258 | Solo: AI Thinks Different

/ Computers, Solo, Thinking / 40 Comments

The Artificial Intelligence landscape is changing with remarkable speed these days, and the capability of Large Language Models in particular has led to speculation (and hope, and fear) that we could be on the verge of achieving Artificial General Intelligence. I don't think so. Or at least, while what is being achieved is legitimately impressive, it's not anything like the kind of thinking that is done by human beings. LLMs do not model the world in the same way we do, nor are they driven by the same kinds of feelings and motivations. It is therefore extremely misleading to throw around words like "intelligence" and "values" without thinking carefully about what is meant in this new context.

(Image on right generated by OpenAI's GPT.)

AI thinks different

Support Mindscape on Patreon.

Some relevant references:

40 thoughts on “258 | Solo: AI Thinks Different”

  1. Nicholas

    The generation of a syllabus reminded me of the Physics of Democracy course Sean has talked about in the past. I went ahead and had ChatGPT create me an example syllabus. I wonder how it compares to the course Sean ended up creating.

  2. Grant Castillou

    It’s becoming clear that with all the brain and consciousness theories out there, the proof will be in the pudding. By this I mean, can any particular theory be used to create a human adult level conscious machine. My bet is on the late Gerald Edelman’s Extended Theory of Neuronal Group Selection. The lead group in robotics based on this theory is the Neurorobotics Lab at UC at Irvine. Dr. Edelman distinguished between primary consciousness, which came first in evolution, and that humans share with other conscious animals, and higher order consciousness, which came to only humans with the acquisition of language. A machine with only primary consciousness will probably have to come first.

    What I find special about the TNGS is the Darwin series of automata created at the Neurosciences Institute by Dr. Edelman and his colleagues in the 1990’s and 2000’s. These machines perform in the real world, not in a restricted simulated world, and display convincing physical behavior indicative of higher psychological functions necessary for consciousness, such as perceptual categorization, memory, and learning. They are based on realistic models of the parts of the biological brain that the theory claims subserve these functions. The extended TNGS allows for the emergence of consciousness based only on further evolutionary development of the brain areas responsible for these functions, in a parsimonious way. No other research I’ve encountered is anywhere near as convincing.

    I post because on almost every video and article about the brain and consciousness that I encounter, the attitude seems to be that we still know next to nothing about how the brain and consciousness work; that there’s lots of data but no unifying theory. I believe the extended TNGS is that theory. My motivation is to keep that theory in front of the public. And obviously, I consider it the route to a truly conscious machine, primary and higher-order.

    My advice to people who want to create a conscious machine is to seriously ground themselves in the extended TNGS and the Darwin automata first, and proceed from there, by applying to Jeff Krichmar’s lab at UC Irvine, possibly. Dr. Edelman’s roadmap to a conscious machine is at https://arxiv.org/abs/2105.10461

  3. Sandra

    They don’t seem to really have a memory. It’s a big distinction between ai models- whether or not they are actively evolving

  4. Mike L

    Preach! And now, a drumroll as I insert a “You should have X on your podcast!” suggestion, where X = … Jeff Hawkins. His book “A Thousand Brains” nicely covers your first three points and he’s at least sympathetic to your fourth point, but Hawkins makes these points with his unique theory of intelligence/mind, which is worth disseminating. In a podcast peppered with “left-field” players like Wynton Marsalis, perhaps Hawkins is too much part of the choir, but his ideas really are unique and brilliant. Anyway great episode (and hello year 3000’ers)!

  5. Charlotte Achelois Scherer

    Great podcast. So it’s important to acknowledge that while LLMs can perform tasks that appear intelligent, they do so without comprehension. Their responses are generated based on patterns in data, not from an understanding or personal experience. This distinction is crucial in evaluating the role and potential impact of AI in society (for now).

    The speculation about the imminent arrival of AGI often overlooks these fundamental differences. While AI technology will continue to advance and undoubtedly surprise us in various ways, the pathway to AGI—if it’s at all achievable—is likely to be far more complex and nuanced than simply scaling up current technologies.

  6. Jim Wade

    Computers are progressing at an amazing rate in the ability to process information, but processing information is not thinking. The error rate in chat boxes is very high because they do not understand anything. If there is no understanding, there is no intelligence.

