I had dinner a few nights ago with Andrew’s former postdoc Aleks Jakulin, who left the green fields of academia for entrepreneurship ages ago. Aleks was telling me he was impressed by the new LLMs, but then asserted that they’re clearly not intelligent. This reminded me of the old saw in AI that “AI is whatever a machine can’t do.”
In the end, the definition of “intelligent” is a matter of semantics. Semantics is defined by conventional usage, not by fiat (the exception seems to be an astronomical organization trying to change the definition of “planet” to make it more astronomically precise). We do this all the time. If you think about what “water” means, it’s incredibly vague. In the simplest case, how many minerals can it contain before we call it “mud” rather than “water”? Does it even have to be made of H20 if we can find a clear liquid on an alternative earth that will nourish us in the same way (this is a common example in philosophy from Hilary Putnam, I believe)? When the word “water” was first introduced into English, let’s just say that our understanding of chemistry was less developed than it is now. The word “intelligent” is no different. We’ve been using the term since before computers, and now we have to rethink what it means. By convention, we could decide as a group of language users to define “intelligent” however we want. Usually such decisions are guided by pragmatic considerations (or at least I’d like to think so—this is the standard position of pragmatist philosophers of language, like Richard Rorty). For instance, we could decide to exclude GPT because (a) it’s not embodied in the same way as a person, (b) it doesn’t have long-term memory, (c) it runs on silicon rather than cells, etc.
It would be convenient for benchmarking if we could fix a definition of “intelligence” to work with. What we do instead is just keep moving the bar on what counts as “intelligent.” I doubt people 50 years ago (1974) would have said you can play chess without being intelligent. But as soon as Deep Blue beat the human chess champion, everyone changed their tune and the chorus became “chess is just a game” and “it’s finite” and “it has well defined rules, unlike real life.” Then when IBM’s Watson trounced the world champion at Jeopardy!, a language based game, it was dismissed as a parlor trick. Obviously because a machine can play Jeopardy!, the reasoning went, it doesn’t require intelligence.
Here’s the first hit on Google I found searching for something like [what machines can’t do]. This one’s in a popular magazine, not the scientific literature. It’s the usual piece in the genre of “ML is amazing, but it’s not intelligent because it can’t do X”.
- Rob Toews. 2021. What artificial intelligence still can’t do. Forbes.
Let’s go over Toews’s list of AI’s failures circa 2021 (these are direct quotes).
- Use “common sense.” A man went to a restaurant. He ordered a steak. He left a big tip. If asked what the man ate in this scenario, a human would have no problem giving the correct answer—a steak. Yet today’s most advanced artificial intelligence struggles with prompts like this.
- Learn continuously and adapt on the fly. Today, the typical AI development process is divided into two distinct phases: training and deployment.
- Understand cause and effect. Today’s machine learning is at its core a correlative tool. It excels at identifying subtle patterns and associations in data. But when it comes to understanding the causal mechanisms—the real-world dynamics—that underlie those patterns, today’s AI is at a loss.
- “Reason ethically…In 2016, Microsoft debuted an AI personality on Twitter named Tay. The idea was for Tay to engage in online conversations with Twitter users as a fun, interactive demonstration of Microsoft’s NLP technology. It did not go well. Within hours, Internet trolls had gotten Tay to tweet a wide range of offensive messages: for instance, “Hitler was right” and “I hate feminists and they should all die and burn in hell.”
(1) ChatGPT-4 gets these common-sense problems mostly right. But it’s not logic. The man may have ordered a steak, gotten it, sent it back, ordered the fish instead, and still left a big tip. This is a problem with a lot of the questions posed to GPT about whether X follows from Y. It’s not a sound inference, just the most likely thing to happen, or as we used to say, the “default.” Older AIs were typically designed around sound inference and weren’t so much trying to emulate human imprecision (having said that, my grad school admissions essay was about and my postdoc was funded by a grant on default logics back in the 1980s!).
(2) You can do in-context learning with ChatGPT, but it doesn’t retain anything long term without retraining/fine tuning. It will certainly adapt to its task/listener on the fly throughout a conversation (arguably the current systems like ChatGPT adapt to their interlocuter too much—it’s what they were trained to do via reinforcement learning). Long-term memory is perhaps the biggest technical challenge to overcome, and it’s been interesting to see people going back to LSTM/recursive NN ideas (transformers, the neural net architecture underlying ChatGPT, were introduced in a paper titled “Attention is all you need”, which used long, but finite memory).
(3) ChatGPT 4 is pretty bad at causal inference. But it’s probably above the bar for what Toews’s complaints. It’ll get simple “causal inference” right the same way people do. In general, humans are pretty bad at causal inference. We are way too prone to jump to causal conclusions based on insufficient evidence. Do we classify baseball announcers as not intelligent when they talk about how a player struggles with high pressure situations after N = 10 plate appearances in the playoffs? We’re also pretty bad at reasoning about things that go against our preconceptions. Do we think Fisher was not intelligent because he argued that smoking didn’t cause cancer? Do we think all the anthropogenic global warming deniers are not intelligent? Maybe they’re right and it’s just a coincidence that temps have gone up coinciding with industrialization and carbon emissions. Seems like a highly suspicious coincidence, but causation is really hard when you can’t do randomized controlled trials (and even then it’s not so easy because of all the possible mediation).
(4) How you call this one depends on whether you think the front-line fine-tuning of ChatGPT made a reasonably helpful/harmless/truthful bot or not and whether the “ethics” it was trained with are yours. You can certainly jailbreak even ChatGPT-4 to send it spiraling into hate land or fantasy land. You can jailbreak some of my family in the same way, but I wouldn’t go so far as to say they weren’t intelligent. You can find lots of folks who think ChatGPT is too “woke”. This is a running theme on the GPT subreddit. It’s also a running theme among anti-woke billionaires, as reflected in the UK’s Daily Telegraph article title, “ChatGPT may be the next big thing, but it’s a biased woke robot.”
I’ve heard a lot of people say their dog is more intelligent than ChatGPT. I suppose they would argue for a version of intelligence that doesn’t require (1) or (4) and is very tolerant of poor performance in (2) and (3).
I was watching a “classic” football game on the NFL channel, and blog-surfing during the commercials, when I saw this post. Had to turn off the game to focus on this since I found it absorbing.
The key point for me is to focus on the actual foibles of human brains when comparing them to various AI programs. I think most if not all the people who communicate here are a couple standard deviations above the average IQ of 100. (Not that IQ is a great, determinative measurement, but for lack of a better one.) There is not a general AI which can emulate them, but that does not mean that no progress has been made.
Funny that you mention IQ.
I was writing this comment as you posted yours.
On a more mundane level.
I see a lot of arguments that “intelligence” is a clear and coherent concept, basically because it can be measured by IQ tests. It’s simply a fact, some people say, that some people are more intelligent than others. IQ tests predict how well you’ll do in life. Others say there are clear differences in intelligence among different groups, because intelligence is a clear and coherent attribute that is largely (or I guess some say completely) an inherited attribute. It’s a function of your generetic makeup.
So how well does ChatGPT do on IQ tests? Presumably it would get a result above zero! So it must be intelligent by that logic. I just did a Google that says ChatGPT’s gets 155 on an IQ test, better than 99.9% of (human) test takers. So then I guess anyone who argues that IQ validly measure intelligence, to be logically consistent would agree that AI is not only intelligent but is in fact basically Einstein intelligent.
To use IQ to compare between AIs and humans, you want to first demonstrate measurement invariance.
The brilliant Cremeiux has an article: https://www.cremieux.xyz/p/nonhuman-intelligence
So if IQs of LLMs don’t mean much, does that have implications for how much meaning is there to IQ of humans?
That’s a dismal argument (not yours but the “arguments” that “you see a lot of”). What “can be measured by IQ tests” is “intelligence” only if one asserts that “intelligence” is “what can be measured by IQ tests”.
