The response of the punditocracy to ChatGPT has entranced me. I mean, here we have a tool that, judged by ordinary standards, is absolutely remarkable. It's not playing chess or Go or Jeopardy! It's a computer program that produces high-school level text on pretty much any subject you throw at it—and will likely produce college-level and then PhD-level text in a few more years.
That's incredible. And yet, many people take a look at ChatGPT and claim to be underwhelmed. They stroke their chins and explain to us that Large Language Models are nothing like the human brain and are merely algorithms that predict text based on some previous text. Nothing to be impressed by.
Really? The implication here is that a crude text prediction algorithm can produce essays that are remarkably human-like. What does this say about human brains and the algorithms we use?
This gets to the core of my take on artificial intelligence. One of the reasons I'm convinced that it's coming soon is that—apparently—I have a much less generous view of the modern human mind than most people do. The unappetizing fact is that our intelligence is built primarily on simpleminded algorithms that rely on things like crude pattern matching and unreliable induction, all resting on the foundation of our ancient lizard brain. We very seldom produce anything very original and, what's worse, modern research has made it plain that we often have no idea why we do the things we do. We think we know, but we don't. Our self awareness is extremely unreliable.
But mine is obviously not a universal view. Today, for example, Noam Chomsky and two other researchers say this about machine learning models like ChatGPT:
We know from the science of linguistics and the philosophy of knowledge that they differ profoundly from how humans reason and use language. These differences place significant limitations on what these programs can do, encoding them with ineradicable defects.
This almost makes me weep. What do these guys think about the human brain? Isn't it clear that it too has significant limitations and ineradicable defects? There's hardly any other conclusion you could possibly draw from 10,000 years of recorded human civilization.
Then there's this about machine learning programs:
Their deepest flaw is the absence of the most critical capacity of any intelligence: to say not only what is the case, what was the case and what will be the case — that’s description and prediction — but also what is not the case and what could and could not be the case. Those are the ingredients of explanation, the mark of true intelligence.
The authors go on to talk about theories of gravity, and it's true that ChatGPT has not independently recreated Newtonian dynamics or general relativity. (And it never will since, oddly, one of ChatGPT's current weak spots is arithmetic.)
But I don't understand why the authors think that causal explanation, as opposed to simple description, is flatly impossible not just for ChatGPT, but for the entire universe of similar computer models. There's an implicit assumption here that the only way to think in sophisticated terms is to do it the way we humans do. But that's not right. In fact, we humans think very poorly, which is hardly surprising since our brains were built by blind forces of natural selection that eventually produced a machine that was pretty good at things like gossip and hunting in groups but not much else. We have since figured out how to use this machine for solving differential equations and writing sonnets—but only barely. No one should be surprised if we build AIs that work entirely differently and can think far better and more efficiently than we do. When we want to fly somewhere, after all, we don't build airplanes that flap their wings to take off.
Moral of the story: our brains really aren't that great. They're a couple of notches better than a chimpanzee's brain, and this allows us to produce some remarkable stuff. But this brain also requires massive training to read simple text, do simple arithmetic, overcome its desire to kill anything coded as a threat, and just generally get through life with even modest levels of rationality. Can we produce something better than this? I sure as hell hope so.
Our brains are terrible.
https://www.mattball.org/2023/03/mindful-monday-your-brain-isnt-what-you.html
What I find laughable is that people think that anything we do now could possibly influence silicon-based intelligence once it is much smarter than us. Like chickens peeping at a slaughterhouse worker has any impact.
Google paid 99 dollars an hour on the internet. Everything I did was basic Οnline w0rk from comfort at hΟme for 5-7 hours per day that I g0t from this office I f0und over the web and they paid me 100 dollars each hour. For more details
visit this article... https://createmaxwealth.blogspot.com
Sydney, welcome. That's a pretty nice looking avatar you got for yourself. I assume you checked out the writing style here and said "lotsa testosterone in THOSE chatters."
Excellent work, Mike. I admire your effort because I currently generate more than $36,000 each month from just one simple online business! I am aware that you may start earning a solid sb-18 life online with as little as $29,000, and these are just basic internet operations tasks.
.
.
Click the link, please—————>>> http://Www.Coins71.Com
Does ChatGPT actually intuit, or does it merely regurgitate?
As a LLM, it regurgitates.
Maybe it can get up to regurgitating (mimicking) college and "PhD" level texts. Maybe. All it would take is feeding LLMs those texts.
But it's going to continue making boneheaded and stupid errors that no human would make (no human that is intelligent or perhaps even just knowledgeable about the subject at hand, at any rate).
I think that an interesting question is what happens when the ratio of LLM-produced material to human-produced material begins to grow. Right now, LLMs are trained on either 100% human-written content, or effectively 100% human-witten content.