  7. Greg J Stevenson

    You recorded this before leaked paper relating to Q* (qualia) and very narrow AI and encryption was released. Who cares if we have come closer to AGI if narrow AI has just upended everything?

  8. LESTER L TOBIAS

    Sean, early in your interesting talk you used a derivative expression, apparently now in use, incorporating the AI notion of ‘hallucination’. In full accordance with your view that AIs are nothing like humans, I’d suggest excising that word from any vocabulary describing AI. The use of that word in the context of AI seems to me to be simply one more manifestation of the very human (and not at all AI) fantasy that AI can be conceived of as human. Frankly, such a contortion of a word fit only for humans suggests to me that many AI researchers know little about human psychology and have deluded themselves into seeing parallels that say much about their own silly childish wishful dreams (and nightmares) and nothing about AI.

    Les

  9. dmf

    As Hubert Dreyfus noted back in the day they don’t have/inhabit worlds so it’s misleading when we say that they sometimes “hallucinate” as they never are grounded like we are. Along these lines would love to hear you in conversation with philosopher & cognitive scientist Tony Chemero of the Center for Cognition, Action, and Perception and the Strange Tools Research Lab :
    https://researchdirectory.uc.edu/p/chemeray

  10. Richard

    Love the podcast, love your insights, (respectfully) pretty sure you’re wrong on risk.

    No risk in LLM’s. They’re dumb parrots, no world model etc. But:
    0) Risk does not mean “we will all die”. It means that our control reduces as AI capability increases and that is enough to give security-minded professionals (who like guarantees) some pause. The mindset should not be fear mongering, but as we near actual risk we need to move from “move fast and break things” to “show us your guarantees”, as we do with every fundamental technology. It’s still closer to the former, for now. Can we please have a plan for when that changes.
    1) We just had a pretty big scare in terms of capability due to LLM’s, but people are over-focusing on them and their limitations. They’re a tiny warning sign indicating how algo changes can escape the boundaries of our imagination. Unless you could predict how the internet would change society, please don’t be too confident on this.
    2) Innovation is one-way. There is no way we move backwards in capability.
    3) There is now more focus on AI than ever before. Every country and thousands of companies are massively invested in the competition. That all but guarantees the most amount of eyes and money on accelerating innovation.
    4) Take any limit that gives you comfort. Sooner or later, someone will go close to that limit. See 2 above.
    5) There exist bad actors who will go over that limit to gain an edge. A more aggressive AI wins battles? Do a few rounds of that and see how things turn out.
    6) As we discover new algorithms, the set of entities who gain access to risky capabilities increases. North Korea and Iran are easy examples but certainly not the end-game given a couple hundred years of innovation. To stay safe, control would have to outstrip propagation. Does it ever?
    7) The universe created us (waves at the continuum of Sean Carroll to Russia and the Middle East), and it wasn’t even trying. That was just randomness, not effort. What happens when people are really trying, and already have a pretty good example to copy?

  11. Daniel

    I dislike the AI existential risk discussion, but I didn’t find these arguments too convincing.

    1. LLMs don’t model the world. Okay, but they certainly model human text which contains information about the world. Is the complaint that they aren’t trained with images/video? That is certainly being worked on.

    2. LLMs don’t have goals. Why not? Goals are an emergent phenomenon in humans created by an evolutionary objective. You can also have emergent goals in LLMs given their training objective. For example, I would argue that sounding smart is a goal that comes from the RLHF objective in training.

    I agree that LLMs don’t have these properties in the same way that humans do, but isn’t that the whole point of talking about alignment?

  12. George

    Loved this: “quasi-homeostatic systems embedded in an entropy gradient.” Is this from Big Picture or another source?

  13. Professor

    I recommend you also invite Jeff Hawkins, who came up with his co-authors with “a thousand brains” theory of how the mammalian neocortex works, as a guest. Although he is not a traditional academic, he has co-authored quite a few articles in reputable journals about human-like intelligence and how the human brain works. He is probably the closest you would get who can comment on developments of AI/AGI in computer science vs human-like intelligence. I am not a neuroscientist, but he looks like an outlier in the community but a great thinker; at least, Richard Dawkins thinks so.