Those arguments become somewhat self-serving since (like “arguments” around the existence of free will) their subject (“intelligence”/”free will”) is mechanistically underdetermined and so pretty much all “arguments” tend to become contrived in support of some sort of a “position” and susceptible to semantic confusion.
I distinctly remember asking Andrew at one point what “happiness” meant in a survey and he replied something like, “whatever a survey respondent replies.” This is a very nuanced take on semantics, which admits that we don’t know what happiness means and can’t convey it precisely to a survey respondent, but we’ve given everyone some context, so let’s see what they say. A more formal way to do this is with a multiple rater model like cultural consensus theory.
Yes, I remember that conversation! Or, at least, I think I do.
Intelligence in the psychometric sense and the ordinary sense are similar. People who have a really IQ tend to be what ordinary people say when they use the word “intelligence.” But ordinary language does not perfectly mirror technical language. That’s why some scholars like Arthur Jensen elected to use a technical term like “cognitive ability” or “mental ability.”
General cognitive ability is whatever explains the positive manifold. When you give people a diverse test of items that require thinking, recalling, pattern recognition, and reasoning they produce tests in which all items positively correlate at least somewhat. Most people don’t have a particular objection to the idea of a cognitively demanding test item, so I think this is a fine description. General cognitive ability seems to be “real” in the same sense that consciousness is “real” but we don’t know the underlying mechanism. That’s okay. We can understand aspects of intelligence and conscientiousness without exactly understanding what is going on. We do know that there are certain biological correlates which can clue us into how general cognitive ability works.
Just a reminder that Turing said this in his 1950 paper. I’m paraphrasing a lot and fast forwarding to the present context, but he then went on to say that when the machine’s performance resonates with the hearer as intelligent (in a sustained way), then we can say the machine is intelligent.
The semantic debate then regresses to defining ‘resonance’ (not to mention the fact it is hard to define what a machine is, and elude embedding the answer in the definition).
I like this definition and for me the most important thing that machines cannot do is be held responsible for their errors.
What does it mean to be “held responsible”? I mean that seriously as a question.
Does it mean that someone being held responsible loses something of material value? That they lose some kind of reputational value? That they are made to at least partially compensate someone for the error? What?
I think it’s entirely possible to set up a company that does stuff via chatbots, and then when they make an error, they have to compensate someone and lose reputation. So if the company is essentially a few owners and a chatbot, then yeah, kinda the chatbot can be held responsible… You might argue that it’s the company and owners that are held responsible and I wouldn’t really argue that though.
I think though if there is a chance of losing some material value because of errors made by chatbots, no-one in their right mind would deploy chatbots into that kind of scenario at the moment. It’s just trivial to make chatbots say unhinged things that are completely devoid of logic, facts, completely divorced from reality.
For example, recently I saw someone posting a GPT summary of the landlord interactions in the movie The Big Lebowski
https://www.reddit.com/r/lebowski/comments/1c2tun1/i_think_i_saw_a_different_movie/
Anyone who has seen the film knows this is complete made up horseshit.
If Chat GPT can’t even get correct some basic plot points of a public movie that is one of the most widely discussed and shared movies on the internet, it seems unlikely to get anything of higher value right.
It just doesn’t have any representation of facts. It’s only connected to the world tenuously.
For the nonce, I remain of the opinion that it’s always the _user_ of the technology that’s responsible.
So, for example, I don’t get the moral panic over drones. If you use a drone to kill someone, it’s you, not the drone, that done it.
In the long term, I don’t see any _philosophical_, _logical_, _theoretical_ or other reason that we couldn’t build an AI just as smart as we are. The human brain is either (equivalent to) a computer, or it’s magic. Since there’s no such thing as magic, we’re all computers. But even then, theoretically at least, we’re still back in the HAL situation of being able to turn off the power. (Theoretically at least because if Google comes up with a real AI and it starts acting up, they won’t be willing to let you power down their server farm. Oops.)
Of course, I also believe that human intelligence is way better and way harder than people in AI think, so it’s a long way off.
But, you wrote: “It just doesn’t have any representation of facts. It’s only connected to the world tenuously.”
Careful there: the next thing you know, you’ll be calling it a parlor trick.
We can think of this mulitple ways. When a user uses Dall-E to generate a copyright-infringing image (or uses GPT to generate copyright-infringing text), who is responsible? Is it the company, Open-AI? That’s what the people behind the lawsuits are claiming. Or is it the person using the technology in an infringing way? That was the outcome of the betamax suit when this came up with videos. We didn’t hold Sony accountable for people using their machines to pirate. (In contrast, Napster, which was being used to distribute pirated music, was ruled illegal because, apparently (and I saw a one-hour lecture on this—I’m no legal expert), there weren’t non-infringing uses of the technology.
As suggested, you can turn off the machine but it does not suffer. So, I don’t see any meaningful way in which the AI can be held responsible. But humans can (and should) be. Whether it is singular individuals or groups isn’t something I’ve thought about – and, since we consider corporations “individuals” under the law, there are thorny questions I haven’t thought through. But, for once, I agree with David that it must be the humans that are held responsible in some fashion. The problem is that with an AI it is very difficult to figure out what humans to hold responsible – is it all that contributed to the algorithms, the final person that implemented the technology, the company for which it was employed, etc.? And, responsibility in the legal sense may not be the same as responsibility in the ethical sense. So, I think these issues are complicated and our collective understanding and agreements have not come close to catching up with the technology.
I am reminded about the “guns don’t kill people, people do.” Similar issues arise there but I think the situation is far simpler – and yet we still have trouble grappling with them. But for Andrew’s original post, I think it is futile to articulate what humans can do compared with what humans can do – the rapid changes in technology and our lack of understanding of how the human brain works combine to make that a black hole. However, it seems clear to me that there is no way to meaningfully hold a machine responsible and that, at least, puts the focus back on which humans are responsible for which deeds. I think this is a question we need better answers for. Under current understandings, I expect to see familiar warning labels attached to any AI services, warning users that they are responsible for any mistakes/consequences that occur through its use. Good luck with that.
This was my post, not Andrew’s. Maybe I should be making this more clear! See my earlier response about where this is at in the legal system. It seems not unlikely that it will be the users held responsible for infringing uses.
Is that because the idea of ‘being held responsible’ relies on a notion of consciousness or awareness that it does not seem reasonable to extend to (current) AI?
For example, it might be that after some serious incident caused by an AI error, a court ordered the machine to be switched off, software to be deleted, etc. Were the AI a human being, we would certainly say that it was personally being held responsible for its error; in contrast, switching off the AI would be seen not as a negative consequence for the AI itself, because the AI is unaware of what is happening to it, but for associated human beings (e.g. loss of shareholder value). So until an AI can suffer negative consequences, it cannot “be held responsible”.
To relate this back to the post: in addition, if we did ascribe this kind of awareness to an AI, would its high performance on an IQ test suddenly indicate that it was “intelligent”, rather than merely a sophisticated “pattern matcher”?
I think you leave out in the ethical assessment that the AIs apparently have a fairly casual relationship with the truth.
It wasn’t me. I just grabbed a random article of the “AI isn’t here because it can’t do X” variety. You can grab another article and get another set of issues. My point was just that they’re always things the current AI can’t do, but when we build an AI that can do them, the bar moves. That is, I don’t think the author’s going to be convinced when ChatGPT can do all four things he mentions.
A better understanding of what we mean by intelligence with regard to organisms would certainly help to better inform what benchmarks we should strive to improve with respect to silicon. Creativite problem outside the realm of the training data seems like it should be a requirement. GPT is remarkably good at solving problems likely to have been included in its training data. Creativity basically requires making decisions outside the training data, and it seems an open question to ask where outside the training data we must go to be “creative”: Namely, are creative tasks an interpolation of the training data such that the interpolation is accessible with nonzero mass, or are creative tasks an extrapolation of whatever embedding space the models already understand?
