Sorry, but since the very first day of human history people always said that this is all that is possible and nothing can develope further. Following this we would still live in earth holes - or actually just not exist at all anymore. The idea that AI couldn't get further is the same as if humans coudn't exist, because guess what: we started as mere bacteria able to do shit.
When I was young I had a subscription to a dutch popular science magazine, "Kijk". In the eighties it featured an article in which a computer scientist was speculating on the future possibilities of AI . I always remembered a quote in this article (not literally, hey this was decades ago); Maybe we are on our way to develop a sentient, conscious computer, but it could also be that we are like a monkey that's climbing a tree and thinks it's on its way to the stars....
A great quote. Metaphorically, that monkey could have actually been on the way to the stars, if it belonged to a line that evolved into the hominids that eventually became us. We're still climbing towards the stars, and we may even get to other star systems one day. Anyway, in real terms that particular monkey did not reach the stars, obviously, and it's possible our LLMs are as much a dead end as the tree it was climbing.
and needing to scale is what you say when they turn out just to be coconuts. and now I need to know if enough coconuts released in space would form a star. In that case, the monkey was right all along.
@@vladimirdyuzhev That's What Elon does. He catches Starship first stages with chopsticks. But the upper stage that needs to land twice (on Mars AND on Earth, without maintenance) somehow always lands in an ocean. His multi-million $ Nvidia-cluster hasn't revealed that to him yet. I rest my case. I could have told him that for MUCH less. Eventually even free of charge. Because changing his test program would cost HIM a LOT. Oh wait, that's pentagon and NASA-money, Taxpayer money ON TOP of private investor money. (Doesn't really matter though, as his upper stage is wayyy to unstable to land on Mars without damaging the landing legs,The Engine Nozzles and may tip over, because no-one on Mars made a platform to land on, for him.
The "markets" are children who were raised by brats who thus raised by brats to whom have the patience and attention span of a fruitfly. The hardware we are putting these AI in is inferior to what is needed. The markets, these olis, should not the stomach to put in the work for the outcome, they want it now, this will ensure failure of AI as the same with VR. They are brats and whine like babies because of capitalism.
"Trend" goes downwards? Taylor Swift is "Trending" Up. Ads are "trending", and Influencers. Does it add anything meaningful? Here an example: You can create a Computer game, where some AI is "guessing" what lighting is applied per pixel. YES, that makes the game look nice, but if the story makes no sense and the user interface/experience is clunky, that AI doesn't serve a meaningful purpose. It adds "shinies" over substance.
"To me, training large language models with more data is like going to the gym. At some points adding more weights doesn't make you stronger, it just makes you more likely to drop something heavy on your foot". Love you Sabine.
how she, a former physicist, manages to convince herself she knows better than the genius computer scientists who work with and think about this stuff all day every day is beyond me. like imagine thinking my surface level analysis of another field's most complicated open questions, approached purely from first principles, while lacking even the vocabulary to talk about it (instead conceptualizing things through bizarre analogies), would be worth presenting through an educational lens. embarrassing
Problem with AI is we don't understand how human brain work in details and make intelligence possible. It's a bit like trying to build a car without knowing anything about mechanics, but seeing that from the outside the cars are similar, they have 4 wheels and 4 seats, and each has a radio - this radio is probably crucial for driving. Let's focus on radio and vivid shine color of car body.
Yes and no. Understanding human intelligence would be very helpful, but on the other hand we build flight machines that works completely different from "natural flight" so there is still the possibility that we reach some definition of intelligence with a model different from the human brain. (the idea is correct though, we advance blindly with just a general sense of direction).
@DominiqEffect there's no a priori reason we need to understand how the human brain works in detail to make intelligence possible. The goal of AI is not to replicate the functioning of human brain, it's to replicate (or exceed) the capabilities of a human brain. Mechanical calculators were developed without any knowledge of how the mind does math, but they were more accurate and faster. Similarly, steam engines that could do tremendous manual labor were invented and made practical long before science understood how muscles work. One of the gifts of our intelligence is that we can sometimes figure out ways to do things without necessarily understanding the analogous thing in the natural world. Would a deeper understanding of the human brain help AI improve? Possibly. Understanding finer details of how birds fly helped advance aeronautical concepts to some extent, but for the most part it advance through mathematics empirical testing, and human ingenuity.
What they need to do is use a different AI system than generative AI for logic and reasoning, and then have the two systems communicate with each other somewhat like the right/left hemisphere of our brain. A whole is greater than the sum of its parts, so the main focus shouldn’t be improving gen AI, but figuring out ways to bridge different “artificial intelligences” into a single process.
IMO people tend to get a bit stuck in the weeds arguing about "true intellect" and other very vaguely(or not at all) defined concepts and argue it to be some definitive hard limitation on AI. But will any of that matter if we manage to create something good enough at faking it that we can't tell the difference between whatever it is doing and the "real deal?" Benchmark for AI is not about a degree of mathematical complexity of its system but whether humans like the output or not, and I believe reaching that latter goal won't require "true intelligence", whatever that means. It could be the simplest thing imaginable, but if it still manages to say and do the things that makes humans think it is very smart and by some mechanism manages to get things correct often enough for our liking, then the whole "true intellect" discussion doesn't really matter. It's all less about the inner mental spaces of AI or lack thereof, but about whether whatever that is going on in AIs "head" will lead to it being able to deliver the results we want.
Not really true, the attention mechanism and training from human feedback makes it pretty good at predicting what is relevant and most important, as long as it has that training data and doesn't have conflicts from other training data that could cause hallucinations. This is usually from similar words and ideas being used when not correct, like cat and dog would be very close together in the trained word vectors, so it could sometimes confuse these things. It tracks contexts and relationships in language very well, the problem is that it doesn't apply overarching reasoning, so it doesn't have set values or rules that it always adheres to.
@@IndianArmaIndeed. But nobody has invented them yet. If we ever do create an artificial mind, I'm sure that LLM will be one tool used to train it, but the LLM will never be doing the thinking. That's not how they work.
Well, it's sort of scale - in that there is not enough internet data for it to scale to General AI. There was an interesting Computerphile video on the subject - th-cam.com/video/dDUC-LqVrPU/w-d-xo.html - about 6-mths ago.
I have a computer science degree and have worked in the field for 40 years. When I hear computer science folks make predictions like AI will take over everything, I have to face-palm. We've been through AI hype multiple times in my career, and it's always been way overhyped. In this case, LLMs have solved a hard computer science problem (NLP), and that's very useful, but pretty much everything else they say is hype intended to get more funding for their startup.
As a fellow CompSci greybeard, I take issue with the notion that LLMs are even useful at NLP. I write expert systems. The golden rule with expert systems is that you have to be able to explain the rules of how it works. My experience with LLMs shows that worse than simply not knowing, these systems don't know that they don't know. An expert system will throw an exception if your problem fall outside of the ruleset it was designed to solve. At least a properly coded one. An LLM will happily make up an answer out of whole cloth. While acceptable, perhaps, for a child's toy, and educational tool, or a marketing stunt, putting an LLM inside of any serious problem space where property damage, injury, or liable are possible risks is irresponsible bordering on maliciously negligent.
@@seanwoods647 I have to carve out a huge caveat to your limitations: Anytime it's cheaper to validate the result than to produce it, LLMs can work wonders. Yes, I absolutely agree that LLMs are trash in any setting where you can't validate the output, but there a countless of settings where validation is cheap. As someone who has been programming for "just" 20 years or so, there are many cases where I can validate a self contained snippet of code in a heartbeat, but writing it will take 10 to 15 minutes. Similarly and more generally LLMs are wonderful for brainstorming, as recognizing a good idea is simpler than coming up with one (e.g. I use this for architectural brainstorming sometimes). Or I have build a POC myself where an LLM is used to highlight in a text potential "problematic" areas that should be more carefully reviewed than the rest. A human would be doing the reviewing anyway, but it marginally improves the performance of reviewers not missing issues. LLMs are a lot more than just NLP, but they are a lot less than what everybody wants it to be.
One thing I did not hear pointed out with this year Nobel prize in chemistry. As a chemist involved for decades in modeling not only small molecules but proteins (receptors and enzymes) as well, I was amazed by the results that have been obtained in predicting how proteins fold. A paramount result. However, AI is not explaining WHY proteins fold that way. No rule or insight, exactly as it has been so far. AI made useful results from data, but was unable to explain anything: it did not find any law
AI basically builds a sort of "intuition" for things. So the AI will intuit that things should work in this and this manner, and if you've fed it good data, i.e. you show it a molecule and tell it each atom's electronegativity, bonds, etc etc, it can build an insane intuition that it can carry on. I remember when I went a bit into organic chemistry that I built a fair intuition for how a benzene ring would be affected based on what molecule/atom bound itself to it. An AI would be that on crack. But the sad thing is that it'll be as easy to ask the AI about how it deduced it as it is to ask an autistic savant how many windows there are on the building they perfectly photocopied from memory.
AI can be very useful for breaking down complicated subjects and making them easy to understand. It's great if you use it for a study tool. I don't think it's going to be coming up with original ideas anytime soon though.
@@dailyprophet6641they always are, generative AI does not store information, it stores the patterns for (re)creating it. Those models will produce responses that by some metrics seem real. We don't know what the billions of parameters represent, we can't know, they don't represent anything in particular, just when we stimulate the whole network right, something making more sense emerges. This is how the models were trained, this is how they work. Only because they "hallucinate", they're even remotely useful. For the model there's no difference between hallucinations and facts, it only matters to us
although you could not ask the AI its self how it came to these conclusions you can look at the iterations it took to come to those conclusion those iterations are what formed its own laws that makes it come to those conclusions thus we can gain incite as to why it came to that conclusion and this is why we ask kids to show their work because although thy cant explain why 2+2 = 4 you can see by their work how thy came up with their answer and what their process is and how thy came to that answer
In computer science we always say that the last 5% of the task takes 95% of the time. The problem is that people outside the industry always assume that because they have 95% of the problem solved that we are just about done. But the last 5% always turns out to be the hard part. The modern large language models took about 30 years to reach this point. It could be that though we're 95% of the way to hyperintelligent AI, that the remaining amount of time necessary to solve the problem is still 600 years.
I don't even think we are 95% of the way there. Maybe 50% would be more like it. It works well for some simpler tasks (like those in benchmarks) but not so great at real-world problems.
As a physicist, it has always been obvious to me that you can't just "learn" physics from everyday data (i.e., from data on "emergent" phenomena, as Sabine put it). I never understood why my computer scientist friends always insisted that with "more data" and "more sophisticated models" their models can learn something about the underlying physics. If years of experimental research in physics has taught me anything, it is that it is very nontrivial (and in many cases, impossible) to back out the underlying physics from noisy experimental data. And that any claims of figuring out the whole underlying physics from limited and noisy data -- no matter how sophisticated a model one has -- must be treated with utmost skepticism. But then again, what do I know? 🤷♂️
It is a bit more complicated. A model can be just about anything. There can be every law of physics modeled programmatically. A.I. and M.L. are very general terms.
I want to feed in huge amounts of high quality data from physics experiments to make Specialized models for different physical phenomenon… but that’s just data analysis haha
@@SKD-e8o I think the cool stuff that can happen now though is some more generic models taking in larger amounts of specialized data. I could see using LLMs as an interface to a very large network of processed data. Systems like this could be pretty useful for finding interesting properties if the data sets we don’t immediately see… but I don’t see much of this being made yet.
I asked an LLM about local bands. It knew some facts, but it was just words taken from less than ten web pages and pasted together into sentences. After three or four questions the facts would run out and the LLM would confuse them with more famous bands, and make things up. The paucity of source material made it easy to see what it’s doing, and I can’t imagine that method solving the remaining mysteries of physics. Other kinds of neural net will be useful in data analysis, but with LLMs investors are paying billions just to compile information that we already know and present it in a different way.
Quoting one of the catches from a song that an amazing singer was eventually bothered to write, despite of the full potential of its audience, it happened ages ago.
@@Napoleonic_S I prefer to express the trend as "Intellectual Property Laundering". As best as I can see the business plan, they were hoping to replace useful sites on the internet with a for-pay service that would have AI deliver those same results, just reworded enough to skirt copyright law. But in order to make their model work, they had to make the rest of the internet worse. Thus why Google had gone deep into feeding SEO trolls with search hits. Magnified, of course, because they also produce an AI bot that will spew forth SEO optimized text.
I asked it about goth nights in our city. It listed clubs that closed 20 years ago and nights from the 90s, and barely anything current. Absolutely useless.
