Debunking the great AI lie | Noam Chomsky, Gary Marcus, Jeremy Kahn

@Mark Thank you for your well thought out response. Regarding AGI I'd love to read up on this; what projects are you referring to? I'm quite jaded on this subject after the repeated snake oil salesman over promising and under delivering. Regarding ChatGPT we can agree to disagree. I think these tools are a distraction at best and dangerous at worst because of statements like, "the disclaimer that follow-up to check facts is important too". What good is it then if it can lie to you? One is much better off learning to ask better questions of the search engines that already exist and learning how to judge sources critically on their own. These two skills are invaluable in the current age.

​​​@Derek Adair agreed with you until your final point. We are a lot closer to an AGI now. Several projects are working at building a cognitive architecture into AI along with some kind of rudimentary value system, such as 'do no harm'. General intelligence has already been demonstrated in systems that can take previous learning to complete entirely new tasks without training. That's more or less the definition of AGI. But I disagree with Chomsky that the models don't contribute beyond being useful. Yes, they don't create new insights on cognition, but they can serve as a kind of Hive Mind to inform about ideas and research. And it can do a reasonable job of this. This can be both informative but also inspiring, because it opens up novel forms of education. Tutors often have little time to deal with questions and students are often shy to ask, but an AI like ChatGPT can fill in as a tutor, answering questions and giving information about debates, theories, and research, with the disclaimer that follow-up to check facts is important too. I've tested ChatGPTs capabilities in several areas where I have high expertise and it's generally quite good. The smarter the questions, the smarter the answers too, so garbage in garbage out still. But if your reasonably smart in how you use it, it you understand even a little of how it works, it can be very good for helping a naive individual discover things about the world, even while the AI itself doesn't really have a model of the world. In that sense, it can be more than just a useful tool, it can actually inspire. In addition, I've used it to exercise my curiosity in fields that I was interested in but never had the time to explore - the problem with books is that too often, the author's wordings raise way more questions than they answer and there is no chance to get clarification. ChatGPT gives a possibility to dig down a bit when ideas are presented. I've enjoyed that immensely.

@Harry Adams I already got the point they're making and I absolutely agree with that assessment. My mistake was (mis-)using the term "false dichotomy". I did not mean the logical fallacy. I meant an erroneous separation of two spheres, like in standard macroeconomic theory, which artificially separates "real" from monetary quantities thereby annihilating the function of money as means of payment. In our case at hand, the information processing needs to be merged into a framework that actually is capable of semantic expression, analysis and novel "ideas" rather than mere stochastic regurgitation of existing ideas, bad and good. I dont think AGI could emerge or be created without either of these. Information processing is integral part of the larger approach as well as the semantics.

The issue is AI has no desire or drive to learn. It won't do anything without being told to. So how can it ever develop into anything more than an algorithm.

@Rob Weaver I see, that's pretty interesting. So I suppose that it's ability to have a good understanding of not only how often one word might follow another, but also like certin patterns and phrasing and relationships that words or phrases might have with one another. Like maybe it could pick up the pattern that if someone begins a sentence with "yeah right" that it's more likely that the next part of the sentence could be written sarcastically as opposed to more genuinely, so the model can use that information in how it responds instead of just using a lookup table. I did something interesting with GPT 3 the other night where I asked it to play a game with me where we would take turns taking 2 different words and putting them together to imply something novel. not only did it create some words that some I couldn't even guess, it also guessed correctly my words, and if it couldn't, often could use a clue I gave it to put the pieces together. I thought that was pretty cool. I've played a game where I've asked it to take an object and rename it to something that describes the object in a unique way without telling me what the object was. It came up with "electron gate" for a lightswitch switch, and "rapid air vibrator" for a speaker which was also pretty cool. It is incredible but you're not suggesting that even GTP4's joke telling ability is actually creativity though are you? I think it's probably in our future but I think it's still just really good at recognizing patterns, but then again, who's to say our brains aren't doing the same thing?

@Alex Czech Generative AI implies that it has the ability to generate unique content. GPT4 can in fact invent a joke that has never been heard in history(ask it to tell a joke around a very specific subject), and it's so good at this that the joke can actually be funny. That's not memorizing.. Creating a look up table AI for statistical 'next word' through "memorizing" is really easy, I've done it, and it's nothing like GPT.