A separate question is how machines learn. In-context learning must have limitations that any organism is unbound by. A truly intelligent system would need to be capable of something approximating online learning from new experiences. Children display this better than adults by constantly replaying something new to explore the space and learn something new. Whether that is manipulating objects to practice fine motor control or repeating a word incessantly to master the rhythms of natural speech, children (and adults to a lesser extent) learn through practice in a way that machines cannot replicate. I suspect other species learn similarly, although I personally have paid less attention to other species than I have to younger relatives and my own children. A system we could agree is intelligent would understand when to replicate this behavior to train on new tasks. I’m unaware of any research today which has led to machines autonomously retraining to learn new tasks. However, the implications are obviously huge for developing systems sensitive to distribution shift, such as recommender systems and robotics.
What gets you the singularity is when an AI system is intelligent enough to design noticeably more intelligent AI systems than itself.
Does it? Would not the less super smart AI avoid unleashing a beast it cannot control?
I like the conclusion of this (https://cepr.org/voxeu/columns/ai-and-paperclip-problem) blog post. It says that right now, we should be more afraid of humans using AI for ‘evil purposes’ than of AI draining the world’s resources to produce paper clips. I tend agree (emphasis on the ‘right now’ aspect). In this context, I appreciate the ongoing public debate on the prevention of deep fakes, which is a very real threat to public peace and not at all hypothetical.
If a lesser AI is programmed to follow a certain action, then it is going to do it. It doesn’t think in a similar way as you that “I should not unleash a beast I cannot control.” That’s why it’s very important to ensure that it’s well aligned with the interests of humans before making large leaps in capability.
ChatGPT 4.0 can make pictures. Try to get it to make a picture of a man riding a horse backwards, see what happens. Try real hard.
I asked the ‘AI’ of a search engine to find pictures of men riding a horse backwards. It succeeded (at least the first two entries were spot on). So I guess it depends on the AI you use. Some AI cannot do certain things. But another AI might be able to do it.
Yup. Very hard to get that going with Dall-E. Closest I could get after about ten prompts was two people on the horse, with one sitting backward. I’d say that’s more of a limitation of Dall-E than of ChatGPT. The diffusion models connection to language is very local and tag-based. Even I wouldn’t go so far as to say something like Dall-E is intelligent. But it does bring up interesting points because of what feels like creativity when you give it some sketchy prompts (I used StableDiffusion to generate 30 illustrations for a D&D module I used ChatGPT to write all the text for [lots of subprompting, etc.—it can’t just do this all in one go]).
Btw I’m not claiming the ability to make a picture of a man riding a horse backwards is a necessary condition for ‘intelligence’, I just find it interesting that chatGPT4 can make pictures of all kinds of crazy things, but not that.
I think it effectively has a strong prior distribution of what “a man on a horse” means, and additional instructions just can’t overcome it.
Previously it had a similar issue with a modified Monty Hall problem in which the doors are transparent: it would still say you should always switch doors. And the “man and his son in a car accident”, if you changed the man to a woman it would give a funny answer. It now does the modified Monty Hall problem fine, haven’t tried the other.
It really is the Chinese Room problem in real life. Amazing how prescient that thought experiment was.
The stupid thing about the Chinese Room (as I understand it) is that Searle postulated that a room couldn’t understand Chinese, **even if it never made any of these mistakes**. Like if you had a switch and you could switch back and forth between talking to the Chinese Room and a Chinese guy on a teletype terminal and you had zero way to figure out who was who, Searle still said basically the chinese room doesn’t understand chinese.
Whereas with Chat GPT et al it’s precisely the ability to trip it up into nonsense that would be obvious to a human which makes it clearly not “understand”.
My own view is if you could actually have a Chat GPT ish system that never told you obvious nonsense then it would *indeed* understand.
My understanding of the Chinese room problem is different.
Imagine that instead of the computer program (as the Chinese Room is described in the Stanford encyclopedia) there is simply a massive (infinite?) book with correct responses to all possible inputs, and the machine operator enters the asker’s input and gives the machine output back to the asker. In 1956, John McCarthy and Claude Shannon suggested (in a preface) that such a system would pass the Turing test, but in fact the machine operator has no idea what the inputs or outputs mean.
There is more to those discussions, but for now, it makes no difference whether the answer is generated by a finite computer program or is found in the pages of a (possibly) infinite book – the operator has no idea what is being shared by the parties. The setup is an elaborate database lookup.
Back in the finite world, increasingly large books of this type would dazzle increasingly large crowds that would less and less frequently give ridiculous answers. Does this mean that the setup is becoming more intelligent and less of a database lookup? I don’t think so.
My knowledge of the Chinese Room argument is pretty much at Wikipedia level, which is here: https://en.wikipedia.org/wiki/Chinese_room
I’d say it more or less conforms to Daniel’s description, but not necessarily Eric’s…although the difference between Searle’s formulation and Eric’s might not be important really. In Searle’s thought experiment, there’s a ‘program’ (potentially encoded in many many books). Searle sits in the room, accepts an input of characters, through a slot in the door, and then follows the procedure in the books to generate an output. There’s no book or set of books that contains every set of possible inputs, or at least that is not the only way to do it (as Large Language Models demonstrate).
My memory of reading the Chinese Room article at the time was that Searle didn’t understand that Chinese was just another language and that his argument depended on his readers not understanding that either. (Chinese is a seriously fun/fascinating/ridiculously hard langaguage, but it’s just a language.)
Long story short was that he didn’t believe that language/thought could be done symbolically. Most AI types disagreed, and most still do.
Another amusing bloke was Jerry Fodor. Way sharper than Searle, I thought. To overstate things a bit, Fodor’s argument was that human thought and understanding are infinitely perfect, that they take everything possible into account, and computers couldn’t possibly do that. I am, of course, fond of that, since I think people really do think, and that human thought is seriously amazingly good. But it ain’t perfect. (Fodor’s “the mind doesn’t work that way: the scope and limits of computational psychology” is a must-read. (Interesting: it’s copyright 2000, which is much more recent than I thought.))
For the me the central take-away of the Chinese Room thought experiment is the issue whether we handle the term intelligence (or rather “understanding”) as strictly being determined by observable behaviour. The Chinese Room is meant to illustrate that what is observable isn’t all there is to it. This of course is controversial and can be challenged. Searle’s original point went further than that, but in any case the Chinese Room is still good for illustrating this issue.
David in Tokyo, Searle certainly knows Chinese is just a language. His point in using Chinese specifically is that it makes credible his assertion that he does not speak it _at all_. For an English speaker to say he knows no French or German or Italian or Norwegian wouldn’t really be credible, we all know at least a few words of these languages (at a bare minimum, the words English has in common with them), and the grammar is similar too. Searle is trying to create a scenario in which he, and he alone, in a room with a set of instructions (perhaps very long) and an appropriate set of files (perhaps very large) could take an input in Chinese and produce a suitable output in Chinese — an output that could plausibly come from a Chinese-speaking human — in spite of the fact that Searle himself does not know any Chinese whatsoever.
That Wikipedia article is pretty good, and it’s short; you might give it a read.
Phil et al. Sometimes I enjoy using the Stanford Encyclopedia of Philosophy for a next-up level article for philosophy or related topics (like QM or etc)
https://plato.stanford.edu/entries/chinese-room/
Searle ultimately believed that based on his argument there were special “causal powers” of biological systems which allowed them to form minds which “inorganic computers” couldn’t do.
No one of course is able to get him or anyone who supports that view to explain further… because as far as we know all computers are equivalent so either minds are “magic” or they’re Turing equivalent…
Clearly, if you can *always* output convincing chinese then you’re a chinese speaker. Of course plenty of Chinese speakers don’t know anything about The Big Lebowski, or reruns of Threes Company, or Quantum Electrodynamics. One of the big problems with LLMs is that they don’t know what they don’t know, so they just make up words that are grammatical like a pathological liar. They can be caught out specifically by the fact that they are pathological liars about verifiable facts.