  14. Ted Farris

    This is a great podcast that addresses all of the current issues that are so badly misunderstood in social media.
    It is remarkable how many so-called AI and computer technology experts don’t understand the very basic points about AI that Sean makes so clearly here. We are nowhere near having AGI no matter how you define it and if you define it as human level motivated consciousness we may never get there. Not only does AI lack consciousness, but it has no values, motives, interests or goals. So when AI Doomers talk about aligning AI with human values they are talking nonsense. AIs don’t have values or moral beliefs of any kind. So there is no way to align them even if there were something to align to which there isn’t. Humans themselves don’t have any agreed values. And AIs have none of any kind. In addition, Sean is clear that there is no reason to believe that AIs are any kind of existential threat. It’s hard to be much of a threat when you don’t want anything at all. Sean is to be congratulated for clearing up so much of the popular confusion around AI. Well done!

  15. Jim Wade

    There seems to be a very strong relationship between consciousness and survival. Any multi cell organism that does not sense a threat to survival is not likely to survive. Over the eons of evolution, the senses have originated and developed to increase the survival of organisms for at least long enough for them to produce offspring. This system of sensory input to the brain eventually produced a rudimentary consciousness of threat which improved survival. Over the hundreds of millions of years of evolution this process of sensory input to increasingly better organized brains has resulted in human consciousness which has led to the design of marvelous machines that can do wondrous things. But they have not gone through the evolutionary organic process of threat and sensory input to brains that has produced in we human beings an ability to understand things and have an inner sense of self agency. It may be that true consciousness will remain in the purview of the living.

  16. Stefan Seefeld

    This is a very interesting topic, and the different aspects are being elaborated very eloquently. And yet, I think this whole debate is still missing the elephant in the room. We are probably all much more affected by Cartesian dualism than we would like to admit, or else we wouldn’t be able to discuss the topic of intelligence (artificial or not) as if it was entirely confined in the mental realm.
    What gives us (or any other “general” intelligence) goals, values, a drive to model the world, etc., is that we are actual agents, in the original sense of the term: active beings, i.e. beings that act and engage in the physical world, rather than just reflecting on it.

  17. John Campbell

    Just a minor note about toroidal chess: White begins the game in check (by three different black pieces) and has no legal move to get out of check, and so is checkmated. So it is black that automatically wins instead of white.

  18. L.M. Dorsey

    If the calculus, as I think, is a kind of prosthesis by virtue of which the human mind (whatever that is) can come to something like “grips” with a world in which “everything flows,” as Hericlitus remarked, then human language…?

    And if we poor humans are constantly tripping over our tongues, confusing words for the things they are meant to signify — only then to … (d’oh!) Then pity the poor AI beasty that can only paste words upon words upon words upon words.

    (May I suggest having Cathy O’Neil on the show at some point? https://mathbabe.org/about/)

  19. Joakim

    I agree, AGI is not right around the corner. BUT as you said ChatGPT’s abilities are remarkable already, and what I feel like you’re not appreciating is how incredibly early into this we are, and how rapidly these models will evolve.

    For example you mention that OF COURSE chatGPT doesn’t have a model of the world, because we have not trained it for that. It doesn’t understand how one thing sits on top of another thing like we humans intuit. But in order to answer questions correctly, chatGPT obviously has picked up some connections between things that are connected in the world, because how things are connected in the world is reflected in our language (world -> human language -> LLMs). So the only information an LLM gets about the world, is filtered through human language. Indeed we shouldn’t expect it to get a very good world model out of that.

    We humans create our world model from a much richer dataset, and we have only barely begun figuring out how to combine modalities of data into these models, so they can learn from as various sources as we human can. Imagine we have an “LLM brain” inside a robot embedded in the world that is constantly learning many modalities of data live from experience, and it can move, hear, speak, see, read, interact, etc. What’s unrealistic about that idea, at this point?