@Rob Weaver what would you say the substantive difference is between them is? Generative AI uses training data that it's "memorized" to then use a set of algorithms to recall it's learned data to do what is essentially predictive word association, and in the case of GTP, to generate seemingly new strings of words based on the new input from the user. I don't necessarily agree with OP here, but for all intents and purpose, I don't know what differentiating between a generative AI and an AI that "memorizes" data is a terribly useful distinction to make, what do you think?

Generative AI is not memorization... It's.. Generative AI..

@lolilol lolilol He was never sharp.

It's clear that he is declining both physically and intellectually. Nevertheless, even at age 95, he is still sharper than the vast majority of us. When he dies, there won't be many intellectuals left in this world who have proved to meaningfully discusss the great societal questions of our time.

Not so much when he was calling for the detainment of those who would not take the jab. Or that the Palestinians should forget their right of return because Israel will not allow it

@blue bay Yup that was it!

@cathalsurfs general AI is what most would consider real.

There is no such thing as "real" AI. Such a concept is an oxymoron and utterly contrived (by humans in their limited capacity).

Oligarchs using flawed ai against mankind scares me the most.

Agreed

So true.

@Maloxi I mean his brain still functions very well even with great age.

@Number Six huh... what ?

the real white mage. amazing. huge respect for him.

He has very good reserve capacity.

@Todd Buckingham Exactly, as discussed and succinctly put by Sartre to pose the necessity of existential consideration, existence _precedes_ essence. Without any critical consideration of meaning, quite simply: garbage in, garbage out. The only "danger", as also discussed, lies in believing it is anything else, that it is "objective" or actually understanding anything ie "intelligent". But of course the _only_ political question is: cui bono? Who benefits? So we can ask "danger _for whom_ ?", which reveals that for some this Mechanical Turk can produce an artifice of an organizing principle of truth/understanding/value ie "god" that "just so happens" to justify the already existing power structure, regardless of the intentionality towards this, as we can see with these psycho silicon valley billionaires and "effective altruists", many of which are unironically labelling themselves as "secular Calvinists". The divine right of "AI" is then but a technological coat of paint over the "divine right of kings/the market/the entrepreneur". _“The ideas of the ruling class are in every epoch the ruling ideas, i.e. the class which is the ruling material force of society, is at the same time its ruling intellectual force. The class which has the means of material production at its disposal, has control at the same time over the means of mental production, so that thereby, generally speaking, the ideas of those who lack the means of mental production are subject to it. The ruling ideas are nothing more than the ideal expression of the dominant material relationships, the dominant material relationships grasped as ideas.”_ - some guy

It can and it can't. Midjourney slays at generating simple albeit impressive images, 'zombie spiderman' 'thor in pixar style' 'human settlers on Mars in the style of Norman Rockwell' but once you begin to describe complicated illustrations with multiple characters in different emotional states with specific likenesses to specific people, wearing specific coloured costumes (and you want consistancy panel to panel) performing specific actions with specific camera/viewpoint angles it struggles, with characters in specific parts of the composition. You can generate those emotions and 'actors' seperately, but still need photoshop to combine them into a complete image. While I guess I should assume Dalle2/Stable Diffusion/Midjourney will get there, after watching this presentation, and after 20k images generated in Midjourney and noticing its sometimes frustrating limitations I do begin to wonder if AI art models lack of language understanding will mean they'll be stuck at 75%. My thought is, the first company to combine Dalle2/Midjourney/Stable style prompting with Nvidia's Canvas like editability/interactivity, will make a much more powerful and efficient tool just by embracing the human brain.

@blackenedblue 😊u😅ihi

@Martin Brock could have, but wouldn't know what it actually means

@噢嗎ØMVЯ All you need to do is give these opposing comments time for them to reveal their own truths. Let me paraphrase: - Chomsky is bad because old - Engineering is good - Chomsky is wrong on AI because I don’t like the vaccine - Here is a quote that ostensibly supports my argument but is little more than banal chest thumping Woof.🤦🏻‍♂️ Interestingly, if everyone disagreeing with you ended up being ChatGPT, my mind would be blown; not because these comments are intelligent, but because they are distinctly human.

