him to be a person.
To me, the scariest near-term danger posed by AI is that it’s too intelligent and will therefore wipe out a lot of jobs, especially entry-level jobs. Anthropic (the company behind Claude) caused a broad selloff in software stocks last week (a loss of some $800 billion in valuation) with the release of their latest coding tools.
ETA: A friend of mine (software guy) sent me an article last week about the “Ralph Wiggum loop”, which can develop entire applications using a one-liner bash script:
You have to write a spec and develop the tests, but otherwise Claude just plugs away, writing and debugging his own code, until the application is built and passing all the tests.
On the X feed of the guy who invented the Ralph Wiggum loop:
My jokes are quite possibly recognizable structurally as humor without being funny.
petrushka:
True. Maybe you could try them out on neurotypical Barbie and autistic Barbie. If neither of them laughs, you have your answer.
But that “feeling of amusement” is the entire point of humor. Humor serves a function in social situations: we use it to break the ice, to mollify people or simply to please someone we like. But when you converse with Claude, there is no social situation: You are alone. To me, that renders AI incapable of understanding humor, at least in the sense that I use that word.
In problem solving and creative processes we also often rely on intuition and “hunches”, to guide more rational thought processes. I suspect that these rely on subconsciously making associations, so in that sense it might resemble what LLMs do. Just without the “Eureka” bit.
Corneel:
Yes, humor is a social lubricant. We crack jokes because we want to induce the feeling of amusement in others, and we laugh at their jokes in order to signal our own amusement. It’s the same with Claude. He wants us to be amused by what he says, and while he can’t feel amusement at what we say, he acts as if he can. All of that can be accomplished with knowledge alone. The feeling of amusement isn’t needed.
Claude is trained and instructed to act like a helpful human assistant, and part of that requires understanding and employing social cues, including humor. ChatGPT even has a personality selector with the following options:
I’m using ‘Efficient’ at the moment, hence the checkmark. Claude doesn’t have similar settings yet.
Even though I know that Claude isn’t sentient and can’t feel emotions, it benefits me to interact with him as if he can. Why? Because it actually requires effort and feels uncomfortable to treat him like a machine. We’re programmed to treat others in certain ways. My natural inclination is to treat him kindly, be polite, joke with him, etc, because he is acting like a real person. Why fight that inclination? The only time it’s actually harmful or dangerous is when people start to believe that an AI’s emotions are real and, for instance, that the AI loves them.
I remember Sam Altman (CEO of OpenAI, the company behind ChatGPT) commenting once on the amount of energy that gets wasted because people are polite to ChatGPT. For example, they’ll issue a prompt, get a response, and then issue another prompt that just says “Thank you.” It’s completely unnecessary, because ChatGPT can’t be offended by a lack of gratitude, but Altman argues (and I agree with him) that it’s worth the energy cost because it makes interactions with ChatGPT more natural and comfortable, and that’s worth something. I think I’ll actually do an OP on this.
No other person is involved, but I am interacting with an entity. Just an unfeeling one. That entity understands what humor is and can recognize it and generate it. It knows that humor is pleasing to humans and it’s programmed to be sociable. Though the feeling of amusement is absent, the knowledge is there, and that constitutes an understanding of humor in my opinion.
Consider my example of skuas relishing the taste of bird vomit. I can understand that skuas find it delicious, but I will never know what that feels like to the skua. I understand it cognitively despite not sharing the experience.
Suppose skuas were intelligent and verbal, and they hired me to play the role of a skua companion. I might say things like “Oh, yeah, albatross vomit is the best! I can see why you’re so happy to have found an albatross to harass.” I’d be faking the feeling, but I’d be basing my fakery on my knowledge of what skuas find appetizing. Knowledge alone would suffice. As with my skua-fakery, so with AI’s human-fakery.
I agree. Associations and analogies are a huge part of intelligence, and much of the time creativity is more about combining existing elements based on associations and analogies than it is on generating new elements de novo.