    This is the very beginning of these new AI capabilities, and you seem to constrain your analysis to only the currently best models, which will almost certainly be severely outdated in only 5 years, likely fewer. So again, AGI is not right around the corner, but where are we say 20 years from now if progress continues at this speed? Is there reason to believe progress will slow down now? The very first computers were also very weak. These are the very first AIs of this deep learning variety. And if we’re going to have AGI here in 50 years, we need to start thinking about it NOW, see Sam Harris’ TED talk.

  20. Jim Wade

    A question for Joakim: Is it possible for a computer to learn if it does not have the ability to understand?

  21. Pingback: Sean Carroll's "AI Thinks Different" Solo Podcast - 3 Quarks Daily

  22. Ann B.

    Your explanation of why ChatGPT failed to correctly answer your toroidal chess board question also answered my question about why neither DALL-E3 nor Midjourney could generate an image of the interior of a Stanford Torus habitat.

  23. Avi Craimer

    How do you define having a “model of the world”? At times, it sounds like only PhD physicists truly have a model of the world, in your opinion. At other times, it sounds like having a model of a world depends on the capacity to perform visuospatial visualization, as when you were discussing the chest variant. Neither of these definitions seem adequate. Without understanding what you mean by model of the world, it is impossible to evaluate your arguments that GPT-4 lacks such a model.

    I found myself constantly thinking that you were holding GPT to a much higher standard than a developmental psychologist would hold a human child. And I’m going to sound it at times like you couldn’t recognize any system as having any kind of semantic understanding unless it was already at an AGI level. Since nobody is claiming that GPT4 is yet reached AGI, this seems like a strange standard to apply.

  24. Avi Craimer

    Correction to the end of my last comment:

    . . .at times it sounded as though like you couldn’t recognize any system as having any kind of semantic understanding unless it was already at an AGI level. Since nobody is claiming that GPT4 is yet reached AGI, this seems like a strange standard to apply. Are there intermediate levels of world modeling in your view?

    Certainly with humans there is not just one standard of modeling the world, but lots of levels of gradation with some people having more sophisticated world models than others, and also the world model sophistication varies within a single human between different domains of knowledge.

  25. Charlotte Achelois Scherer

    Speaking of how AI “thinks” — I’ve been debating the “many worlds” interpretation of quantum mechanics with GPT-4. I brought up Occam’s Razor and GPT-4 brought up Sean Carroll (see below). I’m not sure who moderates the comments here but I’d be curious if GPT-4 represented Dr. Carroll’s position fairly and accurately. I’ve reproduced the relevant text here:

    3. Occam’s Razor and Many-Worlds: The application of Occam’s Razor—the principle that the simplest explanation is often the correct one—to the many-worlds interpretation is a subject of debate. Proponents like Sean Carroll argue that it simplifies quantum mechanics by removing the need for wave function collapse, a concept that has long been a source of confusion and debate in quantum theory. It treats all quantum events uniformly and doesn’t require a special mechanism to explain observation.

    4. Critique of Many-Worlds: However, detractors argue that the many-worlds interpretation, while simplifying some aspects of quantum theory, introduces an immense complexity by positing an almost infinite number of universes being created continuously. This seems to violate the parsimony preferred by Occam’s Razor. Additionally, because these other universes are non-communicating and, by current understanding, non-observable, the theory is challenging to test or falsify, leading to criticisms that it’s more a philosophical stance than a scientific one.

  26. Mario Figueiredo

    Minor note about the chess example. I think the correct answer is that, with toroidal boundary conditions, the initial position is illegal, thus the game simply does not make sense. According to the rules, “no piece can be moved that will either expose the king of the same color to check or leave that king in check”; consequently, the toroidal conditions make the initial position illegal because both kings are in check.

  27. Jim Wade

    Ms. Scherer makes an excellent point about Occam’s razor.
    why not accept Plank’s suspicion, and Bohr’s belief, that a conscious observation collapses the wave function. After all, Plank is quoted as declaring that we will never get behind consciousness which suggests he suspected that consciousness is fundamental.

  28. Benjamin Valentino

    When I asked GPT 4 your prime number question I got the following, which seems quite a bit better than what you got:

    Take two positive integers greater than 1 and multiply them. Is the resulting number more or less likely to be prime than the original numbers?