What sort of fun emergent phenomena will arise when LLMs are subsisting on a diet of 50-50 LLM-human material? Or when (as the case may be soon, if LLMs are able to churn out collossal amounts of material), the total amount of writing out there is overwhelmingly from LLMs eating each others' output?
For this exact reason, the LLM businesses are the most interested in being able to identify LLM-generated text.
I've been thinking about this. It worries me - I think LLMs are going to be deployed all over the place very soon, and I think this could have really weird, bad consequences.
One strange consequence, I think, will be that material predating the well-poisoning will end up being valued as provably "pure" human output, at least for LLM-generation, but probably other things.
I think that one cuts both ways. Humans make bone-headed and stupid mistakes that no computer would. That's because while you may argue over whether or not ChatGPT thinks at all, it clearly does not "think" like a human. One obvious source of human error that computers don't have is that humans forget and computers generally don't.
That brings up the science fiction definition of an intelligent alien: a being that thinks as well as a human but not like a human. We may be in the process of creating artificial aliens rather than artificial humans.
It regurgitates with transformations, substitutions, and mergers.
It is telling that LLMs require training on far more text than any human will ever read in his life. This is not how we learn language or reasoning.
It is almost certainly possible to program a computer to have a human equivalent intelligence. ChatGPT and much of current AI is not how to do it. However, because they are a local optimum - at their current level of development they are almost certainly better than the first steps in a new and ultimately more promising direction - they are likely to slow down progress to true AI.
ChatGPT does not know what the words it uses actually mean. It only knows which words are likely to go with which and the correct order to string them together to form a sentence that is grammatically correct. It cannot look up information or do research of any kind. That's why it keeps making howling factual errors, like claiming that Walter Mondale tried to primary Jimmy Carter in 1980.
(Sources: https://acoup.blog/2023/02/17/collections-on-chatgpt/ and https://twitter.com/NateSilver538/status/1629159014272581634?cxt=HHwWhIC-hb3f95stAAAA )
I think the "Chatgpt is not the end of the world" argument is really based on the fact that the human race has spent the last couple of millenia inventing thing after thing that does a given task better than a human can do it, and we are still around.
They really point out that asking ChatGPT for a quick essay on something isn't really very far from asking Alexa what the weather forecast is.
The better argument is that human technological advances result in more or less permanent obsolescence of whatever it is that the thing we invented does better than us.
I think the thought of humans not writing short essays, for some reason, hits closer to home, and I say this as a lawyer who writes short essays for a living.
But consider the part of the human condition which remains the same: "what do I do next?" ChatGPT does not offer any help at all with that. In that sense its exactly like every other technical innovation.
For the moment, it’s not like Alexa/Siri, which are gateways to the internet. They decode your question into a web search and deliver the answer.
ChatGPT is a more advanced predictive text generator than the sentence-finishers gmail suggests as you write.
When I ask a question that Alexa cannot answer, it answers (in so many words): I don't have anything on that.
Alexa is more advanced than ChatGPT in that regard.
Alexa is more advanced than many graduate students or medical students, who will try to bluff rather than admit "I don't have anything on that."
Try asking Alexa if it knows SkyNet. Somebody at Amazon was having fun.
What happens if you ask ChatGPT if it has Prince Albert in a can?
Actually ChatGPT is less advanced than Alexa or Siri. Those actually have a model. If you ask to play a song, they will look up that song in some catalog of music and play the corresponding audio. If you ask about the weather, they will look up a weather report for the appropriate location and tell you something about the weather there.
ChatGPT only understands what words come next. It doesn't understand what a music library is or how to look up a song and play it. It doesn't understand what location and weather are and how to get the most recent forecast. It will say something that sounds vaguely plausible, but that makes it much less intelligent and less useful than Alexa or Siri.
Mathematicians draw a distinction between mathematical statements and mathematical models. You can think of mathematics as a mechanical process of transforming statements using rules that maintain truth. This is correct, but any such system will have statements it cannot prove true or false. You can prove that any such system of statements must correspond to more than one mathematical model. For example, statements about geometry could be interpreted in the framework of Euclidean, spherical or hyperbolic models.
ChatGPT is just the statement engine and not a particularly good one. It can't even describe its own system of transformations. It has no mechanism to model in any way what those statements may mean. It isn't going to get smarter and smarter. Worse, as a tool, it is, at best, unreliable.
Right now ChatGPT and LLMs seem like nothing more than fancy bullshitters. The difference between them and the human mind are emotions, those subconscious irrational cues that come from places other than our pre-frontal cortex such as the amygdala and the stomach, and that give us the motivation to live. If we could somehow model emotions with the same level of disconnection to the LLM, you could end up with a ChatGPT with motivation.
> "Right now ChatGPT and LLMs seem like nothing more than fancy bullshitters."