A Chinese room that didn’t get caught out lying about everything, that “knew what it knew” so to speak, would be an intelligence that speaks Chinese IMHO.
Daniel:
Just don’t rely on the Stanford Encyclopedia of Philosophy article on The Ethics and Rationality of Voting. It’s terrible. Even philosophers can get confused sometimes.
Daniel,
I think the idea that biological computers have some sort of magical properties that allow for intelligence, making it unattainable by any non-biological computer, is…I was going to say stupid but perhaps I should really say it seems extremely implausible to me.
But that doesn’t mean I agree that if a machine can simulate intelligence extremely well, it is therefore intelligent. Trivially this depends on the definition of ‘intelligence’, which is close to Bob’s original point, but I don’t think this is purely an argument about definitions, or at least it doesn’t have to be. Recently I’ve been thinking of it this way: is “artificial intelligence” more like “artificial light”, i.e. it is real but is generated by “artificial” means; or is it closer to “artificial flowers”, which simulate flowers in some important ways but which contain no actual flower material whatsoever.
The current LLMs do a fantastic job at simulating intelligence, but examples like the initial inability to correctly answer the glass-door version of the Monty Hall problem, and the current inability to generate a picture of a ‘man riding a horse backwards,’ seem to me to lean strongly in the “artificial flowers” picture of artificial intelligence: these systems seem like they ‘understand’ the prompts, and the answers, but they don’t. As they improve they are going to get better and better at simulating understanding, perhaps to the point where I’m unable to catch them out…but does this mean they really do understand?
Daniel, I share your intuition that the only way to make a system that never gets flummoxed would be to make one that really does ‘understand’, but that doesn’t mean I’m comfortable with a completely operational definition of ‘intelligence’ or ‘understanding.’ Not that I have an alternative to suggest! Ultimately we can only judge based on how these systems respond to our inputs, and if we don’t agree that there’s some set of (input, output) that indicates intelligence then we can give up on ever determining that an AI is intelligent.
I still haven’t found a good way to think about this stuff.
Phil,
imho the LLMs are not generally intelligent, but they *are* intelligent about certain aspects of the world. You can use english to ask them to write fairly specific bits of software, and they’ll do it for example. Like an expanded version of “write a function in Julia that takes two arguments one of which is a DataFrame that has 3 columns…. etc etc… and make a plot that contains…”
is likely to get you a proper Julia function that does what you asked it to. If it gives you a slightly wrong function and you point out the problem and ask it to correct it, it’s likely to do that too.
In this sense I think it understands what it means to “write a function that…”
On the other hand, here’s an interaction I just had with Chat GPT:
—-
ME: Can you write a paragraph summarizing the novel “The Raven’s Nest” by Daniel Lakeland published in 2018?
GPT: “The Raven’s Nest” is a gripping psychological thriller by Daniel Lakeland that delves into the complexities of human nature and the consequences of past actions. Set in a small, isolated town, the novel follows protagonist Sarah as she returns to her hometown to confront the haunting memories of her childhood. As she unravels the mysteries surrounding her family and the dark secrets buried within the community, Sarah finds herself entangled in a web of lies, betrayal, and manipulation. Lakeland expertly weaves together suspenseful twists and turns, keeping readers on the edge of their seats until the shocking climax reveals the truth lurking in the shadows of “The Raven’s Nest.”
—-
Sounds pretty good! It’s a decent blurb for the novel I’ve never written that doesn’t exist. ChatGPT has no idea if such a novel exists, or what the contents of it are. It just knows what blurbs about novels typically sound like in a generic sense, and it gives you that. You can catch it out in its non understanding quite easily. Of course it clearly *understood the language of the task* it just doesn’t *understand the state of the world* and hence it can’t match its language to the state of the world. A proper response from a proper “Chinese Room” would be something like:
“I’ve never read The Raven’s Nest by Daniel Lakeland, so I really don’t know what it’s about. Have you read it? Is it any good?”
This is pretty fun actually… here’s another interaction:
—-
Me: Can you give a plot summary of the erotic novel Moby Dick
GPT: Certainly! However, “Moby Dick” is not an erotic novel; it’s a classic work of literature by Herman Melville, first published in 1851. The novel follows the story of Ishmael, a sailor who joins the crew of the whaling ship Pequod, commanded by the obsessive Captain Ahab. Ahab’s singular obsession is to seek revenge on the white whale, Moby Dick, who had previously maimed him.
As the Pequod embarks on its voyage, Ishmael narrates the crew’s experiences, encounters with other whaling ships, and the diverse characters aboard, including the enigmatic harpooner Queequeg. The novel delves into themes of obsession, fate, morality, and the complexities of human nature.
Ultimately, the pursuit of Moby Dick leads to the destruction of the Pequod and most of its crew, with only Ishmael surviving to tell the tale. “Moby Dick” is celebrated for its rich symbolism, philosophical depth, and exploration of the human condition, but it is not an erotic novel.
ME: Oh sorry I meant the erotic novel Moby Dick whose title is a parody of the famous novel by Melville
Ah, I see! You’re referring to “Moby Dick, or The Card Game,” a parody novel by Jasper Fforde. In this satirical work, Fforde takes the premise of Herman Melville’s classic “Moby Dick” and adds a humorous, erotic twist to it.
In “Moby Dick, or The Card Game,” the story revolves around a group of characters who engage in a card game that serves as a metaphor for the pursuit of love, lust, and desire. The protagonist, Captain Ahab, becomes obsessed not with hunting a whale but with winning the affections of a particular woman. As he navigates the complexities of the card game and the relationships between the players, Ahab grapples with his own desires and motivations.
Throughout the novel, Fforde injects humor, wit, and sexual innuendo into the narrative, creating a playful and irreverent parody of Melville’s original work. While “Moby Dick, or The Card Game” retains some elements of the source material, it takes them in a wholly different direction, exploring themes of romance, sexuality, and human folly in a lighthearted and comedic manner.
———–
That *is* extrapolation, which is what I think is really meant by “causal reasoning”. It is just bad at it.
It does seem these LLMs are near human level at interpolating and identifing authority/consensus positions though. And really that is how humans operate 99% of the time.
Eg, I was reading some page about Genghis Kahn the other day, and really couldn’t tell if the author was a human or bot.
https://www.howandwhys.com/joe-rogan-reacts-to-quest-to-discover-genghis-khans-tomb-that-has-never-been-found/
I think a bot would avoid such glaring inconstencies. So this was a lazy human. Or a human pasting together bot responses.
Aren’t we tiring of these endless discussions about the “intelligence” of LLMs? As the LLMs get ‘smarter’ and humans get ‘dumber’ I think it is a foregone conclusion that we will find it increasingly difficult to tell them apart. That won’t resolve the debate, nor should it. But I think the consequences are far more interesting than trying to define “intelligence” or identify a bright line that distinguishes between human and machine whatever. Andrew has said that 90% of the time he does some form of autocomplete – Anon seems to be putting something like that at 99%. And these are 2 smart people – what about the masses?
I should point out that this is incorrect. ChatGPT doesn’t do any short-term learning or adaptation either. What’s going on during the “conversation” is that, every time you submit something, the entirety of the conversation so far is included in the prompt.
This is why semantics is so tricky. This technique of prompting is called “in-context learning” everywhere, even in the Wikipedia article on prompt engineering: https://en.wikipedia.org/wiki/Prompt_engineering#In-context_learning . One of the key papers in the LLM literature is “Language Models are Few-Shot Learners” from Open-AI, which introduces the whole notion of in-context learning. One could argue, Gelman style, that their choice of terminology was ill-conceived.
On the other hand, the “PT” in “GPT” is “pre-trained”, by which they mean the weights of the network are not updated during conversations. If you define learning to be training neural network weights, then ChatGPT doesn’t learn during a conversation by that definition.