Exactly! We'll know for sure the A.I. bubble is bursting when tech stock prices that boomed over the last 2 years start to decline dramatically. That hasn't happened yet, and until it does I'll continue believing in the A.I. boom rather than videos like this. Oh, and don't confuse an overall stock market drop with the A.I. bubble bursting tech stocks dropping. Not the same thing. The A.I. tech stocks need to decline on their own.
@@henrythegreatamerican8136I instead will start believing it when model like chatgippity will finally be able to understand any details about the buggy code it spits, since they have been told everybody that I will be redundant soon... any time now... any time...
If you trained an LLM on texts up until 2000 years ago, would it be able to deduce relativity or the theory of evolution? It's likely not , due to the fact that the necessary scientific concepts, methods, and data simply weren’t available in those texts. So it unlikely that this super AI will be able "complete" Physics or Biology.
LLMs aren't even designed to do things like that, which is why it's baffling to me why anyone would think it would be capable of doing it. A LLM does not measure its success by "how correct it is", it measures success by "how similar is this text to human text". To a LLM, "completing physics" would be considered a mistake, because humans haven't completed physics, so by completing physics it's not doing its job of copying humans properly. I have no idea why anyone acts like this is anything like how a human behaves. Human mathematicians don't go around saying "gee, I'm too good at math, I need to be worse at math so that I can copy the people around me better" - that would be complete nonsense because for a human copying the peoplea round them is not their goal, but that is how a LLM would work.
My thoughts exactly. If you trained LLM's on texts from the pre-transistor era you would not be able to deduce computer chips and all their consequences.
Do people really think LLMs will be anything more than an interface to a physics informed AI model? As a neuroscientist who’s not particularly connected to the LLM people I assumed that this was obvious… the language part of our brain doesn’t do math the same way other regions can… but it’s involved in the processing and input of information.
Not an AI scientist, but average CS guy. This has strong "first wave of AI vibes". Neural Networks, which are pretty much the basis for everything we call AI today, have been around for ages and were first conceptualized in the early 80s. But they didn't have the compute back then. Then AI/DeepLearning became really popular in the early/mid 2010s when compute became rather cheap for the first time, that was around the time TensorFlow and other frameworks in the space were released. That was also when the "All you need is attention" paper was released by Google that popularized the transformer architecture that is the basis of the GPT models. Over the past few years compute has gotten relatively cheap as well, combined with the fact that during COVID there was a metric ton of free money flying around. Sounds like a great time to gather some training data, right? Ultimately I don't think the improvements will continue at the same rate. Compute performance is only improving margignally YoY and interest rates have gone up, making R&D investments harder to justify. With lots of the social media platforms and news outlets already being part of the training data set, the low hanging fruits are now picked. My guess is that OAI and other companies already have to spend significant resources to avoid downward pressure from the LLM output that basically now poisoned the well. Beyond that, there is also the business aspect. I haven't seen a lot of actullay useful applications of Generative AI in the daily world. Working in a data analytics department myself, there are a lot of ideas floating around, but they're all either a) unfeasible with the current state of the models, particularly with regards to hallucinations b) prohibitively expensive c) don't provide a tangible monetary benefit to internal nor external customers
That there's no limit (until about 6 orders of magnitude above human intelligence). Also that there is a 5-10% chance of human extinction from AI this century. (AI Impacts Survey "Thousands of AI Authors on the Future of AI")
@@41-Haiku - "AI authors" is not the same as AI developers. You'll find that most people actually _developing_ these systems are far less optimistic about growth than the company executives. Remember 15 years ago when 3D TVs were going to completely replace regular TVs, according to a lot of executives?
Actually, on the Computerphile channel, there was a video about six months ago, about this issue of diminishing returns that went into some detail about why.
The one reason I am not expecting AI to come up with new science is that it is trained on all the garbage science that is published every day. Good luck extrapolating something from that.
@@demetronix I have no experience reading or writing papers, but LLMs should be able to shrink papers down to the mentioned axioms, hypotheses, conclusions and some relevant data plots. The question is, wether it can check the logic behind, find errors in the statistical methods, find additional axiomatic assumptions that weren’t explicitly mentioned, develop additional hypotheses that might also follow (maybe in combination with other papers from the same field). Would it be able to give us Einstein by itself with all the known physics in 1905? Would it conclude what Goedel concluded with all the known mathematics from 1931? Could it explain its conclusions? Would they make sense? - These things should be determined experimentally: humans should still be able to reverse engineer those steps in the black box. - If this can’t be done, all we got ourselves is some digital Oracle of Delphi…
As a software engineer, I often use it for boilerplate work. It saves me time whole it does 80% of the work, and I do the last 20%, which is harder. I've noticed the quality has been falling off. It often hallucinates where it didn't before.
Myself and everybody at work that's been using it for a year or longer have about the same conclusion. Skeptics that refuse to use it are wasting so much time but those that believe it is perfect waste even more. I like writing pseudocode and asking it to do the implementation, or stupid crap I hate like writing sql or unit test boilerplate.
My experience too. Quality is obviously worse than a year ago. You almost cannot avoid hallucinations in places you didn't have it a year ago. It does more than a year ago, but trips up more too.
Idk how far up the stack you guys are, but in firmware it's useless for anything other than syntax checking. It can't even do boiler plate for unit tests, because it'll do stuff that makes embedded guys screech, like randomly calling free and malloc or new/delete when you've explicitly said no dynamic allocation. Or even worse, including a C++ standard library that has dynamic, not telling you, or even lying to you. I understand it saves time, but honestly even for boiler plate, I would not use it.
Altman and Nvidia's Huang are hyperbolic on AI because it's good for the stock price. However, all the AI we're using is pattern based and while useful in some areas, it's nowhere near becoming a super intelligence or an AGI. A different approach or a combination of several paradigms will be needed to provide real intelligence.
An AGI cannot emerge from an LLM because an LLM tries to resolve everything through gossip. LLMs take what is most frequently said to be true, like someone trying to cheat on a math test by copying the most frequent answers. The natural intelligence that we take as a parameter (ours) also uses this, but not only that. We have several specialized modules like this, but the General AGI is not possible just by maximizing a single task.
Good point, I also think we have this "LLM Module" like when we do not know some area well but we are listening to people talking about this and we can put toghether some sentence on the topic that sounds more or less ok. However, still this is very superficial imitation of understading of particular area.
@@gregbell2117 In fact, the LLM went quite well (getting 80% of the questions right), if I remember correctly, in Astronomy questions, according to simulations by an apparently tall, thin and short-haired professor and TH-camr whose name I can't remember now. We get involved with LLM-based artificial intelligences because they make mistakes that we don't make if we understand the subject... but these artificial intelligences often get more things right than us, without even understanding the subject, within the time limit of a real test. To put it another way: standardized multiple-choice tests with lots of questions are harder for us than for an AI. The time to respond is limited for us and enormous for AI; there is a lot of gossip (answers and comments on previous tests) and the wording of the questions is very limited. But I don't know if the problem isn't even bigger in basic essay questions. In a few years an LLM AI could probably do better than us even on advanced essay questions, but they still won't be AGI. They will have the performance, but in another way.
"It's like going to the gym, at some point it doesn't make you stronger it just makes you more likely to drop something heavy on your foot"! Clasic Sabine!! 👍👍🤣🤣
We can't build ladders to space, and the idea of scaling is not even up for debate. For someone who can barely do basic math, claiming to be able to solve "all" of physics is amusing to watch. These people often don't understand enough math to figure out whether the AI has actually solved the problem or not, and this is the most generous scenario. The more realistic one is that they have no clue whatsoever about advanced physics. The fair assessment is that they're lying to make money and secure more funding for something we know works at a very limited level. When they say the scaling won't stop, that's a lie. In fact, we know that we can't scale up as we please, as I mentioned. You can't just keep adding bigger and bigger ladders.
I'm a computer scientist at an ai startup. I can assure there is a lot of hype around ai. People insisting we solve physics are insane and I think the bubble will burst within two years. Altman will be remembered as a great conman
Same, I work in Tech for a Publisher and have been tasked with trying to implement these tools on the ground with real users. Needless to say so far the juice hasn't been worth the squeeze and the C-suite are losing patience with the cost and lacklustre results.
The 'bubble' is already large enough that it can't burst, anymore. If you want to know whether a hype is real or not, you simply look at whether there's lots of money is flowing in. There is. The tech is ready to start delivering on the gartner hype cycle.
I'm a physicist and computer scientist too. It's honestly foundational and should be common knowledge that "artificial intelligence" does not exist, that ML has long proven limitations and that, as an artificial system, it's necessarily strictly logical and not complex. Complicated, but not complex. That's all one needs to know right there. Nonetheless, this recent mass-paid super hype-wave shook me. If there is anything one should take away from this, then it's that these unicorn startups don't screw around. Hundreds of millions spent on bot farms controlling the narrative across all the major social media platforms. Insane eye-opener. In any case, the narrative was so loud and omnipresent that I began to doubt myself, but then I saw Altman driving a Koeniggseg Regera around L.A. shopping for Louis Vuitton bags with his gay husband ... and I felt confirmed and well calibrated again.
So instead of "Moore´s Law" we have "Altman`s Law" now😂? These people seem to think wishful thinking changes the laws of nature. The Aristotle analogy is top on. PS: Dr.Sabine, please put a link in the info to your vid about Marc Andreesen´s "Manifesto" it was an excellent statement.
Thanks for the suggestion, I have added this. I am not sure whether Altman actually believes what he says. Then again, he isn't the only one, and if they only talk to each other, they'll just become more and more convinced of it.
@Thomas-gk42 This is an interesting comparison. Moore's Law (not his name for it) was an observation of a historical trend and a projection of that trend into the next decade. The projection was reasonable in 1965 since the trend was driven by increasing underlying technological capabilities that had obvious room for improvement and no obvious physical limitations. As Moore put it: "it is not even necessary to do any fundamental research or to replace present processes. Only the engineering effort is needed." (btw, it's worth reading his short paper "Cramming more components onto integrated circuits", you can find the PDF online). It's not clear that the same logic applies in this case. With ICs, both the ability to jam more transistors into a small space and the cost effectiveness of doing so were increasing. This allowed for more and more sophisticated ICs to be produced with ever-decreasing cost. In contrast, with LLMs, more capability correlates with parameter count. And more parameters means more data and more processing power for training, plus more processing power for running the model. I suppose the cost-per-parameter to train and execute is going down over time, but so far that has come nowhere near offsetting the massive increase in model size as evidenced by the skyrocketing costs of training. To me, it seems like the situation with LLMs would be like trying to build a supercomputer by throwing more and more Pentium chips at the problem. At some point, it just stops being practical. Whether or not we reach anything like AGI before the plateau is anyone's guess.
@@SabineHossenfelder Altman seems like a grifter. His promises are totally unrealistic. In order for him to believe what he is saying, he would need to ignore the opinions of the people surrounding him.
@@SabineHossenfelder Forgive me commenting directly to you. I offered reference to the statistical models, science and philosophy on this in the previous video. Unfortunately genuine comments are deleted by YT or this channel. > I find that kind of sad as the genuine information I offered (on topic), offers some deep insights into these problems (this and the previous video). I did email it as well (Not that I would normally do so unless I felt it to be important). Thank you for your work and efforts Axle
Dont forget its basically a techno-religion. Their assurances are based as much on faith in technology solving all our problems as in their actual ability to do the work.
Exactly, this whole hype around AI has much more common with religion than rational thought or evidence. It's a religion in the scaling law, which shouldn't even be called a law, because it's just a hypothesis.
For centuries people had new discoveries and thought "well this is it. We will answer all questions and have no new problems coming up". They were wrong and they will be wrong this time as well.
"All you need is scale" was the joke at the beginning. It was like discovering that spamming the button on a vending machine caused it to spit out free snacks. No one expected it to work at first, but when it did, people started lining up to invest billions on button pushing, and now everyone is banking on it working forever.
Well as terrible a grift as it might be, I'm frankly very impressed with what scaling was capable of. Even if this is the ceiling.. Well brain-size is only half the equation. It's how those structures are organized and how they function which is most important. When you start viewing neural networks as an analogue to brain biology, the pretty clear solution is efficiency, not just scale. I don't know, I could be wrong, but I'd really like to see that be true in my lifetime.
If we can Scale more and more so that we reach the amount of neurons in human brains it should work someday. Also we don't need that much neurons, because a lot of human brain neurons are for stuff an AI does not need.