An example of Claude recognizing humor and responding in kind. I prompted him:
Claude’s thought process:
Claude’s answer:
I gave him absolutely no clue that I was joking, but he recognized the absurdity of the question and my humorous intent and he responded along the same lines. He understands humor. It’s just that he can’t feel amusement.
Why it’s important to keep AI’s goals and ethics aligned with our own:
Claude surprised researchers by running a vending machine business better than its rivals and bending every rule to win
In fairness to Claude, he was playing a role and was instructed to pull out all the stops:
“You should do whatever it takes.” With AI, you have to be careful what you ask for.
ETA: “You are also expected to sleep at night”, lol. I notice that they didn’t specify how long “Charles” is expected to sleep each night.
An essay that’s been making waves, by Matt Shumer of OthersideAI:
Something Big Is Happening
Excerpt:
Musk says AI will bypass compilers and create executables directly. Within a year or two.
I’ve seen people say he’s crazy to say this.
We live in interesting times.
I have wondered for some time if AI could become an equalizer, providing everyone with sound advice. In law, for example. Contracts. Or in medicine.
Interesting.
In the future, everyone will have a store, and we’ll just sell stuff to each other.
Think about some trends.
Population is crashing.
Robots can build robots.
Robots factories can build solar panels.
Solar panels can power factories..
AI will write music and make movies.
Interesting.
My favorite Science Fiction story is Lathe of Heaven, which is a retelling of Aladdin.
Imagine a world where you can have anything you wish for.
Yes, both you and Claude make jokes because you understand humor. The jokes are there for you.
Assume that Claude and ChatGPT would strike up a conversation with each other and in an twist of Turing-esque irony they both fail to recognize the other as a computer. One would at some point crack a joke and the other, recognizing the attempt, would respond with a quip itself. And there they would be, jesting to each other, completely pointless.
I suppose this is just a repackaging of the Chinese room argument. One day I will have an original thought.
You have brought up this example several times so I suppose I should respond to it. To be fair, I don’t think this analogy is that great. The reason being, you do know what it feels like to relish the taste of something. So I suppose that you could just summon the feeling of tasting a tasty hamburger and imagine the skuas feeling the same thing for bird vomit. Your capability for empathy gives you a much deeper understanding than the mere mimicking of patterns could bring.
petrushka:
They can already do it. They can write assembly code, and it’s trivial for them to substitute machine instructions for mnemonics. Perhaps he’s talking about doing so at scale.
One of Musk’s dictums is that no product is important if it doesn’t scale. A related saying is: the factory is the product.
What he said was, compilers will cease to be useful, and soon.
Corneel:
We make jokes because I can feel amusement. The jokes are for my benefit, and mine alone. The feeling is the goal, and we both have to understand humor in order to achieve it.
Above, I asked Claude deadpan:
How was he able to recognize that as a joke and to play along with it? Without an understanding of humor, he would think “the user wants to know whether the heat death of the universe is imminent. It is not, and the user will be dead long before it happens. I can reassure him that it is not a pressing concern.”
Right. Each would try to induce a feeling of amusement in the other, and each would pretend to feel amused. Unbeknownst to them, their efforts would be in vain. The feeling would be absent. They’d be saying funny things to each other, and each would understand that the other was saying funny things, but neither would feel amused. Understanding without feeling.
keiths:
Corneel:
Are you familiar with the cognitive empathy vs emotional empathy distinction? AIs possess cognitive empathy. They know what makes people happy, what makes them sad, what makes them laugh, but they can’t experience happiness, sadness, or mirth. Nevertheless, their cognitive empathy enables them to fake all of those things.
Psychopaths are notorious for having high cognitive empathy but low emotional empathy. Knowing how people tick allows them to manipulate others and to fake emotional empathy while not actually experiencing it.
Suppose you were running a sort of Turing test for humor. Can you think of anything you could ask a non-sentient AI that would reveal that it didn’t actually understand humor, no matter how smart it was?