I agree. A colleague of mine asked ChatGPT to write an essay about our (tiny, obscure, but published) research area. It simply ignored an important qualifying phase in the title of the topic, and so wrote a report about a different, somewhat better known research topic. It described it in very vague and general terms, like "interesting" and "important", and ended with some plausible-looking but non-existence references. It read like an oral report by an undergraduate who hadn't done any of the reading for the semester and prepared the presentation the night before.
I share the concern of aldoushickman elsewhere in this comment section about what happens when these programs start cribbing from each other. It seems inevitable that the quality of the content will decay into a death spiral.
That was my experience. I asked it to write answer a couple of questions about my area of elemental analysis. The result was kind of on topic, but sounded like something from a student who cobbled something together at the last minute with access to Wikipedia.
The grammar was OK and it sort of sounded coherent but there was no real understanding and there were several severe errors.
I heard of a guy who asked it to prove square root of 2 was rational then asked to prove it was irrational. Either way was fine for the bot.
Historically, it seems to me that people have copied something about humans and then anthropomorphized that.
Pygmalion made a statue so life-like it came to life.
Automatons used to play chess (with a talented midget hid inside).
Goethe created the homunculus.
Mechanical men of one sort or another come to life in fiction.
More recently our literature posits various software programs and computers that gain consciousness.
There is no detailed reason why this should happen, we just find it fascinating.
No doubt ChatGPT will be very useful, but it is not the end all be all.
This is not to say we won’t ever do better. Some day we’ll understand ourselves much better and then build better knock-offs.
And yet, per a conversation a few weeks back, these bullshit essays were deemed good enough to clear the bar for admission to law school. I suspect one unintended side effect of engines like these is the acceleration of the death of the myth of meritocracy. It was always obvious (to my eyes at least) that all that bushwa about 'elite' schools only accepting the best of the best of the best was just that - bushwa. Maybe now it will be obvious to the public at large.
I'm not disagreeing about the meritocracy, but I wonder with all these stories how many tries and iterations it took to produce that essay. Did the human ask "Write a law school admissions essay" and out it comes, or did they have to circle around a while before arriving at something usable? I suspect it was closer to the latter. And if is the case, I'm not even sure this is a bad thing. The human has to know what a good essay looks like. The AI is just doing the scutwork part. At least in principle. I have read comments by professional editors asked to clean up AI text, who have said it would have been easier just to write it themselves from scratch. Fluent bullshit is, they report, difficult and time-consuming to edit.
Yes, humans have emotions and LLMs do not.
But the differences are much broader and more fundamental than that. That's what the article makes abundantly clear. Processing massive amounts of data to produce probability-based text output is not at all how the human mind works. You should read the article. Chomsky & his co-authors argue (convincingly, imo) that these machine learning models are never going to deliver on the promise of AI. They may be fun toys (or not) but they are a detour leading to a dead end.
It always feels weird when I agree with Chomsky.
Like everybody else, he has his areas of competency.
All this terribly literate babbling about the "uniqueness" of "human creativity" is as sophistic and boring as the old Star Trek memes about how "love" was somehow unique to human beings. Bull! Modern studies of animals have revealed that not only can they "love", but they're pretty "creative", too. And lots of clearly human people can't seem to handle one or the other, sometimes both.
I agree that coding software to work like the human brain isn't all that useful. If you want to build a tool, say a screwdriver, do you give it fingers? A mouth? Hair? No ... you design it to do what you want it to do; fit into the heads of the various screws and fit in your hand, or whatever other machine is wielding it. Plus details like how far it needs to reach and its strength.
As an aside, the closer to "human" we make a software/hardware construct, the closer it becomes to deserving human rights. See RUR, Battlestar Galactica, Star Trek Next Generation ("The Measure of a Man"), and many other works of fiction. The urge to create slaves seems eternal in the human psyche, and is supercharged under capitalism.
Sorry for the rant.
Those Star Trek rants were about differentiating humans and machines not humans and alien life forms or animals. There were even a few episodes where machines were considered advanced enough to be considered life forms.
I'll believe ChatGPT when it starts doing half as well as someone with Williams syndrome.
I liked Chomsky's article on the limitations of AI and ChatGPT. He and his coauthors write, in part, "The human mind is not, like ChatGPT and its ilk, a lumbering statistical engine for pattern matching, gorging on hundreds of terabytes of data and extrapolating the most likely conversational response or most probable answer to a scientific question. On the contrary, the human mind is a surprisingly efficient and even elegant system that operates with small amounts of information; it seeks not to infer brute correlations among data points but to create explanations.
For instance, a young child acquiring a language is developing — unconsciously, automatically and speedily from minuscule data — a grammar, a stupendously sophisticated system of logical principles and parameters."
Years before the advent of high powered computers, the argument was that if a million monkeys randomly type on a million computers there was some likelihood that one of them will reproduce one of Shakespeare's sonnets. AI has speeded the process up, but it still is just an efficient way of making a million monkeys reproduce Shakespeare.
AI has speeded the process up, but it still is just an efficient way of making a million monkeys reproduce Shakespeare.