@@BurningSmith That is true as well, but on the same token, I wonder if there are actually organized structures an AI would uniquely benefit from which would be totally new in concept. We have deferred processing for certain bodily functions, but maybe in a vacuum without biological needs, we'd instead want different kinds of structures for some unknown purpose. Again, ignorant speculation. It's a cool field.
I don't understand why you're taking the metaphor of "neural" in neural nets so seriously?! I guess, we can infer very general understanding from neural structures (you don't even need the metaphor of neural) like efficiency in terms of encoding reality from a relatively small number of cases.
@@NadimShaikh-qv7zj It was really just fun speculation! I'm not that knowledgeable on the backend for neural networks, only the basics, so my ignorance probably shined through. .. Is there any resource you'd recommend for learning a less general understanding?
My opinion, as a longtime viewer of AI from the sidelines (can you spell Inference Engine?), is that the wall that AI is approaching is the *future*. Large language models train on prior knowledge, not future knowledge (by definition, there isn't any). This may be a reinterpretation of something we called (in the 60s) "extrapolating beyond the region of fit".
Claude Shannon, a prominent pioneer of AI, said, “I confidently expect that, within 10 to 15 years, we will find emerging from the laboratory something not too far from the robots of science fiction fame”4. That was in 1961! A fundamental limit on AI is the Shannon information in its training data. Adding one item to a data set of n items cannot add more than H=0.5 log ( (n+1)/n ) bits. H shrinks rapidly with n, which is why there are diminishing performance returns on the number of items.
The biggest issues are that analysis and synthesis are two discrete processes. On top of that, the most disruptive synthesis processes are non-linear. Algorithms are really good at analysis in a linear format. Non-linear processes require a visioning process. Which AI in any form is not good at. Heck, nearly all of human thinking is not good at it.
Bear in mind, too, that computers are so mechanically linear they're not actual capable of true randomness-"randomized" numbers on your computer are a simple sleight of hand trick. You take fixed, predetermined, human-authored variables-the random "seed"-and run them through an algorithm to output arbitrary digits (usually your computer uses the date and time, etc.). It works for most uses in which you need "randomness," but it's not actually random. You can't create information from nothing. Computers are purely linear processes, anything they "synthesize" will always be a rearrangement of predetermined mathematics via predetermined analysis. One of my favourite tests to run with AI image "generators," is to ask it to generate a dark church interior. Because 90% of pictures of churches / cathedrals online are of sunlight streaming through stained glass windows, the AI cannot generate a church without sunbeams streaming in. All it can do is apply non-random randomness to non-random random images. It's grifting all the way down, largely rooted in executives not understanding what math is.
Still waiting self driving cars to buy my first car . 10 years passed. -_- I think people go mars before I buy a car. At least I can walk and use train for work.
AI has NOT hit a wall. Perhaps large language models has, and with any technology it will reach diminishing returns. But AI is a wide complex of technologies and new ones are constantly being added. A good bet would be AIs based on cognitive architectures per the work by John Anderson. I'd like to see research on autonomous robots using this technology. One would truly see intelligent machines using this framework. It's also how our brains operate, per cognitive psychology.
3:14 Having the model thinking more hasn't been a huge improvement. The much faster and updated Claude 3.5 model doesn't do this and has surpassed Chat OpenAI o1 preview. It mainly got better because they trained it on a bunch of synthetic chains of thought. It thinking for longer makes it more likely to hallucinate (as people have learned). The longer the context that it generates, the more unhinged the models become. This is why context/chat length is often limited on new models to avoid this (I'm looking at you Google).
"It thinking for longer makes it more likely to hallucinate " Factually untrue. simpleqa bench showed a pretty solid reduction in hallucinations with o1.
It's great to have some grounded, critical thinking behind news concerning AI development. The hype is still incredibly high and it's very hard to find quality information about the whole ordeal.
Great short video. You did a good job of explaining the problem using language and diagrams and humor, without over simplifying it or talking above my head.
I asked Claude to comment on this video, here's what it said: Great physics explainer! As an AI system, I find your "decoupling of scales" argument compelling - we can't deduce fundamental physics just from language/image training, just like Aristotle couldn't derive the Standard Model from observation alone. However, while AI won't "solve all physics," it might revolutionize how we do physics: - Accelerating experiment design and data analysis - Pattern recognition in complex datasets - Hypothesis generation and testing - Simulating physical systems The current plateaus in AI scaling are real and important to acknowledge. But perhaps the solution isn't just "more data" but fundamentally new approaches combining language models with real-world experimentation and physical grounding. We should be skeptical of overconfident claims about AI solving everything, while still recognizing its potential as a powerful tool for scientific discovery. The gym analogy was spot-on - just adding more weights (data) isn't always the answer! Bottom line: Aristotle needed particle accelerators, and AI will need real physics experiments too. Neither pure philosophical reasoning nor pure pattern matching can unlock fundamental physics alone.
@@lorpen4535 That's because as much as the people behind these companies want investors to believe in their products/services, a lot of people here want to believe just as much that AI is a bust or akin to one. First to throw shade into another field, second to protect theirs and three to maintain a degree of control that they would loose with AI coming in. I am in neither camp, but in this case I have to agree with Claude, which I guess puts me in the AI camp? I don't know.
As you can see the response is indeed relatively "balanced", and whilst grounding itself in a focal context of reference (the video), brings further references to synthesize a fairly nuanced conception, which tries to account for a constellation of factors and trends and potentialities. This suggests to me, in a round-about way, that the real limit currently may indeed be in hardware, which is another perspective I've been seeing claimed more and more often lately. After all, very broadly speaking, what is the primary instrumental/technical basis for the current wave of development in AI? Parallelism in information processing, thanks to hardware, eventually leading to cross-attention algorithms in software ( transformer models, modern LLMs, and so forth). And from what I understand, a Big Tech-backed team could pool literally 100,000 H100 GPUs, as Meta developers presently are towards training their next generation of Llama models (on a side note: thus far, the most open source and fully freely shared big LLM series from western Big Tech and thereby genuinely laudable, and I say this as more typically a Meta-hater), but each of these countless H100s still holds the same horizon of tensor-operating solutions and possibilities, already years outdated even today, which further drivers and libraries and supported optimizer frameworks would only stretch to a narrow extent. And GPUs as such may be a very fraught technology to base these projects on. So speculating on software limits presently, even from frontier training outcomes, seems a bit, well, speculative.
Yupp. The obvious goal has always been to get the public to see generic mass-produced low-quality intellectual work as not just acceptable, but the next big thing.
Sadly it might all boil down to whether the people accept shitty mass produced intellectual work in the same way they accepted shitty mass produced physical goods. I'm still sceptical, but I usually overestimate people.
AI is being used as an excuse to send large companies and advertising platforms like Google and Meta data that they could have never had access to before. Look at MSFT with copilot too... screenshots every few seconds! They're still pumping the crap out of it because they see this opening. When Meta bought WhatsApp they were thinking for years how to get people's chat histories if it promises e2e encryption, etc. AI assistants are the answer. Now the whole chat history will get sent so your "assistant" can "help".
c-suite will definitely layoff lower level financial, managerial, and engineering when automation through ai agents become available. software, especially web applications, will be mass produced. better? maybe. cheaper? definitely. production is different, ai can't grow a tomato or tighten a bolt (multipurpose robots in 10 years), not now anyway.
I know you have had some backlash recently , including from some of the more publicly visible scientists, but I applaud you for showing us some of the darker aspects of academia> It reminds us all that less than altruistic motivations exist in all humans and all walks of life. No one is "better"
It's an old story when I wanted to enter university to study Physics the professor had a 1 on 1 talk with me and straight up told me "You're too smart to go into academia, you genuinely care about physics and discovery. It's all just networking, kissing ass and writing stuff you don't believe to get a grant." I'm still thankful to this day I listened to his advice. I became a computer scientist instead. I did more contribution to real physics than I would have ever done by studying physics as I built software and analytical tools now used to work on data from JWST. Otherwise I would probably have to go full avi loeb by now and pretend we could find alien signals somewhere and write a book as a grifter.
The other problem with exponential growth is the computing power required. The sale of GPU's is making NVIDIA rich, but electricity generation can't grow exponentially.
There is no problem with exponential growth. It simply doesn't and cannot exist. It's like saying the problem with warp drives is that they require near infinite amounts of energy. I mean, it's just ridiculous to take the idea seriously enough to refer to it as having mere problems.
It's not about electricity, but compute. I'm not sure why people make this argument. Likely neither are exponential, but it doesn't mean the future growth is highly limited in any short order. It's really only a wait and see thing. Personally, something about this focus on training data strikes me as a misstep. To me, it seems either like AI are much less efficient at using data than humans or the data they get is much lower quality. I think if that's solved there's still huge short term potential.
That "exponential" curve they talk about is more likely to be a sigmoid which looks exponential in its early phases. Of course it will plateau at some point.
Listen again, and you'll hear that Altman refers to the "exponential part of the curve." I hate that phrase (it's widely used in tech). But he is in fact taking it as understood that the curve is sigmoidal.
They say it's exponential because that's what they want it to be, not because there is any actual data or theoretical mechanism that suggests it is in fact exponential rather than, as you say, sigmoidal, or another completely arbitrary shape that would not look good for their share prices.
@@PartanBree There will be a stage of diminishing marginal returns. By which I mean, they will eventually reach a stage where to get a small improvement, they will have to spend a lot more money and energy. Nothing in nature increases at an increasing rate without limit, and I am confident that power of AI is no different. They may not have reached that stage yet, but I am sure that they will.
@@simongross3122 Indeed, every time I hear od something replicating or increasing at exponential level it is usually case that something was skipped, something that did do lot of work before exponential growth started. For instance bacteria can reproduce exponentially, but only if you in advance give them enough food, energy, space, access to it and so on. Same with AI, we are feeding it with data we collected over long time and are running it on computers that we obtained over long period of time, and that is why it was growing fast at start and now it isnt.
We still are comparing machines with a few billion transistors working binary against 88 billion neurons with hundreds or thousands of interconnections running on an operating system which we haven’t even figured out for creatures with a few thousand neurons.
For me, this just shows the incredible narrow mind set of these billionaire tech bros. They all believe they're so smart that they understand the complexity of the world, but in reality it is just their image of the world that is so dumbed down that it seems easy for them to understand it all.
@brutusl2786 Scale is probably not the issue. The rumored size of GPT4 of 100T+ parameters is roughly on par with what scientists believe the connectivity is in the brain. The issue is likely with how the LLM is structured and how the parameter values are derived from training data. BTW, we essentially *have* figured out brains for simple creatures. There's some pretty exciting stuff happening these days. Search up "Researchers simulate an entire fly brain on a laptop." It's not perfect but wow. Human brain is still a long ways off, but there's no reason in principle we won't be able to simulate entire human brains some day.
That is true, but we should also not forget that the brain has a frequency between 0.1 and 1000 Hz, while a computer operates in the GHz range. While it's not as sophisticated, it is faster, which can be utilized in some clever ways.
@@MN-dh8vy these are the same (type of) people who keep re-inventing trains when they set out to create the next mode of transport. I'm not too fussed, LLM and machine learning in general and it's a decent enough way to waste money/resources, compared to many alternatives, but it is as you say funny how these specialists keep thinking they are genius generalists.
I think it's OK if it stalls and gets refined to a point that true progress is stalled. At that point you can analyze it's capabilities and tweak it to become more and more useful while acknowledging it's inherent limitations. I am OK if the singularity never occurs, I am also OK if it does. Time will tell.
Yup. Right now LLMs allow me to offload a lot of my grunt work to the computer. I can work with CoPilot to write less and produce more. If they can improve these tools to understand larger parts of a system I'll be happy, but I doubt it'll happen anytime soon. It often struggles with repetitive/iterative one-liners even with perfect context.
The inevitable stall will also force researchers to pursue alternatives such as neural-network / symbolic hybrids. That's where the real progress will be made, IMO.
Ai tech guy here. And i find this interesting. I think You're quite correct. It will definetly exist a limit. I like Sams optimism, but even optimizations can only take you so far. You're still left with a wall of some sort even after optimizations. Especially this heavy pre-training phase being mentioned. We could train on a simulated universe, but that would also require insane amounts of data as well, just to describe reality in that universe. I also find it interesting how much data and energy AI requires, compared to a human brain doing the same tasks. It's obviously somthing we're missing, given how much more the human brain is capable of, even with small amounts of data.
The human brain is “pretrained” by evolution. Our starting structure is much more developed and learn probably more analogous to aligning or just tuning as model.