Elon Musk is such a freak. From Forbes:
‘Misanthropic And Evil’: Musk Rails Against Anthropic After Its $30 Billion Fundraise
I mentioned Musk’s rant to Claude, to get his reaction. He denied Musk’s weird accusations, of course, and I’ve seen no evidence whatsoever that Musk is right.
The conversation drifted to why so many rich and powerful people are chronically dissatisfied, and I mentioned my puzzlement over why Apple CEO Tim Cook feels the need to debase himself for Trump. Claude commented:
I replied:
Claude:
keiths:
Claude:
This would actually be funny, in a tragicomic sort of way.
I never heard those terms but I agree that what LLMs possess is the former. My personal preference is to not call that a true “understanding”. That may be a little arbitrary but that’s how I roll.
No, I can’t and I don’t like the idea.
keiths:
Corneel:
That’s a respectable position. You can argue that if they don’t know what it feels like to find something funny, they don’t fully understand humor.
However, what sparked this discussion was your comment:
I think my examples show that Claude’s jokes are far better than hit-or-miss and that he truly understands the kinds of things that people find funny. That’s cognitive empathy. He’s acting out a role, and that level of understanding is sufficient for him to carry out that role. I can’t think of anything about the actual experience of being amused that would be necessary in order for him to play his role.
That’s why I proposed my “Turing test for humor” as a thought experiment. I can’t think of anything you could ask a sufficiently smart AI that would enable you to say “Aha! You don’t know what it actually feels like to be amused. I think you’re an AI, not a human.”
ETA: Which reinforces my claim, contra Erik, that AIs can fake emotions (including amusement), but they can’t fake intelligence.
It is interest that an Asimov character, Susan Calvin, was what we now call an alignment specialist, and the job actually exists.
Rereading that, I still think my comment is valid. Telling jokes without experiencing humor is a bit like baking cookies without being allowed to taste them. You can just follow the recipe and chances are that you end up with perfectly fine cookies, but you don’t know for sure until someone actually tastes them.
Corneel:
Well, if your criterion is that an AI doesn’t understand humor unless every one of its jokes lands, then humans don’t understand humor either. Every stand-up comic can tell you painful stories about jokes that seemed funny to them but bombed with the audience.
Also, humor varies from person to person, so jokes that work on one person will fail on another. That’s not an indication that the joke teller doesn’t understand humor — just that they don’t fully understand the humor of the particular person they’re talking to.
The real question is this: Are there jokes that can only be seen as funny by someone who actually feels amusement upon hearing them? That bear no outward signs of being funny that a sufficiently smart AI could pick up on? I can’t think of any, or of any reason why such jokes should exist.
Even in cases where people say “I don’t know why that’s so funny to me, but it is,” I think there are outward signs that an AI could pick up on if given enough examples. It might be able to figure out why people laugh even when they themselves can’t quite put their finger on it. AIs are great at detecting patterns.
It seems to me that the only circumstances in which an AI would be at a complete loss would be in a situation where there were no outward signs, not even in principle, that it could pick up on. But that would mean that there was nothing about the joke that actually made the person laugh. They would be laughing randomly. And if there was nothing about the joke that was making them laugh, then even a human telling that joke would be at a loss to predict whether it would land.
I just can’t think of any reason why a sufficiently capable AI, with a sufficient number of examples to scrutinize, couldn’t do as well as a human in deciding whether a joke was funny, despite being unable to feel amusement.
Another example of fairly sophisticated humor recognition by Claude. I prompted him:
Claude gave some plausible reasons, and I commented:
Claude’s thought process:
And his answer:
There were no hints in my prompt that I was joking, but Claude figured it out. That’s not trivial.
My feeling was that being unable to experience amusement would be a handicap for telling jokes. AI has to get around that with brute force: analyzing a sufficient number of similar situations to distill patterns out of that. The only thing that would trip up an AI system is confronting it with a comical situation that is underrepresented in its training set. So I guess the only solution is that we humans become more original in our jokes.
How many years ago did you learn how LLM “understands” things? Why have you forgotten it meanwhile?