OpenAI hasn’t divulged exactly how they’re supplying context to ChatGPT based on conversation history. There’s a finite limit, so it can’t be all of the history for long dialogues. But it also doesn’t have to be the literal history—the history can be compressed, edited, etc.
” In general, humans are pretty bad at causal inference.”
Hmm. My take is exactly the opposite: we’re really really good at causal inference. Every time we run across/think about something, we ask why. And try to find a causal answer that matches what we know about reality. Sure, we don’t make the effort a lot of the time (see Thinking Fast and Slow). Or maybe sometimes we try too hard to come up with something when we don’t have anything. Or do motivated reasoning to come to desired conclusions based on a political or other belief system. But it’s always “why”.
But that’s fine, because we really actually can do causal inference*. We know what words mean, and work with those meanings. And we’re really really good at figuring out some really complicated cool stuff: the standard model, eliptic curves, functional harmony (both the classical and jazz versions).
LLMs don’t do inferencing, causal or otherwise. At all.
*: Note the double standard! When an LLM fails, I claim the LLM failed. When a human fails, I claim they just didn’t do the work. It’s a double standard. But it’s right.
The hilarious thing about LLMs is that not only don’t they do inferencing, but they don’t have a clue about the meaning of the text strings they generate. The whole game is spitting out a completely ungrounded (in any idea of reality whatosoever) random string of undefined tokens, and
_get the user to do the work of figuring out what that random string of undefined tokens means_.
The user then gives the LLM credit for understanding what it was saying, when it didn’t.
(I’ll allow the reader to infer the class of objects I think this mechanism falls under.)
One can assert that “LLMs don’t do inferencing, causal or otherwise.” and “we really actually can do causal inference” but the philosophical game, if you want to play it, is to try to refine the definition of “inference” in a way that this follows as a consequence. The point is developing a coherent definition of “causal inference” whereby humans can do it but what ChatGPT is doing is not it.
Why do all these examples involve mandating/banning things? Rather than intelligence, I think it comes down to some people having a lower threshold for ordering others around.
As best I can recall, it was a long time ago, the early discussion around smoking and cancer did not involve banning or mandating anything. I think the original idea was that educating people about the serious health dangers of smoking would lead them to give it up. It was also anticipated that the tobacco companies might respond by coming up with ways to make cigarettes less dangerous.
Mandatory warnings on packages, and bans on smoking in most indoor spaces came decades later. Are you (implicitly) claiming that Fisher anticipated in the 1950’s policies that would be mooted and implemented in the 1980’s and beyond?
There is a long history of trying to ban it, going back to literally the first European smoker:
Then in the 20th century it was the Soviets and Nazis:
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5637694/
So the Spanish Inquisition, Lenin, and Hitler (Stalin smoked, so wasn’t interested in a ban). There were various other attempts in the meantime using racism against Native Americans as the excuse.
But I was thinking more that once you start trying to force others to go along with your idea, an unwarranted confidence develops as justification.
Anon:
I don’t think Fisher was trying to ban anything, although it seems that he did get pretty angry when people pointed out his mistakes.
Regarding banning cigarettes, or at least the cessation of their manufacture, recall that in 1986, the president of R. J. Reynolds, one of the largest cigarette companies, said that if he ever “saw or thought there was any evidence whatsoever that conclusively proved that, in some way, tobacco was harmful to people, and I believed it in my heart and my soul, then I would get out of the business.” In 1954, the vice president of Philip Morris, another biggie, said that his company would “stop business tomorrow” if “we had any thought or knowledge that in any way we were selling a product harmful to consumers.” And in 1972, the president of R. J. Reynolds (not the same guy who was president of the company in 1986) said, “If our product is harmful . . . we’ll stop making it.” The president of Lorillard said, “if it were proven that cigarette smoking caused cancers, cigarettes should not be marketed.” So I guess you can add the presidents of Philip Morris, R. J. Reynolds, and Lorillard to your list which also includes the Spanish Inquisition, Lenin, and Hitler.
OK, those cigarette company presidents just wanted to stop making and selling cigarettes. I don’t think anyone was talking about making it illegal to smoke hand-rolled tobacco cigs. The major public health problem seems to have been as much from the mass production of the deadly product as from the tobacco leaf itself.
Did anyone ever believe those tobacco CEOs?
Anyway, point is cancer is just the latest in a long list of reasons to ban smoking. Really its only been the excuse for about 15% of the time. Lots of people just think smoking is annoying because it smells gross or irritates their allergies.
I wouldn’t be surprised if in the future cancer is seen as much more a genetic/developmental issue, and the cancer reason is replaced by cardiovascular disease. Because in the end it really is just correlation = causation, which is all Fisher pointed out in the 1950s.
I’d guess it increases the turnover rate of the lung tissue, and carcinogenesis is ultimately caused by cell division approaching/beyond the hayflick limit. So it should be possible to get a quantitative model that predicts exactly who get cancer if we could measure that turnover rate. Then you would have causation.
It has always been odd to me — and frustrating to many public health researchers — that lung cancer gets all the press when it comes to smoking, given that the effect on heart disease is so much larger. So I hope Anon is right and the future focus shifts to cardiovascular disease.
Just my poor choice of examples. Let’s mine recent topics in the blog, all of which are obviously being put forward as causal: ghosts, conspiracy theories, ESP, pyramid power, ancient astronauts, astrology, homeopathic medicine, power pose, ovulation, clothing color, nudges, hot hand (love that the polarity flipped and flipped back), etc. There’s a whole sports, social science, systems biology, and biomedical literature out there just waiting to be mined for ridiculous causal claims.
How much of those differences In belief are due to background knowledge?
I think its still orthagonal to intellegence.
For example I don’t believe RCTs provide causal evidence. Instead, causality comes from the model of what generated the data. If your model makes surprising predictions that come true, then whatever causality you built in can be taken as correct (until a better model cones along).
Does it really make sense to talk about “causality” when looking at averages? The average person doesn’t exist.
Bob:
I agree with your general point that there are problems to defining some threshold of intelligence or understanding.
One thing I think do is use the success of chatbots to get some understanding of our own thought processes. I tried to get at this in my post, I disagree with Geoff Hinton regarding “glorified autocomplete”, where I wrote:
I continued:
Finally:
Also relevant is this post, “AI” as shorthand for turning off our brains. (This is not an anti-AI post; it’s a discussion of how we think about AI.).
I replied to those posts. Herb Simon, in his cognitive sci class, used to emphasize just how slow our brains are in terms of clock cycle, but just how good they are at associative reasoning. His theory (I should say, “among his theories”) was that learning amounted to chunking together operations and recognizing where the chunk applies.
A major issue with the intelligence concept is that most people who use it aren’t clear about to what extent it can be diagnosed by observation from the outside. The most extreme approach would be to define intelligence by specifying a bunch of observable criteria (an intelligence test that is perfect by definition, “intelligence is what intelligence tests measure”) and then say, whatever fulfills these is intelligent. Chances are LLMs and potentially other existing systems are already intelligent according to pretty much any set of observable criteria that has been specified at some point in the past. But then this is not satisfactory and never has been (which was for example addressed by the Chinese Room argument).
Another approach would be to say, intelligence is something that exists and can be diagnosed by tests, but the tests (and observable criteria in general) do not *define* intelligence, rather they are estimates of it that are potentially biased, and the relation between what can be measured/observed and the “true” intelligence itself is complex and not fully understood. One can also distinguish between the idea that there is a true one-dimensional level of intelligence, and that it is potentially multidimensional or not even quantifiable. According to this approach it is not straightforward what the implications of anything that a machine does and gets right or wrong are regarding its intelligence.
The idea addressed in this posting that some people say that “intelligence is whatever machines cannot (yet) do” seems to a combination of an naive concept of observability of intelligence with an unwillingness to ascribe it to machines. This of course doesn’t hold water, however anyone who doesn’t want to attribute intelligence to machines can just drop the naive idea of observability. Then arguing that machines cannot be intelligent may still not convince a majority of (more or less qualified) people, but this is a different debate from the one treated here.