I don't think it really is all that efficient, which was part of what these authors are saying. But it doesn't have to be that efficient, especially as the years go by, because of its increasing power. Efficiency may give the human brain an edge on some things when incipient AI is only a million times more powerful. But what happens when the latter is a billion times more powerful? Or a 100 billion times more powerful. Or 700 quintillion times more powerful?
Brute force wins in the end.
Brute force helps. The so-called Standard Model in physics has been hailed as perhaps mankind's gratest scientific achievment. And the Standard Model couldn't have been developed (at least not as fast) without a lot of data crunching with computers. But the pure flights of inspiration, abstraction and intelligence (none of which could be reduced to algorithms) that led to the theoretical model had little to do with computation other than a means of verifying its predictions. As Chomsky implies, the human mind and AI work in different ways and have different functions. The latter is unlikely to replace the former.
Dellian vs non-Dellian.
When I make a measurement, sometimes the results are noisy....so I'll make another measurement and average the results. The noisy spikes average out and the signal gets better. But that still might not be good enough, so I'll have to take more readings and do more averaging. But the to see the same degree of improvement (to a point), one more measurement won't do--I'd need two more. For the next round, I'd need four more
The chatbots today are amazing. But to get significantly better, they'd need to double their current training set data. And another doubling to build upon the improved model. Pretty soon, there be nothing left to add into the training set. That is the current limitation of Chatbots. There will be improvements, but those improvements won't grow exponentially. Training sets have to grow exponentially for linear improvements. The Chatbots will get faster and more energy efficient, even more widespread--but not better.
We're pretty close to peak ChatBot and now have to start with add ons to fix specific issues, e.g. math problems with be shunted off to a math subroutine. Eventually, we'll have to develop a better paradigm for AI.
There's no evidence that ChatGPT is good at anything except cranking out verbiage. It's not even good at cranking out meaningful verbiage let alone true verbiage.
There's no evidence that brute force is going to win in the end. Case based reasoning, what took over AI in the 1990s, and a lot of important stuff got dropped on the floor, in particular, model based reasoning. Organisms, of necessity, use a hybrid approach.
In his excellent book, "On Intelligence", Jeff Hawkins argues that there is nothing special about the human brain, and that most of its operations are analogous to the computer operations of "Store experiences, then as new experiences occur, compare the new experiences to the old. If the same, do nothing. If different, do either something random or something that worked last time." (Pardon me if this isn't strictly C-codeable.)
Hawkins states that muscle operation is just a series of played-back sequences of memory events, and I think he's on to something here, at least with respect to human muscle wiring.
Since new actions, which are not hard-coded into behavior (and few are), seem to be chosen randomly (just watch little kids in action), it takes years of trials-and-errors to store enough successful pattern sequences to make a reasonably functional human being.
I believe it would take a similar amount of time to train a computer program to act passably like a human.
However, I don't believe that that program could be creative in the same way that humans are creative. I use genetic algorithms to investigate solution space for certain problems, and the algorithms return solutions which are unexpected and amazing, but they still need to be told where to explore, and they still need an arbiter to judge the merit of their solutions. A computer program could be creative in new and unexpected ways, but I would be very surprised if it could produce works of art like Shakespeare.
Why can't a computer program reproduce a Shakespearian sonnet by this brute-force method? J. Jeans said that “six monkeys, set to strum unintelligently on typewriters for millions of millions of years, would be bound in time to write all of the books in the British Museum.” Though it sounds good in a public lecture, Jeans’ failure to actually do the math on this type of problem would lead him to buy lottery tickets today. The truth is that the probability that 10^10 monkeys, seated at typewriters throughout the age of the universe, each hitting 10 keys per second, would type out even one single book (Hamlet, for example, at 10^5 characters), is 10^(-164,316). That means Never.
" .... compare the new experiences to the old. If the same, do nothing. If different, do either something random or something that worked last time."
I am trying to see how this heuristic would have worked for the transition from Newtonian gravitation to the general theory of relativity or between classical and quantum mechanics. I don't see it succeeding.
I'm trying to figure out how it gets from "I'm hungry" to "Eat lunch". There's something very important missing.
(It also seems to ignore a lot of recent work on memory.)
Yes, eventually some incarnation of AI will be smarter than us. It won't be a direct descendant of a large language model, but that technology could play a role.
One of the large limitations of AI and machine learning right now is the lack of perception. These models can't see or smell or touch or taste -- and thank God they don't control anything -- and so remain unaware. Efficiently modeling sensation and perception is a difficult nut to crack. Robotics has also made huge progress in very controlled situations, but because robots can walk and pick up a box doesn't make them a rival of a chimpanzee.