@Peter Stafford “As do his political views” I can now infer your real motivation from this comment. Asking in earnest here: can AI do that?

@Peter Stafford Quantitatively, yes.

GTP-4 is much better at understanding the structure of a word (made of letters, has rhymes, has syllables), but it still struggles at some tasks, where it knows the rules but can't reliably follow those rules, but can immediately tell what it did wrong. It just fails at harder problems. Re predicates: Perhaps the Language Model should have some reinforcement learning early on about formal grammar, just like having an English class for 6th graders. Make sure it codified internally all the language structure we want it to, and eliminate incorrect associations, in contrast to just letting it figure it out by example with no formal instruction. Do that at an early stage in training, e.g. 6th grade, before high school and college reading.

no one is unbiased

@Ivan Leon The younger one didn't seem to be very Intelligent/knowledgeable on the subject. The older one seems to be wise at least.

@António Bento both are very intelligent, one is Noam Chomsky, is not fair to be compared with him.

@Mario ArancibiaI agree, and that's how a real scientist should be. The older scientist seemed to be a very good man of science, but the one sitting live didn't seem to be very bright at all. But maybe it was just me.

Sadly, very true. Yet I have heard scientists claim that they can derive as much or more joy from learning where they have been wrong, than when they seemed to be right.

@KOT "Seems" is the key word here and you provided it yourself. Huge differences in dynamics and efficiency are swept under the "bounces electricity around" metaphorical rug to begin with. Neither the I/O nor the dynamics are even remotely similar. Look up any recent model of single-neuron complexity or synaptic dynamics and you'll realize that single neurons and synapses are far more complex respectively than scalar threshold and static linear functions

@Maloxi The concept seems the same, our brains receive input, bounces electricity around neurons and perceives and then creates output, large language model receives input bounces electricity around its parameters perceives and then creates output, what is so different?

@KOT Wildly inaccurate. Even adopting your flawed perspective for a moment, it's obvious that ANN are way too far from biological neurons right now

@Nobody Quite I think the main takeaway is that the only part of the human brain we need to copy to make AI function is the neurons, that's it everything else about the human brain doesn't matter, all the AI needs is neurons, to be honest that's all our brain needs too, people are over complicating it, you do not need to understand the inner workings and everything that goes on with the brain to make AI ,just replicate the neurons and you will be fine. LOL

​@KOT yes. For start, the hardware in brains and computers is quite different. No massive parallelism, clocking, etc. So mimicking the brains is not the best approach. However researching AI can help in a roundabout way to understand human cognition and more. Details: For instance, I did some experiments with stable diffusion and discovered a lot of very interesting things. First of all its akin to "The Treachery of Images" by Magritte (It displays an image of the pipe, not the pipe). Stable diffusion produces an image of the painting, not the painting - a stochastic visual representation of a given image description within the domain of internet images used for learning. If you use a style of speed-art (realistic very fast drawn paintings), like Craig Mullins, you can have interesting results. The art style of Craig Mullins' sketches omits everything that can be easily imagined by the viewer to emphasize main points of interest or composition. For an artist there's a question "how to effectively omit unimportant, but to present enough believability?" Like "how to put several strokes of brush here and there to portray a lake in a distance, but make the viewer understand, that there's a lake there". If you look at a couple of Craig's sketches, it's hard to get the gist of it. But if you have a thousand believable sketches, you have a better chance to imagine the how his style works. I.e. you look at an image of the painting to understand how it is. Like you look at an image of a pipe to understand what is a pipe.

I donno, it kinda sounded to me that a lot of their arguments boiled down to "well, people said self driving cars would exist by now, but they do not, so these new AI systems are basically just that, probably." These guys are like, wow chatgpt hasn't solved world hunger yet so it's basically trash. The type of artificial intelligence that the current models are were not described to the public as anything more than predictive text based generative AI that could be a tool to help people acquire more information more quickly. It uses some pretty advanced tech to achieve this but it's also likely only the beginning of the potential power of a model in the future. There are going to be a lot amazing benefits to humanity these AI system could bring but it's also extremely important that we start thinking about possible serious problems now, and being dismissive that the first few generations of these AI systems is to pretend that we can be passive as the power of these systems start to explode.