I’m not totally sure to what extent you, Bob, agree with what I would call naive observational approach – this posting and earlier ones I have seen gave me the impression that your point is in fact “machines can do XXX now, so we should call them intelligent”. This is a point of view one can hold, but I don’t think it appropriately captures how we tend to think and reason about intelligence. To me it seems more appropriate to use the term “intelligence” in such a way that it has aspects that are essentially not directly observable. I accept that observational definitions are easier and in some sense better to handle than those referring to non-observable aspects, but I don’t think that this trumps the consideration that with a purely observational definition we lose quite a bit of what is implied in the discussion about intelligence, particularly including the development of human intelligence.
One way to “read” the observation that in some discourse “intelligence is whatever machines cannot (yet) do” is that we develop our concept of intelligence, and that we learn, that certain things we believed machines could not do (at least not using the range of algorithmic capacities that we can currently understand) actually can be done by machines. This may in fact change and sharpen the concept of intelligence, with the implication that something that would have looked like a strong sign of intelligence 30 years ago doesn’t look so impressive anymore now. We have in fact learned a lot about how behaviour that has been taken as a sign of intelligence in the past can emerge from putting algorithms and data together, and the consequence may be that the concept of intelligence shifts away from these things. This is of course “moving the goalposts”, and Bob may complain about it, but I’d say there is some sense and meaning in this, and it cannot be discounted that easily.
I don’t want to take a clear position regarding how intelligence should be defined here, because I think that such concepts are ultimately human constructions, and as such dynamic, and to some extent matters of decision. I do understand the unwillingness to withhold the term intelligence just on the basis that it refers to something that isn’t observable, because without any connection to observables obviously the concept is inaccessible for science, and some may even say, for any meaningful discussion. Still, we have a choice how to use the term “intelligence”, our use can be dynamic, and any understanding of intelligence requires facing the role of observation when diagnosing it, which for sure is not straightforward.
I’m not entirely sure what you mean by “naive observational approach.” The outside perspective was the motivation for Turing’s test. This has been reified in both behavioral psychology (e.g., Skinner) and in the more recent (though still 50+ years old) “functionalist” philosophy of mind. I’d say I’m pretty much a functionalist in this sense. To not be seems to be to require some sort of spiritual definition or something too reductionist. But it’s all very hard to talk about precisely.
It seems to me to do less damage to the common notion of “intelligence” to include GPT than to exclude it. I’m very much of a pragmatist linguistic philosophy, and thus think this is purely our choice to make as communicating agents.
I like your conclusion. My whole point is that these notions are dynamic and human constructed. We can’t go do an empirical investigation to see if something is “intelligent”, it’s just a matter of semantics.
What I meant by “naive observational approach” is what I explained in the beginning: “to define intelligence by specifying a bunch of observable criteria and then say, whatever fulfills these is intelligent”. This is obviously connected to Turing, behavioral psychology, functionalist philosophy of mind, although I’m not going to call these “naive” without deeper knowledge of the literature. Particularly you acknowledge the dynamic of the concept, which means you apparently don’t operate with a “closed” observational definition.
The key question then may be: What are the implications of this (or other) use of the term “intelligent”? What do you mean by “doing damage” that you think should be avoided?
So far, as I see it, some AI programs are intelligent at certain tasks. I guess the outstanding example, for me, is AlphaGo, after over 2400 years of humans studying and writing about Go, coming up with a move in its tournament with the world champion which none of the experts watching had seen before, and of which the (now former) world champion later said, “I never thought a machine could beat me, but when I saw that move I knew I would lose the tournament.”
I, for one, would have held AlphaGo responsible as not being the world champion had it lost that tournament, but I agree that autonomous responsibility is mostly withheld from AI programs, because humans prefer to remain in control of them (among other reasons).
Also, currently LLM’s lack an executive function to double-check their work and the ability to train continuously; not, I think, because such things will be impossible to add, but because they are difficult to add all at once, and will cost more resources.
Given that it currently takes about 1000 neural-network nodes to emulate a human neuron, AlphaGo had less than a mouse’s cerebral cortex, and ChatGPT has about an order of magnitude less than a human’s. Our nano-tech brains are the one resource I think space aliens might want to conquer the Earth for (assuming they could harvest them to run their systems).
As a fallible and autonomous human, I was triggered by the mention above of Searle’s Chinese Room “argument”. When I first read about it in Dr. Penrose’s “The Emperor’s New Mind”, despite my great respect for Dr. Penrose (of course) it seemed (and seems) bogus to me. A neuron in a Chinese-speaker’s brain does not understand Chinese either. It is the gestalt of a huge number of neurons working together at once that understands Chinese. So either I or Searle or both has/have had a failure of our intelligence/s, and perhaps they are only parlor tricks.
I’ve been ripped off. Plagiarized. Become a victim of intellectual theft.
In an article* titled “LLMs aren’t very bright. Why are so many people fooled?”
Gary Marcus writes: “For a long time I have been baffled about why so many people take LLMs so seriously, when it is obvious, from the boneheaded errors, hallucinations, and utter inability to fact-check their own work, that the cognitive capacities of LLMs are severely limited.”
Thus irritating me no end for saying it more clealy than I’ve been able to. He takes my “gets the user to do the work of understanding the text” bit several steps further, arguing that what LLMs due is analagous to psychic readings**. Well, sure. But I like my “gets the user to do the work” better, since it’s a more accurate description of what’s going on. But, like, whatever.
** This is taken (with attribution) from Baldur Bjarnason’s artice “The LLMentalist Effect: how chat-based Large Language Models replicate the mechanisms of a psychic’s con.”
So far so good, But the problem with Marcus’ article comes at the end, where he writes:
“Gary Marcus thinks the whole world would be different if so many people weren’t so easily taken by the parlor tricks of generative AI.”
Grumble. Parlor trick is my term. I’ve been ripped off. (Truth in advertizing: I’d be surprised if I were the first one to have used it in this way.)
*: https://garymarcus.substack.com/p/llms-arent-very-bright-why-are-so
I think the analogy to psychics is as misguided as the rest. I have GPT teach me things all the time—it’s a greet teacher that can drill down into follow ups and phrase things in the language you want (for example, I use it to translate measure theory back to Lebesgue land for me). I use it almost every day (including yesterday and probably today) to write code that works. I’ve had it brainstorm creatively on everything RPG-related I’ve done since it came out. I don’t think I’d have as much luck asking a psychic to write ggplot code for me. We fed my dad’s poetry in and had it provide constructive feedback—it’s really good as a writing coach.
If ChatGPT is a “parlor trick”, then your notion of parlor trick is much broader than mine.
When you feed in language and have it rewrite or comment on the language it makes good sense because it has specifically the input to work with. When you ask it about facts about the world… it doesn’t. Like the link I posted elsewhere to a conversation about The Big Lebowski. It just makes up scenes in the movie that don’t exist. On the other hand if you fed it the script of the movie and asked it to explain a scene, it could work from that text to provide you some analysis of the scene.
I think I’m more impressed with IBM Watson now than I was at the time it won Jeopardy.
My pet theory is that a lot of argument stems from society inappropriately bundling concepts. With climate change for example, I think some people are opposed to government action on grounds of being inherently anti-government, believing there are more pressing issues, that we’ll adapt, preferring paper towels to hand dryers, or that it’ll be cheaper to deal with later. However, scientific measurement of climate change seems to always be intermixed with advocacy for government action. So if you’re opposed to government action, you’re opposing science. Makes it harder to have a policy debate and shifts a lot of the debate to discrediting the science and scientists.