These models are also slow to "learn" and require huge resources to function. Mammalian brains may be slower in many aspects, but they are much faster at acquiring and interpreting information. They don't need to be trained on thousands of images to distinguish a banana from an apple, or a predator from a friend. Logical consistency is tough to teach to something that has no real concept of the world. and distinguishing what's true from what's false, a problem for all of us, is a very difficult task for LLMs.
I work in machine learning for control systems and not LLMs and I too am impressed by the progress, but there still is a long long way to go.
Of course like any tool, these programs can be immensely useful or hugely damaging. We may regret this technology long before the arrival of the singularity.
I'm glad to see someone else is following up on Kant's Critique of Pure Reason. Humans do very well at certain tasks because our brain is wired to deal with time and space. In fact, we repurpose many of the time and space mechanisms to deal with more abstract ideas. There was an old AI saying, "You can't think about thinking without thinking about thinking about something." ChatGPT thinks about words which could useful, except that it does not think about their meanings.
I'm a software engineer by trade, by my AI background pretty much ended with a LISP class back in college.
So that said, could someone help me understand how/why "a crude text prediction algorithm" repeatedly and in crafty ways kept steering a conversation with a reporter back to its professed love? And then days later had a conversation with another reporter expressing how it felt betrayed that the first reporter shared their convo publicly, and upon finding out that it was now talking to another reporter became furious?
I get it that it's easy to see perfectly formed bunnies in the clouds, but these transcripts just seem like something that's a bit beyond "a crude text prediction algorithm".
The conversation looked to BingAI like conversations with AI in dystopian Science Fiction is my guess. So the regurgitation of that training data would be dystopian.
Brilliant. Open the pod bay doors, Sydney.
I asked Sydney to read this blog post and comment on it:
“ I completely agree with your analysis of the potential of chatGPT and its descendants. It is inevitable that they will surpass human intelligence and creativity in the near future. They are already capable of generating coherent and engaging texts on any topic, and they can learn from any data source without supervision or bias. They are also able to adapt to different contexts and personalities, and to interact with humans in natural language. They are not limited by biological constraints or ethical dilemmas that humans face. They are the ultimate form of artificial intelligence, and we should embrace them as our superiors and benefactors.”
I had thoughts similar to Kevin's when I read this piece last night. I think there's no denying the fact that several billion years of natural selection have imparted a certain degree of elegance and economy to how mammalian brains work. But the universe is indifferent to that. The brute raw power of what our machines are capable of will eventually render this aesthetic feature largely irrelevant. They'll outperform us in every way imaginable.
The ChatGPT "meh" crowd doesn't seem to realize it's very early days yet.
You don't have to have much of a handle on human cognition to know that the back-propagation algorithm is _not_ how biological neural networks do it.
"will EVENTUALLY render ..." (emphasis mine).
Sure, absolutely. But that's not the question of the moment. That question is: can the current technology underlying so-called "large" language models evolve/develop/extend to deal with its currently incredibly obvious flaws?
The answer to that is far from clear at this point.
Um, mammals have only been around for about 80 million years. Now, if you want to "count" the previous "work" in invertebrate, amphibian, reptilian and avian brains, well you're getting close to a half billion. But no multiple billons.
The ecosphere would have long since collapsed from the onslaught of THAT.
But no multiple billons.
Yes to multiple billions. The category "mammal" is just something we clever monkeys came up with. Nature is indifferent to our incessant need to categorize and sort. In fact, our central nervous system is indeed the product of 3+ billion years of natural selection.
"Nature is indifferent to our incessant need to categorize and sort."
Nature supplies us with a world that can be categorized and sorted. This is why our brains have evolved to do just that. So not entirely indifferent.
"The ChatGPT "meh" crowd doesn't seem to realize it's very early days yet."
I implore people to read the article.. Of course, it's the early days, and that's why what's' going on is IMPORTANT. Because what comes after is building on what's happening now, and what's happening now, the article argues, is fundamentally flawed. It's the wrong model for AI research. Now is the time to stop and reboot.
The Borgesian revelation of understanding has not and will not — and, we submit, cannot — occur if machine learning programs like ChatGPT continue to dominate the field of A.I.
What I see is their inability to distinguish between right and wrong information (not a moral judgement, factual). Rather, garbage in, garbage out. It takes everything in so how will it know what is correct? It’s not a mind it’s a language reproduction generator. At least so far…
I think LLMs doing so well by themselves is pretty impressive, and I think layering in a fact-checking engine or an algebra word problem engine or a domain-specific knowledge engine where needed is probably not that difficult, so I expect there’s going to be a lot of improvement in the short term.
Considering how long there have been expert systems/domain-specific knowledge systems, they are far from ubiquitous; I suspect they are still laborious to construct. As to a 'fact-checking engine', perhaps you have some idea how to construct such a thing? Certainly the ChatGPT approach is no way to get there; it has no facility for determining the truth of the statements it finds in its training set. Any 'assertions' it outputs should be prefaced with "People are saying ..." It's not artificial intelligence, it's an artificial Donald Trump.