'neuro' has the connotation that any animal with a nervous system can operate or symbolize the platform

@Leonardo Hernandez Cano thanks!

@Aullen Verch I was just clarifying for other folks that do not know about it :D

@Leonardo Hernandez Cano sorry, just a typo..

Just fyi: the approach Gary mentions is "neuro-symbolic AI", not "nero symbolic".

Very valuable and important points made by the two speakers. A call for getting back to the drawing board with AI. I presume the fact that we don't know yet how humans come to understand the worland generalizable principles of language is going be hold back properly useful & smart AI

​@Travis Porco What you are describing is the system abolishing itself.

@David Peppers oh? seems kinda anecdotal to me. Seems kinda like if you've already decided that, than that's all you'll notice.

​@krunkle5136 Cannot make real connections with people hooked into their devices. Even when they poke their heads out for a moment, they are not really present. They are just anxious to get back to their device.

@Mr Sh1tCoin how's that now? Would you say that only the most wealthy *should* have access to the material benefits of cutting edge tech?

@J H I'll be honest, I'm having a hard time understanding what you wrote. Is that a meta argument for clarity of writing? Lol Could you rephrase what you wrote?

​@mago I'd really be interested in your reaction to my response to Shelley in this thread.

​@Jamie Shelley so AI getting valid inferences wrong simply by not being able to parse language _meaning_ ISN'T important to YOU or others that may be ACTUAL AI Developers? The irony of this discussion is blowing my mind right now. If you think a person is wrong for calling you out for your grammar or misrepresenting your position because "well, the person reading this _must_ be human so they can 'get it'" is the most profound reason for this discussion in the video and the purpose behind the debate. When actual humans can't be called out and reflect on why being intelligent and articulate and careful with their words.....well, we will soon enough find out how deeply problematic it will be. Do us all a favor though and realize it's a personal responsibility of *yours* for not being accountable for your own contribution and others that come to your defense using terms like 'language Nazi'.

@Gaulish Realist "I don't know why you think I agree with you." Because we both agree that Noam Chomsky is the subject of the second clause, you literally wrote that. And now you acknowledge the implied object in the second clause, thank god. You just don't seem to like the idea of the subject of the first clause being turned into a dropped object in the second clause, which is a perfectly fine to write. They weren't writing in the most formal or proper way, but we all know that Chomsky isn't an AI developer, so there's no room for confusion unless, again, you're being pedantic.

@mago It would be a projection if I was wrong but I am not. I know there is an implied object and I already told you the object is not relevant. You are making a simple thing unnecessarily complicated. The correct way to phrase it would be "As an AI developer, I find Noam Chomsky to be...". The way he phrased it was wrong. I don't know why you think I agree with you. I haven't changed my argument once. If anything, it's you that agrees with me.

@JossWainwright Cheers!

@Lo Lee Cheers, mate!

Thanks for the helpful replies. I'm not an artist, but I value visual images a lot, hence the interest. But I am a professional in another purportedly AI-threatened field, so what you say makes sense translated into my world.

@I Wonder You just copy/pasted comments that exposed you as a spineless liar in your original slander on what he had said, and you are concluding that from the experience? 😂 The absolute nerve...

@Svalbard Sleeper District “People who refuse to accept vaccines, I think the right response for them is not to force them to, but rather to insist that they be isolated." How can we get food to them?' asks Chomsky. 'Well, that's actually their problem' I've never been called a spineless rat, but find that people who are willing to say such things are often the ones who shy away from confrontation in the real world.

@KOT Yesss. Double down on your silly comments. Online, there is no cost for you doing so. 🙃

Isn't that precisely how human cognition works though? Like a filter, through the lens of memory.

The previous comment is clearly related to "solving" self-driving on actual roads, creating useful humanoid robots, etc (Tesla clearly has their "eye on the ball"). I don't consider there being any "science" here; these are engineering challenges; we're "just" using electrons to do stuff, as are our minds and muscles.

lmao no he is not

he is a boomer

Sorry Dave...