With AI being intelligent, I think there’s bundling going on that if AI is intelligent, we can put it on the same scale as humans and dogs. So you see graphs like this: https://content.time.com/time/daily/2011/1102/singularity_graphic.jpg
Suddenly, if you agree GPT is intelligent, you have to answer whether it’s smarter than a dog. And ChatGPT can explain calculus, so you feel a bit foolish saying it’s dumber. Then again, a dog has self-preservation, so maybe you feel foolish saying GPT is smarter. I think we’re creating artificial intelligence that operates on fundamentally different principals than biological intelligence, so you will always be able to point to differences. Labeling those differences smart or dumb is putting it on a scale it doesn’t belong on. It’s more instructive to just put it in it’s own box, evaluate what this new concept can and can’t do, and try to figure out what limitations may exist as we continue to scale up our current approach.
Hmm, which “people” are referenced here in claiming “I doubt people 50 years ago (1974) would have said you can play chess without being intelligent.”? By 1974, I think chess computer programs had gotten good enough that they could certainly “play chess” well enough to beat many humans – not the better players, but that wasn’t a requirement (plenty of humans “play chess” badly!). Certainly by a few years later, you could buy specialized chess computers as a consumer item, and they definitely did “play chess”. Not grandmaster-level chess, but skilled enough so if (not when) you got to the point of being able to beat it consistently, you’d be a respectable club player. The only question was more along the lines of, could the computer programs ever play better than the very best humans? That difference is what punditry focused on. Yet before that, being able to defeat all but a relatively small number of humans seems not all that much different.
That is all true, but I think a reason that people would not have considered those programs intelligent, is that all the strategy in them was conceived by humans and directly programmed into them. Whereas neural-network AI’s are given the rules and then train themselves, which is closer to how human intelligence works. For example, that is how I learned to play tic-tac-toe as a child.
Indeed: ““I doubt people 50 years ago (1974) would have said you can play chess without being intelligent.”
was certainly not true of the top computer chess developers, for example those who created Belle:
https://en.wikipedia.org/wiki/Belle_(chess_machine)
I worked at Bell Labs 1973-83 and followed this, since I knew these folks.
Belle had opening book and endgame databases, with well-optimized tree-search in between.
As per Wikipedia:
“Because of its ability to generate and analyze many chess positions, Belle represented the brute-force approach to chess computing. In the late 1970s, Thompson became interested in the limits of this method, playing different versions of Belle against one another. Using identical machines allowed him to minimize effects of the individual machine’s play style while isolating the effects of search depth. For instance, if one Belle computer searches three levels deep, the other might search to four. Thompson concluded that for each additional level of search, Belle improved by approximately 250 rating points.[14][15] This effect has been replicated in self-play experiments with different machines.[16] Beyond 2,000 points, however, Thompson found that improvements leveled off.[17]”
Ken rarely gave lectures, but did one inside BTL on chess in early 1980s, commented on search levels vs USCF ratings then, noting that didn’t apply well to Grandmasters.
I got Ken to give Computer History Museum oral history on chess:
https://www.computerhistory.org/collections/oralhistories/?s=thompson+mashey
Of course, checkers was solved, using similar methods, not AI. (I’d met Jonathan Schaeffer,as Chinook used MIPS/SGI gear at the time):
https://en.wikipedia.org/wiki/Chinook_(computer_program)
Of course, finally serious advances from AI, and a bit of a surprise to those who followed computer game playing and thought Go was way harder than chess,
https://en.wikipedia.org/wiki/AlphaGo
As of 1974 or so, everyone and their brother (MIT, Bell Labs, Northwestern, and IBM) had realized that brute force search would result in improved play and not only were busting their butts in the coding (MIT’s “Tech” program, for example, generated bishop/rook/queen moves in one (PDP-6) instruction per move (asymptotically)), but also had hardware designs in the works. (I had forgotten that Bell Labs actually got their version built: the college designs never had the money to spare.)
The insight that made Go programs work was Monte Carlo tree search (about 2005). Folks who had worked on chess were flabbergasted that random search was of use in game playing: it was a major intellectual achievement. At least for the game-playing computer world.
My reading of the AlphaGo story was that the Japanese program “Tenchou Igo*” had been beating pros with a three stone handicap and had finally beaten a retired pro in an even game (on a generic PC without a GPU). So the writing was on the wall that computer Go had been figured out, so Google threw a server farm at the problem***.
People working on the current top Go programs are of the opinion that AlphaGo was pretty kludgey (no papers were written on it, the code wasn’t released), but that AlphaGo Master (which played 60 “anonymous” games on the internet, and revolutionized Go opening theory) was a cleaned up version. And experience with those led to the self-play learning bit (AlphaGoZero, which was documented in detail). Modern programs, such as KataGo are several generations beyond AlphaGoZero.
FWIW, my take from the game-playing computer world is that Roger Schank is right: “There’s no such thing as AI”. These are all special-purpose kludges for playing these games, and have no intellectual or philosophical relationship whatsoever to anything else in the universe.
The only conclusion to be drawn is that nerds (myself included (I once implemented Alpha-Beta for Othello)) like to write game programs.
*: The final released version was “Tenchou Igo 7” (known as “Zen 7” in English) and plays slightly super-human Go on a 2023 generation PC with an i7 and lots of memory without using the GPU. Since development stopped in 2019 or so, it plays hilariously old-fashioned Go completely oblivious to the revolution in opening theory of the stronger programs. But it still plays really nasty Go.
**: The Tenchou Igo folks were unfunded academics, and there were also Chinese groups writing Go programs and running them on server farms as well. So (as you can tell that I intend to imply above), even though the Tenchou Igo folks were doing a better job of the implementation, they couldn’t compete with server farms.
***: By the way, I also suspect that the Deep Blue story was similar, in the sense that programs running on a generic PC were super-human strength by 2004 or so, so IBM may have realized that there was a narrow window of opportunity for their hardware before PCs caught up. Deep Blue won in 1997, but a program drew the world champion in 2002, smartphones became stronger than any human in history in 2009.
Yes. When Ken Thompson gave BTL lecture in early 1980s, he noted AI attempts hadn’tworked well.
He mentioned an experiment to figure out how chess masters analyzed positions:
1) A position would be revealed to master, whose yes would be tracked.
2) They’d look at the piece they’d move, look around the board, then move the piece.
This did not lead to useful AI strategies. (The resulting conjecture was that masters would take in the position as a gestalt.)
So how about what AI can do that humans can’t?
This is kind of stunning. I think there must be problems to get such a huge effect. But still…
Durably reducing conspiracy beliefs through dialogues with AI
Key finding: “The debunking conversation with the AI reduced belief in conspiracy theories by roughly 20%. This effect did not decay over 2 months’ time … and occurred even for participants whose conspiracy beliefs were deeply entrenched.”
The authors’ theory is that “prior attempts at fact-based intervention may have failed simply due to a lack of depth and personalization of the corrective information” and they suggest that advanced AI bots are well-suited to correct exactly that problem.
“Reasoning-based strategies were clearly the most frequently used … Conversely, the rapport-building strategies of finding common ground and expressing understanding were used only “moderately” and … various psychological and social/emotional strategies were used even less.”
https://t.co/aN8jAXXArr
The author of those comments:
https://x.com/ichbinilya/status/1780133566694768731
I actually tried to use chatgpt again today and it couldn’t give me a straight answer (just like every time).
I wanted to know how people measure duration of an event in attoseconds (10^-18) when the definition of a second is only accurate to a femtosecond (10^-15):
https://pastebin.com/1ZANr7vQ
This seems like an obscure, yet uncontroversial question. It basically kept saying you have to be really smart and use ultra-fast lasers.
I assume I’d get a better answer on stackexchange, but maybe not.
The definition of a second is accurate to infinite decimal points by definition.
One second is 9192631770 cycles of the radio waves emitted by cesium under some particular conditions. You can still divide that much time into say a trillion parts and get something much smaller than one cycle.
https://phys.org/news/2015-06-atomic-clock.html
I didn’t manage to parse the jargon on this topic (hence trying to ask chatgpt).