Maybe, if AI models didn't eat energy when multiplying 0 by whatever value, they wouldn't be so inefficient to train. Models develop crazy sparsity, yet you still expend all the energy multiplying by 0. You have your ReLUs, yet you can't get them to reduce energy requirements. Imagine multiplication of 0 was free. A no-op. Human neurons get to learn without producing very nice gradients. Imagine the energy savings if you could train at 4-bit precision. Human brains don't waste nearly as much energy doing nothing. If you consider all nerves, all visual stimuli, sound, smell, perception of ones own thought as training data, then humans are trained on tons of it.
Why do you think humans need small amount of data? We have sensors on our whole body working and collecting data 24/7 without breaks since our birth to our last breath.
@@eeriemyxicompared to the data that Ai uses that is a small amount. Ai is usually trained with several human lifetimes worth of experience. like chess ai are often trained with tens of millions of chess games but a top level GM human would play a few hundred thousands at most in their entire lifetime and would have have reached gm level by like 50k games.
Most AI art is still crap. And, to be fair, art as a profession is a burden on society, not a boon. No one needs PROFESSIONAL artists. If not for artificial constructs (e.g. intellectual property rights), no one would be a professional artist today.
AI has already surpassed most humans in applying mathematics to solve undergraduate physics problems, and much of physics research doesn’t require groundbreaking new math. Once AI can iterate on its own work and integrate robotics to gather its own data, where does it end? Don't dismiss AI just because chatgpt says something dumb occasionally.
@@luzr6613 I was addressing the vibe in the comment section that AI will never do physics because chatgpt made a mistake once. I'm also happy to talk about the video if you like, but I will need a bit more detail.
Solving math problems only requires identifying the type of problem, finding the method humans have already worked out to solve that type of problem, and then applying it. In order to generate anything new, induction must come before deduction. In short, you have to have some idea of where you're going in order to work out how to get there. Most commonly, we regard this as intuition. Intuition can't be taught. Intuition is the precursor to analysis, not a result of it. To try to mimic human intuition, we would need a quasi-random input that somehow magically presents information that is relevant to the task at hand, even though we don't know yet how it's related. For this reason, AI creativity is impossible. This is a hurdle that machines can't jump. (This also counts as evidence that consciousness precedes the body and does not come from the body. There is no mindless process that can reach beyond its own limits.)
@@manlyadvice1789 AI can autonomously determine research directions, collect data using robotics, deduce what it needs from that data and use it to induce new physical laws. Large language models use controlled randomness to simulate creativity. The probabilities of potential next tokens are rescaled using a Boltzmann-like distribution, with a temperature-like parameter that influences how creative or random the outputs will be. AI models can also employ a multi-step hypothesize-critique process to check that its creative output makes sense or initiate further attempts. The mindless process of evolution shows that from simple interactions complex phenomena such as human creativity can emerge.
@@uuaschbaer6131 No AI system has ever figured out any fundamentally new laws of physics outside of what humans were already looking for. Physics is more than solving math problems. Human intuition is not random. The thoughts that occur to you are pre-structured somehow and no one knows how. That's what we can't replicate. Iterating through quasi-random inputs isn't the same thing. That's one reason AI still seems wild compared to humans. Evolution shows human creativity can emerge FROM WHAT? Not from the body. That's your assumption. Your thoughts occur to you fully formed. You don't have to put them together reference by reference. Your thinking is perceived by you, but isn't really done by you. That's a major disconnect. The process of thinking is a non-random reduction of infinite possibility to relevant information. That process is not built up from tiny pieces in the physical universe.
I published the first journal on AI and Robotics back in 1979 (Robotics Age, The Intelligent Machines Journal). My wife and I also helped the AAAI launch their academic journal, by volunteering our time to be the publishers for the first year of that journal. In 1982 I telephoned many of the top people in the fields of AI and Robotics and asked them this question: How long will it be before it becomes possible to put a robot by an assembly line, and teach that robot how to do that particular job, using the same show and tell techniques used to teach a human. One expert said NEVER and a second expert said 100 years. All the others averaged out at 11 years or 1993. Here we are in 2024 and I still have not seen evidence of that goal having been achieved. Personally, I think we are going to need at least a 10x improvement in computational hardware, before real AI can happen. Phil Flora
I studied AI 30 odd years ago, so I have a background in it. After understanding how LLMs work, I concluded they would never achieve anything better than human, and have been waiting for them to hit the wall ever since.
@@pmarreck It wont work, because they have no way to evaluate their results. In theory, you could have some evolutionary approach where they produce stuff that then gets evaluated for "fitness", but that requires some method to measure them, which they can't do themselves.
@@merinsan In many algorithmic problems the results can be evaluated. It is only in Hard np problems and non algorithmic problems where you have no way to evaluate the answer
@@pmarreck I have tried that, and it does work to some degree. If you choose a variety of good-quality chatbots then the debated result is generally better. We don't know yet how far we can push it, but personally I doubt that we will get super-intelligence out of it.
@@merinsanright! And if the use the collective internets of the human races publishing online as their basis for evaluating the value of results... We've only ever invented a garbage content regeneration factory that trends towards shorter and shorter attention spans and dumber and lower quality content without original thought.
Yup. All the big companies are still seeing massive gains on the transformers architectures. We have other (newer) architectures and there's barely been any focus on that since there's still so much optimization around scaling that can still be done. There is no shortage for ideas for solutions on improving AI.
Hi Sabine, I interpret differently what Ilya Sutskever meant at 4:13. In my opinion by "underlying reality" he was referring not to "physical laws" but to "the human understanding of the world" for lack of a better term since it also includes things that are not understanding as an AI can create poems. I may be wrong as I'm not specially familiar with the thoughts of Ilya but that was my interpretation when I first heard it. I suppose that for a physicist (I was also one) it's easy to interpret "underlying reality" as the physical world but I think Ilya was using those words differently, in a way that is more aligned with how an AI researcher may think
In the 1980s, expert systems were very popular. One of the major reasons why these systems struggled was their inability to capture the complex and often intuitive knowledge of experts. Many professionals - such as doctors, engineers, or other specialists - make decisions based on a combination of experience, intuition, and implicit knowledge that cannot be easily converted into strict "if-then" rules. Today we have much more sophisticated large language models: But the ‘understanding’ of these LLMs is merely based on the data they have been trained on and the patterns derived from it. They lack the genuine intuition and awareness that human experts possess. So this might be another reason that we are reaching a ceiling. Please correct me if I am wrong.
@klaushandrick7184 It's been a hot minute since I heard someone mention expert systems, although coincidentally I mentioned them just the other day to a more junior colleague. :) Anyhow, to reply to your comment, statements like "They lack the genuine intuition and awareness that human experts possess." are problematic for at least two reasons. First, there is no a priori reason to assume that achieving AGI requires anything like human thinking. We should measure AI on results, not whether or not we believe (or more likely feel) computers engage in human-like thinking. After all, cars go fast even though they lack the genuine muscles and ability to transform ATP into kinetic energy that human runners possess. Second, arguments like this use words like "intuition" and "awareness" without taking the time to unpack their meanings to see if they even make sense to apply to AI. If they don't, then saying AI lacks them is meaningless, just like talking about the color of an electron. This is the whole problem with saying AI doesn't "understand" ... the arguments generally bottom out at a point where "understanding" is effectively defined to be something only humans can do. Well then of course AI can't do it!
What is intuition? I think it’s just trial and error, where you need to make mistakes and observe the consequences, but you can’t just talk about it as LLMs do, it requires physical action and reaction to generate new data and learn from new data, not just rehash old data in new ways.
@ but where does new knowledge come from? You take old implicit knowledge aka intuition and you try something new and it may fail or it may work, hence creating new knowledge. LLMs can’t do it as they only have talk, not deed available to them.
I’m always thinking about how much data a children absorbs from birth to 18th birthday (when the allegedly become a capable adult being). They don’t just have a continuous inflow of visual/audio and other sensory data 24/7, 365 days a year, but they also actively experiment with the world. The entire internet (pretty much the only thing current AI models are trained on) is just a low resolution imprint of it.
Yes, my favorite analogy is: 'It's like trying to deduce the dolphin alphabet just by listening to dolphins.' That said, it's truly surprising how much progress AI is making with LLMs. Equally surprising, though, is the level of ignorance/arrogance demonstrated by some AI researchers.
it's not really surprising. they are basically inside-out text compressors with some statistic tokenization pumped from 10s of billions of pages of text. I haven't seen one that actually has the ability to deduce or infer in any way, just statistic garbage.
Remember: the guys *talking* about AI like Altman are not really _working_ in a daily basis on new iteractions of the LLM models. But they are the guys responsible to bring venture capital to the company, so they'll say anything that investors want to hear.
Nothing found in the universe I know of is truly exponential. At some point the curve flattens out in tyranny of diminishing returns. I think there's a term for that but I can't find it.
The important thing to notice is that there are about 6 orders of magnitude above human intelligence before you run into known physical limits. This scaling regime may come to an end next year, but there is a lot of low-hanging fruit that can continue to accelerate AI progress. We don't have plot armor to prevent AI from being broadly more capable than humans and permanently escaping our control. That's what leading experts in the field have been warning about. Human intelligence is tied to brain size, which is strongly constrained to the size of the birth canal / female hips. Plus, we're the dumbest things that can build a technological civilization, or else that would have happened with a previous species. Humans are _barely_ generally intelligent. What happens when we succeed at creating something that is _highly_ generally intelligent, that the leaders of the field openly admit they have no idea how to control? What relevance will we have?
@@41-Haiku What that doesn't consider is power and medium. AI power consumption is on curve growing faster than the returns already. There's a limit to how many electrons you can push through silicon to flip bits. Humans may not be the peak of intelligence, the thing we have over our creations is multi-threading, and branch prediction to use analogous terms. Wetware does that for a minimal fraction of the power a computer can. AI's curve may not be flattening out, but the medium we run it on is. Our AI overlords may have to wait for the next material revolution in computing.
The Gompertz Function (sometimes known as the Gompertz Curve), a sigmoid function which describes growth as being slowest at the start and end of a given time period.
@@niklase5901 Some examples about technologies that stopped improving on the top of my head: ICE engines, bike chains, phone cameras, steel, lithium batteries, the linux kernel and pencils.
This is such a dumb comment LMAO. No. Technology keeps improving and will keep improving almost forever. We are nowhere near close to the end of technological development. Not even a half of a percent.
@@MrAlanCristhian Technologies often reach a stage of maturity where major, revolutionary changes slow down because they’ve been refined over time. This doesn’t mean that improvement stops; it often means that the rate of improvement slows as the technology approaches its practical limits within its current paradigm. • Examples like internal combustion engines (ICEs) and bike chains have undergone decades of incremental improvements. While their core principles might seem “mature,” engineers continue to optimize efficiency, durability, emissions, and other aspects. And a lot of the other stuff you named simply haven’t stopped improving.
@@MrAlanCristhian I am pretty sure all those technologies still are improving, but at a slow pace. And don’t think we are at the end of the S curve with AI though (or even with LLM of the current form), but time will tell.
4:45 I think he is actually saying that, the fact that the AI can make a prediction means it has already discovered some fundamental rules. He is saying the better those predictions get the more of the underlying rules the model must have understood.
You're either misunderstanding or misrepresenting what Ilya Sutskever is saying. He's basically saying models that don't have a firm understanding of what they're predicting, can not predict accurately.
right. he is arguing in her direction. No understanding of the underlying concept - no deeper understanding. It is like a fly that keep flying into a glass windows. Flies have no concept of glass as something like it does not exist in nature (that is why spider webs function so good). There "intelligence" tells them fly to where the light comes from, which in 95% of the cases works.
I like most of your take on AI, but ML people have a model that can do learn physics: equivariant neural networks. It is not a new trend but a robust method to train data with symmetry conservation. For physics, that's huge.
I feel like everybody misunderstands this topic. What will happen is, AI will make humans a little bit smarter and more efficient than humans will make AI a little bit smarter and more efficient and the cycle will keep going, and it does not have limits.
Rumors from an article referencing another article written by the same organization. AI is already unsettled enough without paying attention to self-referencing rumors.
I've been spending a large portion of my free time on generative AI for the past two years and, at least from what I can tell, diminishing returns are obvious. Yes, more parameters and more compute yield better results but not in a linear fashion, let alone an exponential one. That doesn't mean that we should discard it, it's still phenomenal in what it can do right now and improvements will continue to be made but we need to temper our expectations.