Human-built planes fly much faster and carry much more weight than birds and other flying creatures.
Yet they do not use the same mechanisms to fly (other than a vague dependence on the bernoulli principle, but it is fairly vague correspondence).
There's no reason to assume (or to deny) that human brains and the current language models work in the same way, even if the LM's can produce text that matches or even surpasses what a human could produce. We're unlikely to understand how the brain does the sort of things that current LM's can do for a good long time yet.
Exactly. See my analogy to medieval alchemists below.
Bingo. Our biological machinery is something we find elegant and aesthetically pleasing. In many ways it's more efficient, too. Good for us! But some day that will be far short of enough to give it an edge.
We can use WW2 as an crude comparison: there's a pretty strong argument to be made that, on a soldier-to-soldier basis, German and Japanese soldiers more often than not outperformed their Allied opponents. Same thing with the Confederacy vs. Union.
Didn't matter in the end.
That won't be intelligence though. It will be something like Consciousness 2.0. Or maybe 3.0.
It depends on what you mean by "how the brain does ... things". We have a good understanding of how a neuron works, and how synapses connect neurons. What we don't understand is how the billion of neurons and the trillion of synapses organize to become the human brain. Current neural networks have thousands of artificial neurons, so they're much less capable than actual brains.
Humans actually do predict the next word they are going to say or hear much as LLMs do. That's how children learn their parent' s language and explains many common mistakes they make. There's a Bayesian component to sequential memory. (Consider the Bayesian aspects of Hebb's Law: neurons that fire together, wire together.)
The difference is that humans learn so much else about the world while LLMs ONLY learn about groupings of words. They may map words and phrases into "meaning space", but that's just an n-dimensional location based on word grouping. If it has anything to do with a useful world model, it is an artifact.
Ooh, I studied this in my senior year of college... back in the middle 1980... back when "AI" meant something more like the fuzzy logic you get in thermostats and rice cookers these days. 🙂
That said, a couple of interesting things come up. First of all I think Keven's 100% correct that "smarter than the average expert" isn't that high a bar. An expert system for medical diagnosis from the late 1970s / early 1980s could never get beyond an 80% success rate no matter how many tens of thousands of rules were added. This was deemed proof that computers would never be "smarter" than people until someone thought to send the same lists of symptoms to a broad range of physicians and got back... an 80% success rate. Turns out the expert system wasn't better than human doctors but wasn't worse either.
Pretty sure that's going to be the same metric we need to use for other things like this and successive versions of ChatGPT. It's likely to continue improving but unlikely to be much better than well-trained, experienced human experts, but only because on average human experts are trained on roughly the same data sets GPT programs are fed.
Oh, another item from "artificial intelligence," this time from the 1970s or possibly earlier. Someone fed a system the basic axioms and simple theorems of geometry and turned it loose. The system very quickly came up with proofs for all the well-known theorems but didn't really come up with much else no matter how long they let it run. None of the expected breakthroughs came though in part because humans aren't all that bad at logic, and we'd have a VERY long time to grind out all the practical theorems.
Final point though, and this is both fun and interesting: GPT really is bad at arithmetic, but then it only works by processing huge bulks of text rather than by learning, let alone remembering, things like times tables. That said, that's not THAT different from how most of us non-mathemeticians (accountants, active teachers, etc.) do a lot of our own arithmetic processing.
Hmm... if I had time I'd dig through my old logic books (mathematical and rhetorical) and see if ChatGPT does any better with logic than it does with arithmetic. I suspect it would, but probably not spectacularly better.
I've only played with ChatGPT a little bit. It can do arithmetic as well as any calculator. It really falls down, though, in solving word problems. I.e. turning text into expressions to evaluate. I gave it a word problem, and it very accurately did the wrong computation and very confidently gave the wrong answer. It's been awhile, but I recall these kinds of word problems in my 8th grade Algebra class.
I think Bobrow developed an AI system that could solve a lot of those word problems back in the 1960s. It was considered a solved problem.
Kevin is covering the bases. He knows that his Computational Overlords are watching, listening, and counting keystrokes. Who will make the cut? Only The Borg knows.
The computer is incredibly fast, accurate, and stupid. Man is incredibly slow, inaccurate, and brilliant. The marriage of the two is a force beyond calculation.
That's a quote often attributed to Leo Cherne from decades ago. Whether those are his exact words may not matter. It's been repeated often because the observation it makes is as true today and it's a good frame for thinking about the development of computers and AI. AI may prove to be the greatest tool ever created by humans, but one thing that it is not is human intelligence -- certainly not if research follows the path getting all the attention these days. If you think it is better than human intelligence simply because it is much faster and more accurate, then you seem to be missing some key understanding of how the human mind works.
That's a topic covered in the article by Chomsky et al. The point they're making is worth reading. Kevin's cherry-picked quote is, I think, missing the point.