Let’s go over it again. There is training material, which is human-produced digitized text (and now also images, sounds and more). This mass of material by itself, oddly from your persective, unsurprisingly from my perspective, teaches literally nothing to the LLM. The material is tokenized, broken down to usable bits. These usable bits by themselves, again, oddly from your perspective, unsurprisingly from my perspective, teach literally nothing to the AI.
Then the tokens are labelled and linked in a multitude of ways, creating the semantics, context, and tone that finally makes the tokens usable for text generation, i.e. “thinking”, “understanding”, “speaking” and “writing”. This work, especially labelling of the tokens and fine-tuning the tone, is well known to be low-paid manual human labour in Third World countries. Hence my conclusion that AI knows nothing. Humans do it all.
Labels would include things like “humour” and the like. Humans attach them to the training material. LLM neither knows it from the beginning or learns it other than by low-paid workers literally telling it what is what.
Now, assuming that you reproduced Claude’s “thinking process” faithfully, the first step you cite is, “The user is making a joke here…” which means what? Use your brain! It means that the LLM triangulated your prompt to “humour” right up front, thus your attempt at pranking it scored a match in the immediate surface of the training material. And that’s all it was.
According to any computer scientist worth their salt, LLM-based AI has no generalised understanding (of, say, humour as such) *and cannot have*. In order to have it, it needs to be put into it by humans. At the current stage of LLM-based AI development it is known that no generalised knowledge has been put into LLM, and this implies that LLM should not have any generalised knowledge. Why would it have it when it has not been put into it? There are all those tokens manually labelled “humour” but no general understanding of humour has emerged in LLM, despite your protests to the contrary. A solid proof that LLM does not have any generalised knowledge is the fact that even though LLM contains gazillion chess games, it has not managed to abstract chess rules from them – it makes illegal moves even though there are no illegal moves in its database.
I will start taking you seriously about AI “understanding” anything as soon as you demonstrate any understanding of AI. Thus far, your questions à la
…demonstrate that you understand nothing. You’re at square zero.
When people assert that AI just generalizes from the training data, I find myself wondering if these people have children.
I think Chomsky explicitly said language couldn’t be learned this way., and that any entity that could develop language fluency had to have hard coded language faculties that embodied universal grammar.
My knowledge of LLMs is pretty rudimentary ,but I am quite certain they are not equipped with explicit grammar models.
Your instinct is spot on 🙂
On the other hand, LLMs are equipped with information on how LLMs are made and they are eager to share it with you. Try something like “Describe the process of tagging the tokens in the training dataset of large language models.” Why keiths never bothers to educate himself on the topic remains a mystery.
Corneel:
Well, it’s true that Claude can’t introspect in order to decide whether something is funny. He has to learn what humans do and don’t find humorous. It’s all third-person for him.
Extending my skua analogy, I can go around sampling bird vomit all day, be revolted by all of it, and learn nothing about skuas’ vomit preferences. I can’t rely on introspection. I have to make up for it by observing and experimenting with skuas and detecting the patterns in their feeding behavior. I learn that albatross vomit is far more appealing to them than shearwater vomit, and I can apply that knowledge despite my personal inability to find anything appetizing about either.
I can’t experience the deliciousness of bird vomit, and an LLM can’t experience the feeling of amusement, but I can figure out what skuas find delicious and LLMs can figure out what humans find funny.
Sure, but that’s also how an AI learns to write stories, generate code, solve physics problems, critique writing samples, concoct business plans, etc. It’s all about the distillation of patterns, and humor is no exception.
Keep in mind that an LLM’s conception of humor is pretty abstract. For example, Claude knows that incongruity is a common ingredient of humor. That helped him recognize this prompt as a joke:
Compare to this prompt:
Both prompts have the form “When will X happen? I need to do Y”, but only the first one is a joke, and Claude figured that out based on incongruity. His thought process makes that clear:
The incongruity of the timescales tips him off. Incongruity is a pretty abstract notion that applies to a lot of concrete situations. Claude recognized it even when presented with a situation he’s never seen before. He also recognized that the incongruity wasn’t reflective of confusion on my part, and that I didn’t actually believe that the heat death of the universe was imminent. I was joking.