@TomTsu Let's just say I'm skeptical.

Don’t be negative

This is called the 'AI alignment problem' and at this point not only is there no solution, everytime we reassess the problem it becomes more insurmountable. It is one that I personally believe is not solvable either. Humans generally fall under the same alignment issues (we're mortal for example), and at least in theory AI would be immortal if we're able to save its state and copy it to a new machine (or it's able to do that itself). If humans could copy ourselves into a new body, we would, why would an AI not do that once we formulate artificial willpower and desire to have a continued existence?

It has been a powerful and dangerous weapon for years already, and has shaped the situation we're in now. It will likely get much worse. Will meaningful democracy survive? It's hard to say. But much of the "smart" money seems to be betting against it. Young people today face challenges greater than any generation has in a long while. Will they be able to preserve the relatively egalitarian-ish societies that were built over the last two hundred years? Or will they see it all slip away as bullies and strong men, AI in hand, clear out their opposition?

With my limited knowledge of AI it just occurs to me that any technology has a human base, a human has to program it, give input at some point and as we all know humans are incredibley flawed...I think AI is just another dream being chased that could go horribly wrong

​@Bluest i left out and forgot ten thousand others atrocities humans daily perpetrate by bare hands and knives. Firearms are nothing compared to this, the discussion about them as a danger is indulgent and irresponsible.

The powerful guys are usually the bad guys. And the bad guys will use whatever they want to do whatever they want. It seems to me we are skunked. But Jesus has promised to be with the faithful and true hearts to the end of the world.

@cathalsurfs i believe you may be right. we find we're currently feeling hope that it might eventually all be possible.

​@Number Six Yup. I left out and forgot around ten thousand other hazards and atrocities humans daily perpetrate and propagate on this planet. AI is nothing compared to this, and the discussion of it as a danger is indulgent and irresponsible.

Yes...ironic. The sound problems here came from human error...not doing a proper sound check. Perhaps having an AI doing the sound engineering would be an improvement.

I’d put all my money on this if I had any money 😅

@Andrew Rigg Standard model + speculation on the physics of information,. Donald Hoffman offers a falsifiable hypothesis of 'conscious agents' which build a representation of reality. Leonard Susskind has conjectured, wrt to nonlocality, that space might not be fundamental - also that one of the most important questions of physics is the definition of an observer. Roger Penrose has other ideas concerning quantum states of microtubules. Intriguing stuff for those indifferent to tenure.

@James Carter I agree with you but people that are unaware for example Paul Feyerabend's critique of the hegemonic, almost fascist insistence that 'science' is the foundation and basis for all knowledge, will be incapable of seeing this

I remember Fenyman comparing man made flight to birds and pointed out that they achieve the same outcome but those machines don't fly like birds. I think the same thing will probably happen to AI and true AI will look nothing like what we ever imagined.

@James Carter based on which rules of physics?

You cannot produce consciousness no matter how complex your machine. Consciousness creates the universe not the other way around.

@Bill Corley I see this as a positive, it would make sense for a lesser intelligence to struggle with more complex problems, this seems far more feasible than assuming it has no intelligence and is just guessing the solution. PS: it can solve a very wide range of physics problems, definitely more than the avg adult human.

@Bill Corley I have encountered this as well, it makes sense though looking at as a lesser intelligence, it should struggle with more complex problems.

@KOT Not really. Asking an LLM to solve a complex physics problem *often* leads to simple math errors. I've tried it a couple of times to amusing results.

@M80 Get to work dammit. Just kidding. No I'm not. Maybe a little I guess. But it has to embodied and self-referential in a causal spacetime. And feel pleasure and pain too.

Object-oriented neuro-symbolic knowledge representation

This was kinda funny but then I noticed you've got a Yoda avatar while you are doing Vadar. I stopped laughing.

At this point the AI would be a dismal failure. Ask me in about a month.

@Number Six represent the best and the worst aspect of human language

Odd that a lawyer should be the ultimate human...

@yes You need to read Chomsky's computational papers. It's not really a world of math, but computations are the substrate of cognition. Can it be modeled mathematically? is the coming problem.

@M80 Embodied and self-referential.