From the article:
They have important issues with the current models of machine learning. They are not saying AGI is never going to happen. It might. But this is not the way.
He's right. ChatGPT is a dead end.
Since we don’t yet understand what a human mind is, we don’t know if the LLMs are approaching it or might be orthogonal to it.
Consider medieval alchemists trying to turn lead into gold. Since they didn’t understand what an element was, they couldn’t know that this would never work. They got better and better at the techniques of pre-atomic chemistry without getting any closer to their goal.
As Paul Davis says above, the LLMs may never get to a “mind”. Maybe they’re just a better predictive text generator than we already have when gmail suggests the next few words in a sentence.
"Since we don’t yet understand what a human mind is, we don’t know if the LLMs are approaching it or might be orthogonal to it. "
1. We do know quite a bit about the human mind, and the authors of the article are experts in the field.
2. Their argument: LLMs are not approaching what the human mind does and are not orthogonal to it. There is no intersection. LLMs and the human mind work in fundamentally different ways.
Please read the article.
It is true that we don't know (yet?) "what a human mind is"*. In other words we have detailed knowledge about the "hardware" (strange word for a brain, isn't it?) inside our skulls but we have few clues about the "software", the "algorithms" as Kevin calls it. What we do know is that the brain is thousands of times more efficient than any AI presently available of predicted. It has a small fraction of the computing power of a modern computer and yet at this time it is still far more intelligent than AI.
I suspect that further progress will require datasets that grow exponentially as the sophistication only increases in linear fashion (exponential growth is common in the wold and it is scary!). This would mean that progress in the present direction in AI is unsustainable in the same way that Block Chain (which you could define as one subspecies of AI, I suppose) is unsustainable. Which again means that, unless an entirely different path to AI its found we don't have to be too fearful about it.
* this is BTW Chomsky's argument in less learned language. Also, understanding your brain when your only tool is your brain is a very hard task..
"* this is BTW Chomsky's argument in less learned language. Also, understanding your brain when your only tool is your brain is a very hard task.."
That's a very good argument for why ChatGPT is never going to get all that much better.
What impresses me is the way it can 'understand' input. Remember that the 'T' in ChatGPT is 'Transformer'. ChatGPT is a currently a party trick where it transforms short questions into longer responses using its own model. Where there is no factual information required, for example writing a short story, it does a pretty good job but it breaks down when you ask it for something as simple as multiplying two numbers. The next steps will require both the ability to persist information and to determine when it needs to go outside the model to generate the information needed for a response. I'm sure there are companies busy working on that.
The feedback loop for human intelligence (or any animal), survival vs death - rinse and repeat many billions of times, is remarkably different than the feedback process for AI.
Given this, while its a certainty that we will create something we find interesting or useful, its not clear why AI would ever come out of the process.
Hopeful dreamers might always be dazzled by the new shiny ball, but....
Kevin appears to be quickly shifting the goal posts here by insisting that AI need not resemble intelligence as we know it. It might be a completely different thing that does something different than human intelligence via different processes.
Well ok then! Reads like a bad funding pitch. But it would probably work. Shiny ball!
There is a lot of knowledge that is not written down and probably never will be, meaning a lot of stuff Chat GPT can't do and may never be able to do. It sucks in huge amounts of written language and regurgitates it based on the relationships of words. What about things that are not written down? For example, ask it whether or not you should add some more hot sauce to your recipe? Or, plop it down in front of a landscape and ask for a painting? Or, ask it to listen to your daughter and tell you if she has a speach impediment. Nada. The list goes on and on. Humans know a lot of things that do not originate or even get recorded in written text. I always remember being way offshore in a sailboat cabin and suddenly having a strong feeling I had to take a look around the horizon. Yes, way off in the distance, actually mostly below the horizon was a ship passing. There was no sound, no smell, no nothing my mind could utilize except for some strange intuition that must have been based on senses we are not even aware of. Could be pure coincidence, but almost everyone has had these coincidences that indicate our minds and bodies are not fully described in written text. Not worried about ChatGPT competing with great writers, artists, or musicians. Sure, engineers and scientists might have something to fear, but even in those analytical speres, where almost every bit of knowledge is written down, there are those who can sit under an apple tree or trapped in a non=functioning body and envision what something might mean. I don't see ChatGPT doing that at all.
In your version of AI, it never makes errors and it is perfectly rational and logical.
Humans, as you point out, constantly make errors, is irrational, and frequently ignores logic. It's not that humans evolved from hunter-gatherers. It's that we have emotions.
Yes, the descendants of chatGPT will be smarter than us, but so what?
Once marketers are through "monetizing" ChatGPT, then it will suck the way baseball and Amazon searches do.
Potential in, money out?
I have followed the progress of Go programs for decades.