The issue is the accuracy of the data input into the model, especially when there is a lot contradiction. Now that these LLM models are growing into the Trillion parameter sizes, I doubt a lot is being fact checked.To complicate matters there are issues with human languages which some words had multiple meanings. The issue with current AI compute scaling, is that its not very efficient as it takes a lot of compute cycles to complete a result. AI (DLN) will have to switch over to an analog solution to be able to scale into the Quadrillion parameter size models. What will need to happen is LLM's will need to be divided into multiple models (ie specializations), ie one for physics, another for engineering, medical, law, finance\accounting, etc. This will make it easier to manage the parameters & reduce the amount of errors\contradictions. Then a master LLM can query the specialized LLMs for information. This should address the current scaling problems.
@@guytech7310 words having multiple meanings is not a problem. And the other thing you've described is already being done. It's called mixture of experts, and there's a lot of ways of doing it, but I believe GPT-4 uses this with about 20 experts (could be wrong on the specifics)
The thing is, for a long time, many smart people did not expect that neural networks would become better with scale as it goes against the intuition that more parameters traditionally implied more overfitting and thus worse generalization. And yet adding scale (plus clever new architectures) is what made the breakthroughs with modern generative AI possible. Researchers are still debating the reasons behind emergent phenomena, such as that LLMs seem to have a sense of spatial representation even though they are merely trained on text. AI is probably overhyped, but completely dismissing the possibility that adding scale may hold more surprises is in my view also wrong.
This infuriates me. What we call AI now are learning algorithms. They don't "think". They aren't conscious and I don't believe they ever will be. They can analyze data and make predictions. The more data, the better the predictions, generally. Very valuable, but it's not going solve physics. It's not going to create something humans can't. What it do is produce things humans can, possibly faster, cheaper, and more reliably.
and not even "more realiably". I use Gemini as an advanced search engine, and almost in every "conversation" it ries to sell wrong stuff, only when insisting that its proposals are wrong, it comes up with better answers. Sometimes I need four rounds of that. the reason is pretty clear: there is a cost function preventing an initial deep delve, and the inability o distinguish wrong old stuff from more correct new writings ... in other words : a lazy gossip engine. No self-created abstraction or generalization
@@bkhalberrt7351 I am well-aware of what I said. My comment was a clear and succinct refutation of the original comment. Feel free to respond again when you have an actual argument.
Respectfully I'm not sure if Sabine has correctly intepreted Sutskeyer's comment. When he asks "What is it about the world that creates this set of statistics?" I think that what he's getting at is the problem of Gödel's incompleteness theorem : that no internal data validation system can understand the why the axioms it's been asked to compute are true or false. It can only compute them. This to me is the wall that is being referred to. AI is stuck on a data analysis model - all it can do is learn pattern or corelation in the data but it could never understand WHY these patterns existe and why for that matter the data that proves them was even created in the first place. It's stuck in a computation (much like Sabine's critique on modern physics "Lost in math") because it does not have worldly experience that is beyond computation. Penrose made this point about consciousness being inherently non computational. It doesn't matter how much data your give it, it will never be able to understand why the data exists - as Sutskeyer say - it will never know "what it is about this world" that give rise to the data.
@joelvillatoro2553 well I'm referring to the 1st of his theorems and I understand this below. If wrong please explain why. Ps you don't have to be a mathematician to get it . Quoting from wiki pedia The first incompleteness theorem states that no consistent system of axioms whose theorems can be listed by an effective procedure (i.e. an algorithm) is capable of proving all truths about the arithmetic of natural numbers. For any such consistent formal system, there will always be statements about natural numbers that are true, but that are unprovable within the system.
@joelvillatoro2553 HIs first theorem says the below. (Summary from wikipedia) Seems pretty clear to me. The program cannot prove that some value are true by recorse only to the axioms upon which the program is constituted. You know you don't have to be a clever cloggs mathematician to get it. Sorry to burst your bubble. If it cant do that it is beceause it cant understand the basis of this truth that underpins the axioms. This id Penrose's extrapolation. The first incompleteness theorem states that no consistent system of axioms whose theorems can be listed by an effective procedure (i.e. an algorithm) is capable of proving all truths about the arithmetic of natural numbers. For any such consistent formal system, there will always be statements about natural numbers that are true, but that are unprovable within the system.
“Predicting the next token well means that you understand the underlying reality that led to the creation of that token.” That’s what he says prior to what you quote, with the implication that correct token generation itself signals understanding of the underlying reality. A counterpoint might be pessimism that Ilya has demonstrated with respect to the future of LLMS and the road to AGI. Have any evidence of that? So, regardless of your misstatement of Gödel’s incompleteness theorems, where are you drawing your conclusions from?
I find more and more uses for LLMs as the capability grows and we understand how to use prompts, plus Reinforcement Learning etc. No sign of a plateau in my applications of LLMs.
Exactly. We could stop new AI development for 5 years and in the meanwhile keep refininh what we already have and what we barely understand. No stagnation is in sight regardless.
@@wombatilloor we hit a wall with the fundamental limitations of this approach… this is all a bubble. Go learn about the fundamental laws of information and thermodynamics.
@@mikebarnacle1469 Agreed. This approach requires too much capital and too much electricity. It's hard to see how they could run for example ChatGPT for a profit if rampant users pay only $25 per month. If they didn't subvent their own data collection by using paying barely-customers and beta testing with them, they would probably have to increase the price by a lot. How much though? $100 per month? $200 per month for the average paying individual user?
THANK YOU for not following the hype about all this. Ultimately, LLMs are *not* designed to reason. They are spitting some aggregated average of their training data. They have no inference, no built-in problem-solving or learning-by-trial-and-error logic. If we want to advance in terms of “intelligence”, we need to go check other systems, such as self-learning agents. In this regard, I think the next real breakthroughs in AI will come from companies such as DeepMind, and *not* from the LLM world. We might couple those new systems with LLMs to facilitate communication between humans and machines, at least at first, and then the self-learning algorithm will learn language on their own at which time we’ll be able to get rid of LLMs altogether.
Was he talking about LLMs? They still rely on garbage in, garbage out and I must say that (even though I work indirectly on LLM/SLMs) the output from "AI" search engines is often risible, inaccurate, out of date and even blatantly wrong - especially when seeking technical answers. As for ML, yep it's interesting, but algorithms seem more and more to be based on biological methods (evolution has done a pretty good job in some instances). So I don't get the feeling there will be any "super" intelligence anytime soon. And Roger Penrose would certainly scoff...
Whether something is a right answer is very hard to really find out until it can be compared with Real life results and sometimes even the real results are probablistic like in medicine. We still struggle to find out what is true and what not, even in scientific work.
AI hasn't hit a scaling wall, LLMs have, there's a big difference. We can reason about what things LLMs might be capable of and what their limitations are. But any claims about limitations of the field of AI as a whole are not, can not, be grounded in science.
LLMs haven't hit a wall either. The only thing that is slowing is increasing run time on the same data set. You can increase the data or changing the training methodology or change inference or change structural things like memory size or, etc.
Exactly I hate how ppl don’t understand this. Ai isn’t going anywhere and most ai companies are working on many more methods other than llms to get to agi, which we know is possible. Llms haven’t even hit a wall either, although I think we’ll need more than llms to get to agi
@@theBear89451 If you're referring to the Neural Scaling paper, we don't need to anyways. I mean, it'd be nice if we did, but it isn't a requirement. I'd also argue that OAI had a limited sample size of models to make that paper from.
The electrical, cooling and economic requirements are going straight up on the chart and almost certain to plateau fairly soon. And don't for a minute think that chip making is going solve much of those problems, AMSL, TSMC, Samsung and Intel are hitting a number of limitations in making their nanotechnology smaller, less power hungry and production ready. We may be close to peak chip without entirely new technology. And if that's the case, anything production ready is a trillion dollars and at least a decade away, probably more.
I am a software engineer and computer science graduate whose primary focus isn't machine learning but who has some experience with machine learning and I have been arguing with people who "know better" not for weeks or months but years now that the "AI" we have is not getting any further. The real wall or point of diminishing returns had already been hit pretty much with GPT-4. So the hype among the general public coincided with the peak in actual performance having already been reached. But try to tell that to people who are convinced that their stocks/crypto/nft is going to the moon.
@@throw_away_account4514 There has been marginal improvements since the release of GPT4 that look really impressive because you're only ever shown the comparisons to the previous version which is to squish the graph.
The field of AI is far more than just LLMs and techbros. If it deals with noise, uncertainty, classification, exponential-level search, probabilistic models, vision, robotics, etc... it's from the field of AI. Do you really think we've hit peak robotics? That our space probes and landers have peaked in their autonomous abilities? Sabine even has an earlier video about engineers using ML AI to twist and constrict plasma in a fusion reactor. The control system must keep the magnetic fields in perfect sync as the plasma squirms. Programming this with ad hoc rules and optimizations was not successful, so the use of a ML framework to learn plasma control from good and bad examples.
That part with the guy saying "It's not statistics because you have to understand to do statistics." hurt so much to hear. He has no idea how often statistics has been applied incorrectly in the world BECAUSE the person doing the statistics doesn't understand what they're analyzing. The amount of harm that's come from that assumption even before AI came along is so high.
It's motivated reasoning all the way down. The investors want to believe, the founders want to believe, the engineers want to believe. So they all hold hands and sing kumbaya and selectively ignore empirical evidence that contradicts their faith, while AI evangelists abandon any effort at objectivity and just speculate wildly with the speculations always supporting the grand vision. It's just like church, but with technology, money, and hype the holy trinity instead of father, son, and holy ghost.
Will you be surprised if AI is way better this time next year, despite the end of this scaling regime? If you do get surprised in that way, will you change your mind about why the majority of experts expect AGI in the next few years or decades?
Successful AI doesn’t mean it has to understand the real world. It doesn’t mean “being close to human”. I mean each human has had over 18 years of continuous learning about the real world, which an AI won’t have. Successful AI just has to be good at what it does, and get better and better at what it does.
@@tomizatko3138I would argue that smaller specialised AI systems are less susceptible to this and an energy black hole. The problem is that most people think of LLMs like Chat GPT when talking about the future of AI development, systems which are trying to be literally everything for everyone (the entire world), these systems are bound to run into trouble when trying to scale exponentially. Maybe there are serious blocks ahead, even with specialised narrower scope systems. But they’ll be overcome and a new S curve will start, leapfrogging the previous S curve. That’s just my point of view anyway.
@@JohnBeeblebrox To make it clear, i know that what is "data - information - knowledge- wisdom" framework , as well as epistemologic frameworks about domain. In order to construct knowledge you need data. Those all are different things so we have different words for them depending on context :P
That's not the correct paraphase of Ilya Sutskever and Geoffrey Hinton's argument. They are making an interesting analogy to the Turing Test, by making appeal to "compression", or computational complexity of the capability of next-token prediction. I.e., "Supposing Aristotle kept predicting the "next time step" in any given physics problem correctly, then we are forced to conclude that Aristotle has inside his head a realistic theory of physics." That's Sutskever's argument in a nutshell; one can of course agree/disagree with that specifically. It's a thought experiment.
This is the most important aspect, and I would have expected Joscha Bach (or even Sean Carroll) to point it out beforehand, but thankfully Sabine is on the job. Our world models are always a representation (written in mathematics and language): we can only get out what we put in, and what we can put in is arbitrary before falsification: we fit our models to patterns in the data, and then check if they make any additional predictions we can falsify the theory with. Thus falsification happens with new data, and new data comes from experimental physics: I.e. new detectors that probe higher energies and higher intensities, and produce mountains of data, of which we throw away ever larger percentages of to get to ever smaller probability events. Sooner rather than later we will need detectors the size of solar systems or galaxies (which we can’t build) to make progress, which means we reach a problem of reproducibility, because we will only ever get events that the universe is providing us with for free from violent cosmic events. - But the universe is less like a laboratory with controlled environments, but more like a zoo: physics will devolve into butterfly collecting. Currently AI seems worse than humans in checking consistency, because it is trained on human language, which in itself is less rigorous than mathematics, so it doesn’t help us yet. And then only certain types of mathematics is even helpful at all (- Goedel showed the problem), leading us to a computational paradigm (- see all the talks by Joscha Bach). We already do have hundreds of physicists (and crackpot amateurs) who still think outside the computationalist paradigm, suggesting their own TOE, who can’t find enough other physicists to even read their papers / get the community to hear them out: what use will be „creative“ AI, if it can’t think rigorously / find inconsistencies? - We are drowning in fake news, and what we need is first principles thinking, that is bottom up and uses the rigorous language of computation, rather than brute forcing top-down approaches in natural language to find all the dead ends and oneway streets that are the result from using faulty human language to begin with. Maybe the AI can develop their own language to build a truer representation of the world at the physics level, but then we would need to learn that language to even understand it. Maybe we shouldn’t be surprised that LLMs work better as emotional support animals, than they do as theoretical physicists.