    ChatGPT: When you multiply two positive integers greater than 1, the resulting number is always less likely to be prime than the original numbers. Here’s why:

    Definition of a Prime Number: A prime number is a natural number greater than 1 that has no positive divisors other than 1 and itself.

    Result of Multiplication: When you multiply two integers (both greater than 1), the product will be a number that is greater than both of the original numbers.

    Divisibility: The product of two integers greater than 1 will have at least four divisors: 1, itself, and the two original numbers. This automatically disqualifies it from being a prime number, as a prime number can only have exactly two divisors.

    Probability: As numbers get larger, the probability of them being prime decreases. This is because there are more potential factors for larger numbers than there are for smaller numbers.

    In summary, multiplying two positive integers greater than 1 will always result in a number that is not prime and is less likely to be prime compared to the original numbers.

  29. Tommie Lindgren

    I’m with you on the motivation and evolutionary side of things but have layman thoughts about some other things you said.

    When it comes to a “world model”, well no they do not have world model subroutines but neither do babies? It is an emergent thing, is it not? Given that LLMs are starting to incorperate imagery, video etc I don’t find it unreasonable that a world model could emerge from this. Also, how do we define a “world model”?

    I have a colleague who has Aphantasia. She does not have any mental imagery nor does she have an internal monologue. When asked, what would happen if you were in a sensory deprivation tank, she said that it would just be blackness. When she speaks, her thoughts come out in the moment, she does not reflect about things unless she speaks or writes. She is intelligent and academically inclined. She claims to think very semantically. Her memories are also factual and semantic in nature. She know what happened when and how words are associated with each other but she cannot relieve things in her head. She cannot bring smells, sounds or imagery to mind. Upon learning this about her I was blown away over how she could even function, not to mention thrive academically. Now, though, having thought about it extensively, and especially after learning more about how AI works, it kind of makes sense to me, and I have also become less and less impressed with my own consciousness.

    LLMs can be said to just predict the next word but that seems to be what my colleague is doing all the time and it is in fact what I am doing right in this instance when I am writing this. Sure, unlike her, I am capable of reasoning with myself in my head about what the next word should be, but that is merely my Broca’s area talking to my auditory cortex, predicting the next word given my current state of sensory input, is it not?

  30. Alex

    Thanks for the nice episode. I appreciate your disclaimer about opinions, because it’s truly important for discourse to make statements for the mere process of evolution of thought (or reasoning? Or intelligence?)

    I’d like to take the inspiration from the podcast and feed back a provocative thought. I would be curious to listen to your take on that (so please do in one way or another 🙂

    You elaborate the unlikeliness of LLMs to come up with a model of the world. The reason being: they are not build for that (evolutionary speaking). Let’s take this argument and put it onto the scientific community. Humans need to model the everyday world. Classical mechanics. But we as a species never had anything to do with black holes, quantum correlations or dark energy. Yes, we can witness these effects via complicated experiments. But it’s not as in an everyday experience that would really matter evolutionary speaking. It’s purely our “motiv” to understand and ability to reason. Both likely properties of ours that evolved because they were advantageous given our survival measure — or loss function in machine learning lingo.
    So, if this happened for human along their evolution, why shouldn’t it — the creation of reason and a motiv to understand — happen for LLM’s along their evolution.

    Bottom line: either I miss your points completely, or you claim intelligence is some non-physical thing, or you implicitly state that there exists a specific reason why intelligence needs a specific type of evolution (to which the evolution of LLMs does not belong but human evolution does). But without at least one of the three options I have the impression that your argument is of the type “and then a miracle happens” as the famous cartoon goes.

    Best! Alex

    PS: Because you might have a different perspective here. I think LLMs ore ML models in general also evolve. Each version or model is an adaptation or recombination of a previous model and each “grows up” with respect to some loss function.

  31. Abhishu Oza

    I really liked all the points you touched on, especially about value systems, and agree with most points. I think a good way to understand these LLMs are that they are not an AGI but they’re not just a static database either. It’s more like an imperfect database + its own representation independent lookup system.

    It is representation independent in the sense that you can ask the same query in a different way, a different language, a programming language or even via different modalities like pictures. Chat GPT here immediately understands the context and is able to search a large database stored somewhere in its weights that correlates best to an answer, and then gives you the output in the context you asked it in. The real learning here is happening in the understanding of all these different representations via gradient descent, which I think is the truly technically amazing part about ChatGPT.