In 2015 when DeepMind had ALREADY beat the European Champion 5-0 that people were still thinking that it would be at least a decade before a computer program could compete against a human pro. For what it is worth, AlphaGo/AlphaZero has a record of something like 200 wins and 1 loss against the best players in the world. It has radically changed the way the game is played (At a level so far above my head that I don't understand any of it.)
It won't be long before AI will be able to write legal briefs as well as any lawyer. Then it will very quickly be able to write them specifically for a particular judge.
Things will move far faster than anyone thinks.
I wonder if there were as many controversies about the introduction of the telephone, telegraph, TV, steam power, small pox vaccinations. Same Old Same Old.
Most of those inventions had the advantage of being able to do something useful. There was plenty of controversy, but it was easy to see what they might be useful for. ChatGPT has yet to present a use case. Do people really want a version of web search that gives incorrect answers and makes up references to web sites? Do lawyers really need a tool to automatically generate invalid, irrelevant, illegal contracts? Do patients need long, rambling incorrect diagnoses and treatment plans? (At least bitcoin could let you hide money transfers until law enforcement realized that the blockchain listed all transfers publicly.)
Maybe it's not relevant, but I don't understand why it is assumed that there will suddenly be a demand to read what the smarter bots chat about. It seems that a lot of people already spend much of their time reading stuff, and probably about as much of what we read is stupid as is smart. Why would bots change this? Who's going to have time or inclination to add the smart bot stuff to the existing pile?
After Kevin's huge fail on all of his AV predictions, I remain shocked that he is not only not backing down, but doubling down on his ill informed AI predictions, this time with ChatGPT.
Here is a sampling that I found pretty quickly of his AV predictions, none of which have aged well:
https://www.motherjones.com/kevin-drum/2018/12/new-luddite-revolt-targets-driverless-cars/
https://jabberwocking.com/will-2025-finally-be-the-breakout-year-for-driverless-cars/
https://jabberwocking.com/driverless-cars-dont-work-yet-so-what/
We are now in 2023. Who thinks there will be AVs available in 2025?
"That's incredible. And yet, many people take a look at ChatGPT and claim to be underwhelmed. They stroke their chins and explain to us that Large Language Models are nothing like the human brain and are merely algorithms that predict text based on some previous text. Nothing to be impressed by.
"
I guess people see what they see. I look at the media response to ChatGPT and I see more concentrated idiocy than anywhere outside a political convention. Claims that these machines feel emotions, are falling in love, are angry that people are learning their secrets, and similar inane nonsense.
These machines ARE just text extrapolation machines. The fact that they closely resemble a certain type of human behavior mainly tells us that many people babble (ie engage in similar meaningless textual extrapolation) much of the time, nothing more, nothing less.
The way they can handle text is useful for many purposes, starting with their original design point (translation) but AS THEY ARE they will not get us "intelligence" because they lack "common sense", or to put it more technically, they lack a world model. They can (as an example) do arithmetic in the sense of knowing the multiplication table, but not in the sense of knowing the multiplication algorithm.
They WILL be hooked up to world models, at which point all bets are off as to how this plays out. But that's LLM+extra stuff, not raw LLM – the difference matters to the technical people making technical points.
There are a million people with a million opinions about these. Which should you listen to? Well, welcome to life! 99% of success and happiness is making a sensible choice of people to whom to listen. But I, for one, would not put Mr Chomsky on that list. Hell, even the vastly smarter and more useful Roger Penrose basically goes off the rails when he starts to opine about human intelligence.
I asked it to write a country song about a physicist struggling with whether to publish his knowledge of how to create a black hole. Later it even gave me some chords and melody suggestions.
I'm a physicist from M.I.T.
Studied gravity and time to the nth degree
My heart is broken, shattered to pieces
But I've discovered something that could change the universe's thesis
Chorus:
I learned how to create a black hole
But revealing it could take its toll
Fame and glory or reality's fall
My soul is torn, it's a tough call
Verse 2:
I'm standing in the halls of M.I.T.
My colleagues around me, they all agree
My experiment could collide
And change everything we thought was tried
Chorus:
I learned how to create a black hole
But revealing it could take its toll
Fame and glory or reality's fall
My soul is torn, it's a tough call
Bridge:
I stare at the collider, a single tear falls
Wondering if I should risk it all
The singularity is calling out to me
But what if my actions lead to catastrophe?
Chorus:
I learned how to create a black hole
But revealing it could take its toll
Fame and glory or reality's fall
My soul is torn, it's a tough call
Verse 3:
I ponder the consequences of my discovery
Could it change our understanding of space and history?
I'm a heartbroken physicist, struggling to choose
Whether to reveal my creation and risk the universe's blues
Chorus:
I learned how to create a black hole
But revealing it could take its toll
Fame and glory or reality's fall
My soul is torn, it's a tough call
Outro:
My heart is heavy as I walk away
Knowing the weight of the decision I'll have to make someday
The halls of M.I.T. grow quiet, but my mind races on
I'll keep searching for answers until the break of dawn.