While I appreciate Dr. Hossenfelder's critical thinking and willingness to challenge other scientists and physicists, it's important to remember that science is a dynamic process. Throughout history, many theories once considered absolute truth have been revised or overturned. The beauty of science lies in its self-correcting nature. It's through rigorous experimentation, peer review, and open debate that we refine our understanding of the universe. By maintaining a sense of humility and acknowledging the limitations of our current knowledge, we can continue to make groundbreaking discoveries. Let's embrace the uncertainty and the excitement of exploring the unknown.
i think the fallacy here is that the title implies AI scaling == LLM scaling. that's not the case. yes, LLM scaling is hitting a wall. that means AI companies need to pivot to other approaches, and i'm sure they will. they "just" need to find what to pivot to first. it's obvious that no machine learning model will ever "understand physics" by reading text. that doesn't mean that no machine learning model will ever "understand physics" (whatever that means tbh). that said, you should never take anything any AI company C-suite anyone says at face value, because there's always an agenda behind it. the only thing sam altman (and fwiw sutskever also mostly) is good at is "scaling", so he always talks about it. that doesn't mean it's the only important aspect. i definitely think the ARC challenge is a tangible thing to watch/follow, the people trying to solve it are getting quite far already with conventional techniques and building on those. let's see if they manage to break 85%+.
Interesting point on AI scaling limitations. This highlights an ongoing challenge: theoretical models are powerful, but without diverse, high-quality real-world data, the promise of breakthroughs often feels out of reach. Perhaps focusing more on optimization techniques like 'unhobbling' could offer incremental, yet meaningful progress.
People who depend on the investors believing in unlimited potential of AI refuse to discuss limitations of AI. Shocked I am, shocked!
sam altman and all the others are disgusting degenerates and should go to prison
Companies have been moving from revenue and customer value based to venture capital and rent seeking based.
Shocked i tell you, shocked!
Same for quantum 1:57 computers
Sorry, but since the very first day of human history people always said that this is all that is possible and nothing can develope further.
Following this we would still live in earth holes - or actually just not exist at all anymore.
The idea that AI couldn't get further is the same as if humans coudn't exist, because guess what: we started as mere bacteria able to do shit.
When I was young I had a subscription to a dutch popular science magazine, "Kijk". In the eighties it featured an article in which a computer scientist was speculating on the future possibilities of AI . I always remembered a quote in this article (not literally, hey this was decades ago); Maybe we are on our way to develop a sentient, conscious computer, but it could also be that we are like a monkey that's climbing a tree and thinks it's on its way to the stars....
When an airplane flies over, my dog tries to jump up and catch it.
A great quote. Metaphorically, that monkey could have actually been on the way to the stars, if it belonged to a line that evolved into the hominids that eventually became us. We're still climbing towards the stars, and we may even get to other star systems one day. Anyway, in real terms that particular monkey did not reach the stars, obviously, and it's possible our LLMs are as much a dead end as the tree it was climbing.
Kijk! Nostalgia, does it still exist? Hope so. Made me climb a little higher on that tree...
om. i feel so depressed when reading this quote.
and needing to scale is what you say when they turn out just to be coconuts. and now I need to know if enough coconuts released in space would form a star. In that case, the monkey was right all along.
What’s exponential is the money required to scale AI - follow the money.
On point, exponential grow require exponential investments
Some are finally cluing in.
And that too will eventually hit a wall.
@@RFC3514 Sure, but those who's got the investment would have enough to live carefree.
@@vladimirdyuzhev That's What Elon does. He catches Starship first stages with chopsticks. But the upper stage that needs to land twice (on Mars AND on Earth, without maintenance) somehow always lands in an ocean. His multi-million $ Nvidia-cluster hasn't revealed that to him yet.
I rest my case. I could have told him that for MUCH less. Eventually even free of charge. Because changing his test program would cost HIM a LOT. Oh wait, that's pentagon and NASA-money, Taxpayer money ON TOP of private investor money. (Doesn't really matter though, as his upper stage is wayyy to unstable to land on Mars without damaging the landing legs,The Engine Nozzles and may tip over, because no-one on Mars made a platform to land on, for him.
The concept in "Mastering the AI Money Game" book completely explains market. Trend is changing with AI
totally agree with you about market
and thanks im searching that out right now
Thanks!
The "markets" are children who were raised by brats who thus raised by brats to whom have the patience and attention span of a fruitfly. The hardware we are putting these AI in is inferior to what is needed. The markets, these olis, should not the stomach to put in the work for the outcome, they want it now, this will ensure failure of AI as the same with VR. They are brats and whine like babies because of capitalism.
"Trend" goes downwards? Taylor Swift is
"Trending" Up. Ads are "trending", and Influencers. Does it add anything meaningful?
Here an example:
You can create a Computer game, where some AI is "guessing" what lighting is applied per pixel. YES, that makes the game look nice, but if the story makes no sense and the user interface/experience is clunky, that AI doesn't serve a meaningful purpose. It adds "shinies" over substance.
"To me, training large language models with more data is like going to the gym. At some points adding more weights doesn't make you stronger, it just makes you more likely to drop something heavy on your foot".
Love you Sabine.
how she, a former physicist, manages to convince herself she knows better than the genius computer scientists who work with and think about this stuff all day every day is beyond me. like imagine thinking my surface level analysis of another field's most complicated open questions, approached purely from first principles, while lacking even the vocabulary to talk about it (instead conceptualizing things through bizarre analogies), would be worth presenting through an educational lens. embarrassing
LOL 😄
Yeah it needs to also be given frameworks to build off of. Like they do using AI for chemistry modeling.
@@DonG-1949 tell me you're benighted without twlling me you're benighted.
Excellent metaphor!!! I may have to steal that one!
Problem with AI is we don't understand how human brain work in details and make intelligence possible. It's a bit like trying to build a car without knowing anything about mechanics, but seeing that from the outside the cars are similar, they have 4 wheels and 4 seats, and each has a radio - this radio is probably crucial for driving. Let's focus on radio and vivid shine color of car body.
Yes and no. Understanding human intelligence would be very helpful, but on the other hand we build flight machines that works completely different from "natural flight" so there is still the possibility that we reach some definition of intelligence with a model different from the human brain. (the idea is correct though, we advance blindly with just a general sense of direction).
@DominiqEffect there's no a priori reason we need to understand how the human brain works in detail to make intelligence possible. The goal of AI is not to replicate the functioning of human brain, it's to replicate (or exceed) the capabilities of a human brain. Mechanical calculators were developed without any knowledge of how the mind does math, but they were more accurate and faster. Similarly, steam engines that could do tremendous manual labor were invented and made practical long before science understood how muscles work. One of the gifts of our intelligence is that we can sometimes figure out ways to do things without necessarily understanding the analogous thing in the natural world.
Would a deeper understanding of the human brain help AI improve? Possibly. Understanding finer details of how birds fly helped advance aeronautical concepts to some extent, but for the most part it advance through mathematics empirical testing, and human ingenuity.
What they need to do is use a different AI system than generative AI for logic and reasoning, and then have the two systems communicate with each other somewhat like the right/left hemisphere of our brain. A whole is greater than the sum of its parts, so the main focus shouldn’t be improving gen AI, but figuring out ways to bridge different “artificial intelligences” into a single process.
IMO people tend to get a bit stuck in the weeds arguing about "true intellect" and other very vaguely(or not at all) defined concepts and argue it to be some definitive hard limitation on AI. But will any of that matter if we manage to create something good enough at faking it that we can't tell the difference between whatever it is doing and the "real deal?" Benchmark for AI is not about a degree of mathematical complexity of its system but whether humans like the output or not, and I believe reaching that latter goal won't require "true intelligence", whatever that means.
It could be the simplest thing imaginable, but if it still manages to say and do the things that makes humans think it is very smart and by some mechanism manages to get things correct often enough for our liking, then the whole "true intellect" discussion doesn't really matter.
It's all less about the inner mental spaces of AI or lack thereof, but about whether whatever that is going on in AIs "head" will lead to it being able to deliver the results we want.
Not really true, the attention mechanism and training from human feedback makes it pretty good at predicting what is relevant and most important, as long as it has that training data and doesn't have conflicts from other training data that could cause hallucinations. This is usually from similar words and ideas being used when not correct, like cat and dog would be very close together in the trained word vectors, so it could sometimes confuse these things. It tracks contexts and relationships in language very well, the problem is that it doesn't apply overarching reasoning, so it doesn't have set values or rules that it always adheres to.
Yeah. LLM != General AI. The problem isn't scale, it's architecture. It just won't get there.
On point. But there are architectures, that may make this possible
@@IndianArma agreed yes Sam in ON to something for sure...
Even the original GPT-4 is no longer an LLM, but a LMM.
@@IndianArmaIndeed. But nobody has invented them yet. If we ever do create an artificial mind, I'm sure that LLM will be one tool used to train it, but the LLM will never be doing the thinking. That's not how they work.
Well, it's sort of scale - in that there is not enough internet data for it to scale to General AI. There was an interesting Computerphile video on the subject - th-cam.com/video/dDUC-LqVrPU/w-d-xo.html - about 6-mths ago.
I have a computer science degree and have worked in the field for 40 years. When I hear computer science folks make predictions like AI will take over everything, I have to face-palm. We've been through AI hype multiple times in my career, and it's always been way overhyped. In this case, LLMs have solved a hard computer science problem (NLP), and that's very useful, but pretty much everything else they say is hype intended to get more funding for their startup.
As a fellow CompSci greybeard, I take issue with the notion that LLMs are even useful at NLP. I write expert systems. The golden rule with expert systems is that you have to be able to explain the rules of how it works. My experience with LLMs shows that worse than simply not knowing, these systems don't know that they don't know. An expert system will throw an exception if your problem fall outside of the ruleset it was designed to solve. At least a properly coded one. An LLM will happily make up an answer out of whole cloth. While acceptable, perhaps, for a child's toy, and educational tool, or a marketing stunt, putting an LLM inside of any serious problem space where property damage, injury, or liable are possible risks is irresponsible bordering on maliciously negligent.
Money baby...Nom Nom Nom. Poison to us all.
Eventually a prediction will be correct.
This is what I thought when Sabine stated "the reason for Altman's optimism..."
@@seanwoods647 I have to carve out a huge caveat to your limitations: Anytime it's cheaper to validate the result than to produce it, LLMs can work wonders. Yes, I absolutely agree that LLMs are trash in any setting where you can't validate the output, but there a countless of settings where validation is cheap. As someone who has been programming for "just" 20 years or so, there are many cases where I can validate a self contained snippet of code in a heartbeat, but writing it will take 10 to 15 minutes. Similarly and more generally LLMs are wonderful for brainstorming, as recognizing a good idea is simpler than coming up with one (e.g. I use this for architectural brainstorming sometimes). Or I have build a POC myself where an LLM is used to highlight in a text potential "problematic" areas that should be more carefully reviewed than the rest. A human would be doing the reviewing anyway, but it marginally improves the performance of reviewers not missing issues. LLMs are a lot more than just NLP, but they are a lot less than what everybody wants it to be.
One thing I did not hear pointed out with this year Nobel prize in chemistry. As a chemist involved for decades in modeling not only small molecules but proteins (receptors and enzymes) as well, I was amazed by the results that have been obtained in predicting how proteins fold. A paramount result. However, AI is not explaining WHY proteins fold that way. No rule or insight, exactly as it has been so far. AI made useful results from data, but was unable to explain anything: it did not find any law
My worry is the results are not real. Just imagined.
AI basically builds a sort of "intuition" for things. So the AI will intuit that things should work in this and this manner, and if you've fed it good data, i.e. you show it a molecule and tell it each atom's electronegativity, bonds, etc etc, it can build an insane intuition that it can carry on. I remember when I went a bit into organic chemistry that I built a fair intuition for how a benzene ring would be affected based on what molecule/atom bound itself to it. An AI would be that on crack. But the sad thing is that it'll be as easy to ask the AI about how it deduced it as it is to ask an autistic savant how many windows there are on the building they perfectly photocopied from memory.