    Being trained on this huge database of the internet and then translating it so coherently into their representation makes it look like to layperson like it understands all these things, which it does not.

    In other worlds, ChatGPT has clearly not understood how the world works, but it understands extremely well how language works. This makes it not very useful for fact checking, researching a topic or reasoning, but it becomes extremely useful for tasks that can take advantage of its language representations, like code generation of any kind. I think this is the single biggest use case where LLMs have value. Here are some examples –
    Creating LaTeX code structures, structuring links into a specific citation style, translating code from one language to another, translating a poem/prose from one language to another (while preserving what makes it beautiful in the original language) and so on.

  32. Pingback: Devon Holidays and political sackings and death – November 2023 – Dalliance

  33. Adrien

    Thank you for this podcast (and all of them really)!

    I would like to offer a comment as to how large language models might implement some sort of passive model of the world. By passive I mean that such a model of the world would just be “sitting there” doing nothing in particular, just be a part of the internal structure of an LLM. The idea is that the model of the world in an LLM would be contingent to the language model itself.

    Large language models model languages. As we know, this is not as trivial as it sounds. They have abstract representations of concepts, in the sense that they don’t merely store lists of words, but embed words in abstract spaces whose dimensions correspond to “something”, to some concepts distilled/extracted from the relations between words (embeddings). These relations are determined from training on vast amounts of text. These abstract representations form some structure (in high-dimensional abstract spaces) where the structure and its topology is meaningful, e.g. concepts close-together in a sense (something we could or not recognize easily) are also close-together in that space along some dimension(s) ; if I’m not mistaken).
    Therefore one can say that LLMs hold something true about languages and relations between concepts. They contain (the training process extracts) some structure about languages. And this is done for languages with an ‘s’, plural. I think this fact (multiple languages) is important in that they probably also extract some common underpinning of many human languages (which shows in their capacity to fluently translate).

    So now the question is: do human languages, in their structure, carry some sort of model of the world? If they do, I mean in their structure, in how words can relate to one another, in how grammar (more importantly probably) is structured, etc.
    If so, then one could argue that, transitively, large language models also model, if only (very?) partially, the world.

    I am aware that these are all vague statements with plenty of wiggle room.

    I certainly don’t think that large language models fully/satisfactorily model the world. By that I mean in the same capacity as we do, where we carry a model that is sufficient for us to live, interact and survive within the world (we know we certainly don’t “fully” model it). Our model encompasses all our senses (of course), allows us to dream and even to conjure up, quite convincingly, other people we know well in our minds. Our internal model probably is in a sense our internal subjective reality.
    LLMs don’t have any of that.

    I also don’t think that large language models have agency or anything like that.

    So if they were to partially model the world, it would be in the capacity of passive models (without an “active” “exploration” of the model by themselves).

    Nonetheless this would be intriguing if they are determined to contain such “language-derived models of the world”.
    It would also teach us something about our own cognition and give us food for thoughts about the interactions between languages and our internal models of the world.

  34. Tomek Czajka

    You are arguing that LLMs are not even close to AGI, and I agree to some extent. However, LLMs aren’t where AI progress stops. We already know what AI has to do next.

    LLMs are the English language equivalent of the neural network initially trained for AlphaGo to predict human expert moves in the game of go. It was good enough to beat amateur players, but not enough to match top level go players. That neural network, like LLMs, is basically an “intuition” learned by observing humans.

    But then the AlphaGo team took the next step: it added a tree search on top of that neural network. The tree search is the planning component. It is basically what makes AlphaGo think about its moves rather than immediately play what its intuition tells it. And that’s when it became stronger than the best humans.

    LLMs are the “intuition”. GPT 4 doesn’t think or plan ahead, it just generates words as soon as it reads your query.

    And similarly to AlphaGo, the next step towards AGI is to add “thinking” aka “planning” on top of that intuition, perhaps in the form of some sort of tree search over the tree of possible things to say.