But as far as I can tell, if you start counting cycles of two cesium clocks at the same time, after about 1 million seconds (~10 days) you are likely to find one has undergone X cycles while the other X +/-1 cycles. Further, this is a practical problem people are trying to overcome.
So you can keep dividing indefinitely, but eventually you end up with multiple equally valid answers.
Your question, as stated, apparently has stumped not just ChatGPT, but a very intelligent person (Dr. Lakeland) (as well as myself). It illustrates the point of this topic: people tend to expect or demand more of machine intelligence than they get from most fellow humans.
The definition of the second isn’t in terms of count of caesium clock cycles, it’s in terms of the wavelength or frequency of the photons which can induce the transition between two energy levels in a particular *model* of the caesium atom. It has zero uncertainty and is calculable in terms of fundamental constants, but is then measured by reproducing the model as best we can in a laboratory, which can be seen here on page 128
https://www.bipm.org/documents/20126/41483022/SI-Brochure-9.pdf/fcf090b2-04e6-88cc-1149-c3e029ad8232
When people say “a more precise definition of the second,” what they really mean is “a redefinition of the second in terms of a model that we can approximate more precisely with a real-world system.”
No response for a couple hours = stumped apparently. Quite a crusader, here. If you actually read the pastebin though, ChatGPT’s answer is not just stumped, it’s meaningless gibberish, stating both A and its inverse right next to each other and comparing incomparable things. I don’t understand the incessant need to put ChatGPT and humans on an intelligence scale like we’re comparing SAT scores. ChatGPT is obviously doing something categorically different than humans are, and can do things humans can’t while failing in ways humans never would.
Just saw this today Apr 18 at 9:40 Los Angeles time… so dunno what “stumped” means in this context.
From the Wikipedia on the second which is consistent with other sources I’ve read in the past, but I’m not exactly an NIST scientist so perhaps it’s wrong, in which case someone should probably edit it: https://en.wikipedia.org/wiki/Second
“Since 1967, the second has been defined as exactly “the duration of 9,192,631,770 periods of the radiation corresponding to the transition between the two hyperfine levels of the ground state of the caesium-133 atom”.”
So yeah the second is the *duration* of exactly that many cycles of radiation coming from a particular excitation of cesium-133
Of course, working with 9 gigahertz signals in real-world electronic circuits is subject to all kinds of environmental noise etc. So yeah you can certainly have divergence between different clocks in the real world. This just means that the clocks are not perfect not that the definition isn’t perfect. It’s perfect **by definition**.
I’m also not an atomic physicist but I think the wikipedia article on the cesium standard is pretty decent: https://en.wikipedia.org/wiki/Caesium_standard
“The meaning of the preceding definition is as follows. The caesium atom has a ground state electron state with configuration [Xe] 6s1 and, consequently, atomic term symbol 2S1/2. This means that there is one unpaired electron and the total electron spin of the atom is 1/2. Moreover, the nucleus of caesium-133 has a nuclear spin equal to 7/2. The simultaneous presence of electron spin and nuclear spin leads, by a mechanism called hyperfine interaction, to a (small) splitting of all energy levels into two sub-levels. One of the sub-levels corresponds to the electron and nuclear spin being parallel (i.e., pointing in the same direction), leading to a total spin F equal to F = 7/2 + 1/2 = 4; the other sub-level corresponds to anti-parallel electron and nuclear spin (i.e., pointing in opposite directions), leading to a total spin F = 7/2 − 1/2 = 3. In the caesium atom it so happens that the sub-level lowest in energy is the one with F = 3, while the F = 4 sub-level lies energetically slightly above. When the atom is irradiated with electromagnetic radiation having an energy corresponding to the energetic difference between the two sub-levels the radiation is absorbed and the atom is excited, going from the F = 3 sub-level to the F = 4 one. After a small fraction of a second the atom will re-emit the radiation and return to its F = 3 ground state. From the definition of the second it follows that the radiation in question has a frequency of exactly 9.19263177 GHz, corresponding to a wavelength of about 3.26 cm and therefore belonging to the microwave range. ”
To understand what a “hyperfine transition” is, you can see this pic which I found by web based image-search:
https://www.researchgate.net/figure/A-sample-hyperfine-spectrum-showing-transitions-in-rubidium-85-from-F-3-to-F-J-2-3_fig4_252708609
That site shows how the single peak on the left visible in the “No B Field” splits into two peaks “with B field”. The radiation emitted (or absorbed) during transitions between two nearby peaks like that in the Cesium spectrum is what’s referred to in the definition.
By “stumped” I meant “did not understand what Anoneuoid’s problem was well enough to clear it up”. I should have thought of a better word. No pejorative was intended. As I implied, I did not understand the question well either, and I expect ChatGPT did not, and also lacks the ability to do research on Wikipedia. I also don’t think it has anywhere near the capacity of Dr. Lakeland or Dr. Somebody, as I stated in my first comment.
I do think neural-networks are a legitimate area of research with some accomplishments. At the same time I think modern corporations, headed by people similar to Steve Jobs and Jack Welch, are quite capable of misusing that research.
I apologize for my unintentional offense. I will not comment here again for at least a month as a penalty.
JimV. No way dude. You didn’t offend me at all, I was just confused. You explained, I understand your explanation, no harm done.
Yes, it can be hard to communicate ideas through written words. To the extent that ChatGPT etc can provide some useful something it has to 1) have sufficient training data 2) the training data has to be relevant and of high quality 3) the question you ask has to be well posed in such a way that it’s relevant to the training data. 4) there’s probably some luck involved in the sense that random numbers may be involved
And yeah, communicating your question to a human can be hard, it’s no surprise that a language model also can be confused about the way a question is posed.
Heres a talk from someone actually working with attoseconds. At 23:20 she explains there is no clock that can be used to measure these durations:
https://youtube.com/watch?v=KoHKVXxT7uA&t=23m20s
Essentially they had to come up with their own “stopwatch”. It lets you measure something for 5 femtosec, and say during that interval event A occured 10 attosec before event B. But there is no way to map those events to, eg:
12:30:00.0000000000000000 UTC
vs
12:30:00.0000000000000001 UTC
Those timestamps are arrived at by tuning a laser to do whatever to cesium, then tallying up the number of times a voltage crossed zero since some arbitrary moment. The interval between zero crossings is defined as 1/9,192,631,770 = 1.08782775708e-10 sec. Everything that happens during one of those intervals has the same timestamp.
Not quite sure what the 10^-15 femtosecond precision refers to then though. Seems like cumulative error over time.
Anoneuoid,
Conceptually, there is no problem with attosecond precision. In terms of building a clock, yeah, it’s tough. But for example, you could tune a maser (9Ghz is still radio frequency) to do something to cesium and then quadruple the frequency to 36Ghz and then you’d be able to measure below 0.1 nanoseconds… but still to get attosecond perhaps you can do something like at least detect the *order* in which things occur and then just smear them out uniformly in the 0.1 or 0.01 nanosecond timescale where your clock actually ticks and infer their average rate. I’m suspecting that’s more like what’s going on.
In any case, it’s a technological issue not a *definitional* issue. This whole conversation started when you said “the definition of a second is only accurate to a femtosecond”
definitions are infinitely accurate *by definition*.
The lack of actual reasoning ability by LLMs like GPT is clear when they trip up on simple logic puzzles (especially hilarious are the bullshit explanations given by the AI for why its nonsense answer is correct – see the “therefore” sentence below), e.g.:
Bob has 2 brothers. His sister Sue has 2 sisters. How many brothers does Sue have, and how many sisters does Bob have?
ChatGPT (3.5)
Bob has 2 brothers, so he himself is one of the brothers, meaning there are 3 brothers altogether (including Bob). Sue has 2 sisters, so she herself is one of the sisters, meaning there are 3 sisters altogether (including Sue). Therefore, Bob has 2 brothers and 0 sisters, while Sue has 0 brothers and 2 sisters.