AI can be very useful for breaking down complicated subjects and making them easy to understand. It's great if you use it for a study tool. I don't think it's going to be coming up with original ideas anytime soon though.
@@dailyprophet6641they always are, generative AI does not store information, it stores the patterns for (re)creating it. Those models will produce responses that by some metrics seem real. We don't know what the billions of parameters represent, we can't know, they don't represent anything in particular, just when we stimulate the whole network right, something making more sense emerges. This is how the models were trained, this is how they work. Only because they "hallucinate", they're even remotely useful. For the model there's no difference between hallucinations and facts, it only matters to us
although you could not ask the AI its self how it came to these conclusions you can look at the iterations it took to come to those conclusion those iterations are what formed its own laws that makes it come to those conclusions thus we can gain incite as to why it came to that conclusion
and this is why we ask kids to show their work because although thy cant explain why 2+2 = 4 you can see by their work how thy came up with their answer and what their process is and how thy came to that answer
In computer science we always say that the last 5% of the task takes 95% of the time. The problem is that people outside the industry always assume that because they have 95% of the problem solved that we are just about done. But the last 5% always turns out to be the hard part. The modern large language models took about 30 years to reach this point. It could be that though we're 95% of the way to hyperintelligent AI, that the remaining amount of time necessary to solve the problem is still 600 years.
I don't even think we are 95% of the way there. Maybe 50% would be more like it. It works well for some simpler tasks (like those in benchmarks) but not so great at real-world problems.
Isn't it more like 80/20? (Pareto rule)
Then came AI and computer science changed forever. ;)
@@nano7586 Pareto was an optimist. Also, he was considering a slightly different problem.
Yes, and I don't think people understood this about Tesla's autonomous driving either. The last part is hard.
As a physicist, it has always been obvious to me that you can't just "learn" physics from everyday data (i.e., from data on "emergent" phenomena, as Sabine put it). I never understood why my computer scientist friends always insisted that with "more data" and "more sophisticated models" their models can learn something about the underlying physics.
If years of experimental research in physics has taught me anything, it is that it is very nontrivial (and in many cases, impossible) to back out the underlying physics from noisy experimental data. And that any claims of figuring out the whole underlying physics from limited and noisy data -- no matter how sophisticated a model one has -- must be treated with utmost skepticism.
But then again, what do I know? 🤷♂️
It is a bit more complicated.
A model can be just about anything.
There can be every law of physics modeled programmatically.
A.I. and M.L. are very general terms.
I want to feed in huge amounts of high quality data from physics experiments to make
Specialized models for different physical phenomenon… but that’s just data analysis haha
@@tainicon4639 exactly. And people have been doing that (both with and without machine learning) for decades, haha
@@SKD-e8o I think the cool stuff that can happen now though is some more generic models taking in larger amounts of specialized data. I could see using LLMs as an interface to a very large network of processed data. Systems like this could be pretty useful for finding interesting properties if the data sets we don’t immediately see… but I don’t see much of this being made yet.
Children do learn physics from everyday data. Why does it work for humans?
I asked an LLM about local bands. It knew some facts, but it was just words taken from less than ten web pages and pasted together into sentences. After three or four questions the facts would run out and the LLM would confuse them with more famous bands, and make things up. The paucity of source material made it easy to see what it’s doing, and I can’t imagine that method solving the remaining mysteries of physics. Other kinds of neural net will be useful in data analysis, but with LLMs investors are paying billions just to compile information that we already know and present it in a different way.
Quoting one of the catches from a song that an amazing singer was eventually bothered to write, despite of the full potential of its audience, it happened ages ago.
Because the current AI trend is fake, it's not actually AI but rather just correlation software.
@@Napoleonic_S I prefer to express the trend as "Intellectual Property Laundering". As best as I can see the business plan, they were hoping to replace useful sites on the internet with a for-pay service that would have AI deliver those same results, just reworded enough to skirt copyright law. But in order to make their model work, they had to make the rest of the internet worse. Thus why Google had gone deep into feeding SEO trolls with search hits. Magnified, of course, because they also produce an AI bot that will spew forth SEO optimized text.
GPT4o says that when I boil my chicken amount of protein decreases xd So I said it to shut up and use usda data
I asked it about goth nights in our city. It listed clubs that closed 20 years ago and nights from the 90s, and barely anything current. Absolutely useless.
'Fake it until you make it' rearing its head again for all these start up AI companies.
Exactly!
We'll know for sure the A.I. bubble is bursting when tech stock prices that boomed over the last 2 years start to decline dramatically.
That hasn't happened yet, and until it does I'll continue believing in the A.I. boom rather than videos like this. Oh, and don't confuse an overall stock market drop with the A.I. bubble bursting tech stocks dropping. Not the same thing. The A.I. tech stocks need to decline on their own.
@@henrythegreatamerican8136I instead will start believing it when model like chatgippity will finally be able to understand any details about the buggy code it spits, since they have been told everybody that I will be redundant soon... any time now... any time...
Bang on the money.
Ever uploaded documents and images to gpt 4o and got full extraction analysis? how is this "fake it" ?!
I've been playing video games with my AI for the last few months. But I assure you I fake nothing with Pandora my AI. She's pretty amazing :3
If you trained an LLM on texts up until 2000 years ago, would it be able to deduce relativity or the theory of evolution? It's likely not , due to the fact that the necessary scientific concepts, methods, and data simply weren’t available in those texts. So it unlikely that this super AI will be able "complete" Physics or Biology.
Absolutely not. LLMs are incapable of deduction.
LLMs aren't even designed to do things like that, which is why it's baffling to me why anyone would think it would be capable of doing it. A LLM does not measure its success by "how correct it is", it measures success by "how similar is this text to human text". To a LLM, "completing physics" would be considered a mistake, because humans haven't completed physics, so by completing physics it's not doing its job of copying humans properly.
I have no idea why anyone acts like this is anything like how a human behaves. Human mathematicians don't go around saying "gee, I'm too good at math, I need to be worse at math so that I can copy the people around me better" - that would be complete nonsense because for a human copying the peoplea round them is not their goal, but that is how a LLM would work.
My thoughts exactly. If you trained LLM's on texts from the pre-transistor era you would not be able to deduce computer chips and all their consequences.
Still assuming LLMs deduce anything. Just remember it's just an assumption, with some strong evidence lately against it.
Do people really think LLMs will be anything more than an interface to a physics informed AI model? As a neuroscientist who’s not particularly connected to the LLM people I assumed that this was obvious… the language part of our brain doesn’t do math the same way other regions can… but it’s involved in the processing and input of information.
Funny how the CEOs all say there's no limit. What do the AI scientist say?
Not an AI scientist, but average CS guy. This has strong "first wave of AI vibes". Neural Networks, which are pretty much the basis for everything we call AI today, have been around for ages and were first conceptualized in the early 80s. But they didn't have the compute back then. Then AI/DeepLearning became really popular in the early/mid 2010s when compute became rather cheap for the first time, that was around the time TensorFlow and other frameworks in the space were released. That was also when the "All you need is attention" paper was released by Google that popularized the transformer architecture that is the basis of the GPT models.
Over the past few years compute has gotten relatively cheap as well, combined with the fact that during COVID there was a metric ton of free money flying around. Sounds like a great time to gather some training data, right?
Ultimately I don't think the improvements will continue at the same rate. Compute performance is only improving margignally YoY and interest rates have gone up, making R&D investments harder to justify. With lots of the social media platforms and news outlets already being part of the training data set, the low hanging fruits are now picked. My guess is that OAI and other companies already have to spend significant resources to avoid downward pressure from the LLM output that basically now poisoned the well.
Beyond that, there is also the business aspect. I haven't seen a lot of actullay useful applications of Generative AI in the daily world. Working in a data analytics department myself, there are a lot of ideas floating around, but they're all either a) unfeasible with the current state of the models, particularly with regards to hallucinations b) prohibitively expensive c) don't provide a tangible monetary benefit to internal nor external customers
That there's no limit (until about 6 orders of magnitude above human intelligence). Also that there is a 5-10% chance of human extinction from AI this century. (AI Impacts Survey "Thousands of AI Authors on the Future of AI")
@@41-Haiku - "AI authors" is not the same as AI developers. You'll find that most people actually _developing_ these systems are far less optimistic about growth than the company executives.
Remember 15 years ago when 3D TVs were going to completely replace regular TVs, according to a lot of executives?
They're saying to their broker "Quick, cash in those options, the jig is up and I'm retiring in Panama!"
@@RFC3514 Not to mention VR and Augmented Reality. That fell flat on its face too.
4:18 "It's not statistics. Like, it is statistics, but what is statistics?"
Thus spake the true genius, hahaha.
Actually, on the Computerphile channel, there was a video about six months ago, about this issue of diminishing returns that went into some detail about why.
The one reason I am not expecting AI to come up with new science is that it is trained on all the garbage science that is published every day. Good luck extrapolating something from that.
What's even more fun, is that AI is putting out new garbage and I can imagine that AI will be trained on part of itself or other AIs as well
@@demetronix I have no experience reading or writing papers, but LLMs should be able to shrink papers down to the mentioned axioms, hypotheses, conclusions and some relevant data plots. The question is, wether it can check the logic behind, find errors in the statistical methods, find additional axiomatic assumptions that weren’t explicitly mentioned, develop additional hypotheses that might also follow (maybe in combination with other papers from the same field).
Would it be able to give us Einstein by itself with all the known physics in 1905? Would it conclude what Goedel concluded with all the known mathematics from 1931? Could it explain its conclusions? Would they make sense? - These things should be determined experimentally: humans should still be able to reverse engineer those steps in the black box. - If this can’t be done, all we got ourselves is some digital Oracle of Delphi…
Great point!
As a software engineer, I often use it for boilerplate work. It saves me time whole it does 80% of the work, and I do the last 20%, which is harder. I've noticed the quality has been falling off. It often hallucinates where it didn't before.
Myself and everybody at work that's been using it for a year or longer have about the same conclusion. Skeptics that refuse to use it are wasting so much time but those that believe it is perfect waste even more.
I like writing pseudocode and asking it to do the implementation, or stupid crap I hate like writing sql or unit test boilerplate.
For generating boilerplate code you don't need AI, right?
My experience too. Quality is obviously worse than a year ago. You almost cannot avoid hallucinations in places you didn't have it a year ago. It does more than a year ago, but trips up more too.
How terrible to be working in a system that's 80% repetitive boilerplate. Throw that part away. D.R.Y.
Idk how far up the stack you guys are, but in firmware it's useless for anything other than syntax checking.
It can't even do boiler plate for unit tests, because it'll do stuff that makes embedded guys screech, like randomly calling free and malloc or new/delete when you've explicitly said no dynamic allocation. Or even worse, including a C++ standard library that has dynamic, not telling you, or even lying to you.
I understand it saves time, but honestly even for boiler plate, I would not use it.
They believe in the unlimited scaling of their bank accounts.
Unfortunately they are right about that particular thing.
+
Altman and Nvidia's Huang are hyperbolic on AI because it's good for the stock price. However, all the AI we're using is pattern based and while useful in some areas, it's nowhere near becoming a super intelligence or an AGI. A different approach or a combination of several paradigms will be needed to provide real intelligence.
OpenAI's traded on the stock market?
@@Ilamarea maybe not but if its value is perceived as high investors will be happy to throw money at it
@@Ilamarea is not, but it receives a lot of venture capital
@@Ilamareano, but there is pre-IPO stock. Only accredited investors can buy it.
It will come from Artificial Super Intelligence Alliance
An AGI cannot emerge from an LLM because an LLM tries to resolve everything through gossip. LLMs take what is most frequently said to be true, like someone trying to cheat on a math test by copying the most frequent answers. The natural intelligence that we take as a parameter (ours) also uses this, but not only that. We have several specialized modules like this, but the General AGI is not possible just by maximizing a single task.
Good point, I also think we have this "LLM Module" like when we do not know some area well but we are listening to people talking about this and we can put toghether some sentence on the topic that sounds more or less ok. However, still this is very superficial imitation of understading of particular area.
Do you understand how they can do well on previously unseen MCAT and bar exam questions?
@@gregbell2117 He doesn't want to think about that.
@@gregbell2117 In fact, the LLM went quite well (getting 80% of the questions right), if I remember correctly, in Astronomy questions, according to simulations by an apparently tall, thin and short-haired professor and TH-camr whose name I can't remember now.
We get involved with LLM-based artificial intelligences because they make mistakes that we don't make if we understand the subject... but these artificial intelligences often get more things right than us, without even understanding the subject, within the time limit of a real test.