@Juhan Leemet They are all a comparison computation.

@M80 gradient descent does some really crazy stuff mathematically when the network is large and deep. It is being actively researched right now _how_ these networks learn, because as it turns out we do not understand how deep learning actually can perform this well. Big GPTs can often learn facts from a single gradient descent step, which sounds quite extreme if you have worked with gradient descent algorithms for large systems.

Isn't that precisely why we call it 'artificial' ? This is the best way we've found to mimic human intelligence in certain domains. It might be similar to how we do it but that doesn't really matter - it's artificial, if it wasn't we'd just call it intelligence.

@Francesco Rizzo if the AI doesn’t understand anything, it literally can’t answer anything correctly 100% of the time. And there are many questions that do not have a correct answer where it’s useful to be able to understand the subject matter (ChatGPT is horrible at music). Yes, the AI responding correctly is desirable, but we’re not getting a lot of that right now, except for incredibly common knowledge. I’ve asked ChatGPT to do basic polynomial math and it failed hard. I also asked it to write an essay on biological scaffolding and lab grown meat, and again, it failed hard. These models MUST understand language or we can’t guarantee that they’ll spit out a right answer. You could really brush up on epistemology. It’s the field where we ask questions like “What is knowledge?” That’s a pretty damn important question if you’re going to outsource your thinking to a robot.

@MusicDev if an AI does not understand language, but answers correctly 100% of the times, then it's only of matter of semantics. What counts is the result. Also, an AI not understanding stuff but responding correctly is desirable, since it has no consciousness

You missed the point. The point of bringing that up is that these models fundamentally do NOT understand language, they’re just parrots

I'll believe it when I see it in production. Cherry picking success for presentation purposes is not sufficient. I say this as an avid TMP subscriber, someone enthusiastic about text2img

Two words? Really? How could you possibly know what constitutes a good or bad discussion? What precisely are your standards?

@Number Six this was a one-time event. It was socially and economicallly unsustainable. Duh

@miraculixx and what has the u.s. done to them?

I work as a lawyer specialising in housing, debt, and employment. There is plenty of critical thinking to be done that AI is incapable of replacing

@miraculixx I would, but unfortunately it's been sanctioned to death.

@Paul O. I do advocate that one should regularly take time to explore sensational externalities or get out of your mind. However, no amount of alcohol will get you there. Edit: response to "what the fuck are you talking about? are you drunk?"

@PaulOzag Yes I will try, thank you for the reminder. I do appreciate your attempt at analogy and abstraction, because that presents the opportunity for you to question the matters that artificial intelligence can never fill regardless of how well integrated the ideas between neurons and nodes are. Such as whether having a large homeless population is good, which any five year old can answer.

@噢嗎ØMVЯ do you have enough understanding of neuroscience to understand what you flippantly refer to as an "aggregation of statistics" could aptly describe the nature of dendritic connection and action potentials between neurons? Are you saying a mechanical turk is incapable of generating emergent properties? Would you argue thermodynamic architectural intent with a termite mound? You are so sure of what you are so sure of. get over yourself.

@PaulOzag facilitative communication is still possible with inexact terms. Approximation of sense is a task humans excel in undertaking despite information loss or poverty in stimulus. As you have kindly made your profession known, you might be informed that I am a linguist and conduct engineering tasks on simple machines. You are well aware that the artifice is an aggregation of statistics, hardly deserving the name 'learning'. Most of thr effort put into the industry is thr maintenance of a mechanical turk, rather than genuine inquiry between definition and extrapolation.

Sad!

@TaylorCarman Thank you so much!

George Lakoff

And where it comes from

@KOT Sounds like it's time to don our tinfoil hats.

U better get ready ASI is right around the corner, AGI came and went without many people noticing, but ASI is 5 years out max at this point.

"If it is used ethically".....boy thats a BIG if.... most people when given the opportunity to be in a position of power generally abuse it

What doesn’t kill you makes you stronger, Nancy

When have humans ever not used technology ethically?

I don't mind the "lazy" jobs the AI will give us, as long as it can employ as many people as possible and the pay is great.

Modern programming compilers are derived from Chomsky's context-free grammars. There's no irony, only your own ignorance.