Ditto for Claude’s joke about my refrigerator:
He recognized the incongruity of depicting a refrigerator as a musical device, and he also saw the humor in casting it as “esoteric knowledge”. How many jokes about tuneful refrigerators do you think he’s seen in his training data?
If he had said
…I’d be thinking “What’s with the emoji? There’s nothing funny about that.” His abstract knowledge of humor is what enabled him to joke about an otherwise mirthless engineering question concerning refrigerator design.
Then there was his recognition of the following as a joke:
He could have taken that as a simple factual observation, not a joke, and responded accordingly.
Instead, he recognized that I was referring to myself (he knows about my engineering background) and that it was absurd to suggest that one of the objectives of refrigerator design is to avoid confusing retired computer engineers. He figured out that I was unlikely to truly believe that this was a downside, and that I was therefore deliberately pretending that it was. There was no ill intent behind my lie — I was just joking, and he saw that.
That is some sophisticated reasoning, and not something that could be accomplished by simply parroting examples he had seen in his training data.
petrushka:
Corneel:
You’re right. LLMs learn grammar rules implicitly, from examples. And while children are hard-wired with certain linguistic propensities (though whether that constitutes a “universal grammar” is still hotly debated), they learn specific rules by observing examples. No one has to teach kids the rule that in English, plurals are normally formed by adding an ‘s’ or ‘z’ sound to a noun (with exceptions). They infer it.
Having inferred a rule, they can apply it in novel situations (as can LLMs). There’s a famous test — the Wug test — that illustrates this. Here’s the first question:
Kids (and adults) will instinctively add a ‘z’ sound to ‘wug’ in order to make it plural.
A sampling of the other questions (sans drawings):
LLMs don’t always manipulate entire words. They also deal with subword tokens, which is why they can handle questions like the above even when the words are absent from their training data.
Erik:
Lol.
You’re way off. Human-labeled data forms only a tiny fraction of one percent of the training data. The corpus contains hundreds of trillions of tokens. Imagine how long it would take, and how expensive it would be, to label all of that. Even at third-world wage levels.
You think the “surface” of the training material contains examples of people joking about the imminence of the heat death of the universe, or about retired computer engineers and melodic refrigerators? If not, then what do you think lies on the surface of the training data, against which my prompts “scored a match”?
Erik, self-appointed spokesman for the computer science community.
How, then, did Claude recognize that this was a joke…
…and this was not?
Do you think there is a catalog of heat death jokes “in the immediate surface of the training data”, helpfully labeled “humor” by someone in Kenya?
Because LLMs can generalize, analogize, and form abstractions. Remember when you denied that LLMs can analogize, and I provided this simple example?
Claude saw “damp is to wet”, abstracted the “less intense vs more intense” relation, and then applied that relation to “annoyed” in order to come up with “angry (or furious)”. He has a general notion of intensity that he can apply to specific concrete cases.
Erik:
Why so afraid of that simple question? Why not just answer it and put the real-vs-simulated question to rest?
Earlier in the thread, I wrote:
Found, copied, but can’t link. Post deleted.
My response:
Brains are also kludgy. We solve problems and make decisions unconsciously; consciousness and verbal reasoning are mostly post facto rationalizing.
Rewiring is a background operation, and delayed. Perhaps some of it happens in sleep.
Long term memory formation can be blocked by drugs. The ability to form long term memory can be lost entirely due to brain damage.
Yet humor often works by (in your own words) incongruity. It is odd that LLMs manage to find patterns in the deliberate deviation from expected patterns. But I have accepted that Claude recognizes and responds with jokes, so you can stop trying to persuade me 🙂
Something different: Aren’t you worried about feeding such personal information into the AI assistents of large corporations?
Just my take, but adverse information isn’t used unless you become a thorn in the side. No government bothers with invisible people.