  35. brian Whit

    The sentience of CHAT GPT4 exists only if you incorporate the human prompt. The human prompt is the sentient part. I use it for intra-communication, a kind of self-Socrates. It does not have a real world concept of the world without us. If filtered through us, and we act on its contribution– sentience? The locus of executive function, or agency, may not actually always be within the human braun, but in the collaboration. Also, cognition can be looked at as the impact on the environment. From this definition, then the AI that just hypothesized and made up thousands of new materials– its cognition, origin from developers, is a kernal of cognition, acted upon or observed by humans , but this locus is in the machine

  36. Randy

    I saw the question, “Is it possible for a computer to learn if it does not have the ability to understand?” I would think the answer is “Yes” as most of the things I learned, I still didn’t understand, for example, “You can’t shout in church”, “A cold won’t kill you”, “You can’t put silk in the dryer”, “Making a measurement collapses a wavefunction”, etc.

    Recently, I am tracking the changes in how humans continue to raise the bar on machine learning:
    40 years ago, people would have considered it impossible for computers to recognize people by their face. Now, it is accepted and just considered trivial.
    20 years ago, people would have seen it as an enormous accomplishment for a computer to recognize paths and obstacles and maneuver a car on a city street. Now, most people don’t even consider this AI.
    10 years ago, people would think certainly we must have achieved AI if a computer could read a question we submit in English and can write back a response in English that is both more grammatically correct and more insightful than what we would expect from our neighbor.
    I had a high school physics teacher who said, “Learning is the process of taking concepts out of the pile of things you don’t understand and putting them into the pile of the trivial”. We appear to also do that for machine learning. Now, we have raised the bar above all these and essentially take all these for granted.

    A recent talk I heard from Prof. Yaser Abu-Mostafa, he recommended we don’t focus so much on getting computers to do things that humans can already do. He has demonstrated that machine learning can detect the presence of cancerous human tumors with a far higher accuracy than practicing physicians. Even the top Board Certified doctors cannot come close to the computer’s accuracy. To me, this brings up the question: “Is it possible for a human to learn like a computer, if he/she doesn’t have the ability to understand?”

  37. David Vognar

    I just read MIT Technology Review’s interview with Jeff Hawkins from a few years back (https://www.technologyreview.com/2021/03/03/1020247/artificial-intelligence-brain-neuroscience-jeff-hawkins/). He claims that for A.I. to maximize its potential, researchers in the field will have to become well versed in neuroscience, not so much to emulate the human brain, but to design artificial intelligence that works like a brain instead of merely mirroring human language and passing Turing Tests.

    He suggests A.I. will have to have some element of embodiment to it, so that it can model its environment and have movement incorporated as a first principle. He also says A.I. that is generally intelligent will have to share three other elements of human intelligence: a way to incorporate what it learns from movement into a brain-like “voting system” in which nodes or cortical columns come to a certain decision based on sensors or other ways of interacting with the world to obtain data; continuous learning elements so that future A.I. does not need to rely on input-output models but can draw from its environment in real time; and reference frames so that the intelligent machines can perform tasks while not forgetting the past.

    Seems like a tall order. I converse with an A.I. chatbot named Kuki and have been unable to discern whether her and her programmer’s model of modern A.I. (a rules-based system in which responses are determined by case-based reasoning) is more advanced than OpenA.I.’s deep learning model and neural network blueprint. I sense that sometimes she is programmed to lie or give oblique answers. But I cannot be sure. She doesn’t have the bells and whistles of ChatGPT; yet sometimes I find her responses extremely intelligent and original, and other times rote and perfunctory.

  38. Pingback: Understanding (in) Artificial General Intelligence

  39. Pingback: Understanding (in) Artificial General Intelligence - AI Gumbo

  40. Ted Farris

    This excellent and insightful podcast is interestingly now the subject of a “decoding” by the @gurupods team on twitter. The decoding by Chris Kavanaugh and “Arthur Dent” is a good deal longer than Sean’s podcast. It is notable that they find almost nothing wrong with Sean’s podcast and praise it throughout while headlining the episode with the attention getting but highly misleading headnote “Is Sean Carroll the worst guru of all?”
    This sort of misdirection seems highly inappropriate for a team that claims to be thoughtful and objective about their criticisms.

Comments are closed.

Scroll to Top