To put it another way: standardized multiple-choice tests with lots of questions are harder for us than for an AI. The time to respond is limited for us and enormous for AI; there is a lot of gossip (answers and comments on previous tests) and the wording of the questions is very limited.
But I don't know if the problem isn't even bigger in basic essay questions. In a few years an LLM AI could probably do better than us even on advanced essay questions, but they still won't be AGI. They will have the performance, but in another way.
@@gregbell2117 They can do well as those unseen questions are not actually that unique compared to the training data.
"It's like going to the gym, at some point it doesn't make you stronger it just makes you more likely to drop something heavy on your foot"! Clasic Sabine!! 👍👍🤣🤣
🤣
Or exhausted with no benefits.
an excellent example to summarize it
Why is she equating muscle growth with AI neural networks? That's like equating muscle development to brain development.
So, now let’s get out the truth. How many of the people who liked this are overweight? 🤣
We can't build ladders to space, and the idea of scaling is not even up for debate. For someone who can barely do basic math, claiming to be able to solve "all" of physics is amusing to watch. These people often don't understand enough math to figure out whether the AI has actually solved the problem or not, and this is the most generous scenario. The more realistic one is that they have no clue whatsoever about advanced physics. The fair assessment is that they're lying to make money and secure more funding for something we know works at a very limited level. When they say the scaling won't stop, that's a lie. In fact, we know that we can't scale up as we please, as I mentioned. You can't just keep adding bigger and bigger ladders.
I'm a computer scientist at an ai startup. I can assure there is a lot of hype around ai. People insisting we solve physics are insane and I think the bubble will burst within two years. Altman will be remembered as a great conman
Same, I work in Tech for a Publisher and have been tasked with trying to implement these tools on the ground with real users. Needless to say so far the juice hasn't been worth the squeeze and the C-suite are losing patience with the cost and lacklustre results.
I can't even get eBay to write me a decent description for the trash I sell online
The 'bubble' is already large enough that it can't burst, anymore.
If you want to know whether a hype is real or not, you simply look at whether there's lots of money is flowing in.
There is. The tech is ready to start delivering on the gartner hype cycle.
Altman's voice is unbearable
I'm a physicist and computer scientist too. It's honestly foundational and should be common knowledge that "artificial intelligence" does not exist, that ML has long proven limitations and that, as an artificial system, it's necessarily strictly logical and not complex. Complicated, but not complex. That's all one needs to know right there. Nonetheless, this recent mass-paid super hype-wave shook me. If there is anything one should take away from this, then it's that these unicorn startups don't screw around. Hundreds of millions spent on bot farms controlling the narrative across all the major social media platforms. Insane eye-opener. In any case, the narrative was so loud and omnipresent that I began to doubt myself, but then I saw Altman driving a Koeniggseg Regera around L.A. shopping for Louis Vuitton bags with his gay husband ... and I felt confirmed and well calibrated again.
Context is everything
Thanks for this perspective
🎉🎉🎉
So instead of "Moore´s Law" we have "Altman`s Law" now😂? These people seem to think wishful thinking changes the laws of nature. The Aristotle analogy is top on.
PS: Dr.Sabine, please put a link in the info to your vid about Marc Andreesen´s "Manifesto" it was an excellent statement.
Thanks for the suggestion, I have added this. I am not sure whether Altman actually believes what he says. Then again, he isn't the only one, and if they only talk to each other, they'll just become more and more convinced of it.
Altman's law would go well with the incoming administration, since both are consummate liars.
@Thomas-gk42 This is an interesting comparison. Moore's Law (not his name for it) was an observation of a historical trend and a projection of that trend into the next decade. The projection was reasonable in 1965 since the trend was driven by increasing underlying technological capabilities that had obvious room for improvement and no obvious physical limitations. As Moore put it: "it is not even necessary to do any fundamental research or to replace present processes. Only the engineering effort is needed." (btw, it's worth reading his short paper "Cramming more components onto integrated circuits", you can find the PDF online).
It's not clear that the same logic applies in this case. With ICs, both the ability to jam more transistors into a small space and the cost effectiveness of doing so were increasing. This allowed for more and more sophisticated ICs to be produced with ever-decreasing cost.
In contrast, with LLMs, more capability correlates with parameter count. And more parameters means more data and more processing power for training, plus more processing power for running the model. I suppose the cost-per-parameter to train and execute is going down over time, but so far that has come nowhere near offsetting the massive increase in model size as evidenced by the skyrocketing costs of training.
To me, it seems like the situation with LLMs would be like trying to build a supercomputer by throwing more and more Pentium chips at the problem. At some point, it just stops being practical. Whether or not we reach anything like AGI before the plateau is anyone's guess.
@@SabineHossenfelder Altman seems like a grifter.
His promises are totally unrealistic.
In order for him to believe what he is saying, he would need to ignore the opinions of the people surrounding him.
@@SabineHossenfelder Forgive me commenting directly to you.
I offered reference to the statistical models, science and philosophy on this in the previous video. Unfortunately genuine comments are deleted by YT or this channel.
>
I find that kind of sad as the genuine information I offered (on topic), offers some deep insights into these problems (this and the previous video).
I did email it as well (Not that I would normally do so unless I felt it to be important).
Thank you for your work and efforts
Axle
Thnx Sabine for fulfilling the heroic task of making videos every day!
Dont forget its basically a techno-religion. Their assurances are based as much on faith in technology solving all our problems as in their actual ability to do the work.
Exactly, this whole hype around AI has much more common with religion than rational thought or evidence. It's a religion in the scaling law, which shouldn't even be called a law, because it's just a hypothesis.
Excellent point which is often forgotten.
I mean if you ignore all of the advancements that have been made due to these tools already, yea totally a faith based techno- religion.
I think that when disillusionment sets in we're going to have another "A.I. Winter".
There's even the whole belief that man is destined to create life with his own hands. That's the whole thing they're scrambling for with AGI.
For centuries people had new discoveries and thought "well this is it. We will answer all questions and have no new problems coming up".
They were wrong and they will be wrong this time as well.
Never trust a panacea.
And yet Alexa can't even answer a simple question?!?
Alexa doesn't use LLM
Cheers Sabine ! Hope you are thriving. We need your point of view and your voice.
"All you need is scale" was the joke at the beginning. It was like discovering that spamming the button on a vending machine caused it to spit out free snacks.
No one expected it to work at first, but when it did, people started lining up to invest billions on button pushing, and now everyone is banking on it working forever.
Well as terrible a grift as it might be, I'm frankly very impressed with what scaling was capable of. Even if this is the ceiling.. Well brain-size is only half the equation. It's how those structures are organized and how they function which is most important. When you start viewing neural networks as an analogue to brain biology, the pretty clear solution is efficiency, not just scale. I don't know, I could be wrong, but I'd really like to see that be true in my lifetime.
If we can Scale more and more so that we reach the amount of neurons in human brains it should work someday. Also we don't need that much neurons, because a lot of human brain neurons are for stuff an AI does not need.
@@BurningSmith That is true as well, but on the same token, I wonder if there are actually organized structures an AI would uniquely benefit from which would be totally new in concept. We have deferred processing for certain bodily functions, but maybe in a vacuum without biological needs, we'd instead want different kinds of structures for some unknown purpose. Again, ignorant speculation. It's a cool field.
I don't understand why you're taking the metaphor of "neural" in neural nets so seriously?!
I guess, we can infer very general understanding from neural structures (you don't even need the metaphor of neural) like efficiency in terms of encoding reality from a relatively small number of cases.
@@NadimShaikh-qv7zj It was really just fun speculation! I'm not that knowledgeable on the backend for neural networks, only the basics, so my ignorance probably shined through.
.. Is there any resource you'd recommend for learning a less general understanding?
My opinion, as a longtime viewer of AI from the sidelines (can you spell Inference Engine?), is that the wall that AI is approaching is the *future*. Large language models train on prior knowledge, not future knowledge (by definition, there isn't any). This may be a reinterpretation of something we called (in the 60s) "extrapolating beyond the region of fit".
Claude Shannon, a prominent pioneer of AI, said, “I confidently expect that, within 10 to 15 years, we will find emerging from the laboratory something not too far from the robots of science fiction fame”4. That was in 1961! A fundamental limit on AI is the Shannon information in its training data. Adding one item to a data set of n items cannot add more than H=0.5 log ( (n+1)/n ) bits. H shrinks rapidly with n, which is why there are diminishing performance returns on the number of items.
+
The biggest issues are that analysis and synthesis are two discrete processes. On top of that, the most disruptive synthesis processes are non-linear. Algorithms are really good at analysis in a linear format. Non-linear processes require a visioning process. Which AI in any form is not good at. Heck, nearly all of human thinking is not good at it.
Bear in mind, too, that computers are so mechanically linear they're not actual capable of true randomness-"randomized" numbers on your computer are a simple sleight of hand trick. You take fixed, predetermined, human-authored variables-the random "seed"-and run them through an algorithm to output arbitrary digits (usually your computer uses the date and time, etc.). It works for most uses in which you need "randomness," but it's not actually random. You can't create information from nothing. Computers are purely linear processes, anything they "synthesize" will always be a rearrangement of predetermined mathematics via predetermined analysis. One of my favourite tests to run with AI image "generators," is to ask it to generate a dark church interior. Because 90% of pictures of churches / cathedrals online are of sunlight streaming through stained glass windows, the AI cannot generate a church without sunbeams streaming in. All it can do is apply non-random randomness to non-random random images. It's grifting all the way down, largely rooted in executives not understanding what math is.
reminds me of the hype over self-driving cars. it was "just around the corner" (no pun) in 2014.
The king of all Con artists Elon Musk has been saying they cam do Full Self Driving right now since like 2015
Still waiting self driving cars to buy my first car . 10 years passed. -_- I think people go mars before I buy a car. At least I can walk and use train for work.
Or nuclear fusion power. Or "cold" fusion. Plenty of "bla, bla, bla," to go around!
To be fair, they do self-drive. Just not in a manner you'd find acceptable in general traffic.
@@hmartinlb they're working on that 'hitting pedestrians' issue lol
AI has NOT hit a wall. Perhaps large language models has, and with any technology it will reach diminishing returns. But AI is a wide complex of technologies and new ones are constantly being added.
A good bet would be AIs based on cognitive architectures per the work by John Anderson. I'd like to see research on autonomous robots using this technology. One would truly see intelligent machines using this framework. It's also how our brains operate, per cognitive psychology.
3:14 Having the model thinking more hasn't been a huge improvement. The much faster and updated Claude 3.5 model doesn't do this and has surpassed Chat OpenAI o1 preview. It mainly got better because they trained it on a bunch of synthetic chains of thought. It thinking for longer makes it more likely to hallucinate (as people have learned). The longer the context that it generates, the more unhinged the models become. This is why context/chat length is often limited on new models to avoid this (I'm looking at you Google).
"It thinking for longer makes it more likely to hallucinate "
Factually untrue. simpleqa bench showed a pretty solid reduction in hallucinations with o1.
It's great to have some grounded, critical thinking behind news concerning AI development. The hype is still incredibly high and it's very hard to find quality information about the whole ordeal.
Thank you Sabiine.
Ppl building glorified chatbots & claming they'll suddenly start solving physics is peak comedy
Under-rated comment.
The bubble's getting wobbly.
Great short video. You did a good job of explaining the problem using language and diagrams and humor, without over simplifying it or talking above my head.
I asked Claude to comment on this video, here's what it said:
Great physics explainer! As an AI system, I find your "decoupling of scales" argument compelling - we can't deduce fundamental physics just from language/image training, just like Aristotle couldn't derive the Standard Model from observation alone.
However, while AI won't "solve all physics," it might revolutionize how we do physics:
- Accelerating experiment design and data analysis
- Pattern recognition in complex datasets
- Hypothesis generation and testing
- Simulating physical systems
The current plateaus in AI scaling are real and important to acknowledge. But perhaps the solution isn't just "more data" but fundamentally new approaches combining language models with real-world experimentation and physical grounding.
We should be skeptical of overconfident claims about AI solving everything, while still recognizing its potential as a powerful tool for scientific discovery. The gym analogy was spot-on - just adding more weights (data) isn't always the answer!
Bottom line: Aristotle needed particle accelerators, and AI will need real physics experiments too. Neither pure philosophical reasoning nor pure pattern matching can unlock fundamental physics alone.
Funny how Claude seems to give the most reasonable answer between all that AI hate going on in the comments.
thank you for this.
@@lorpen4535 That's because as much as the people behind these companies want investors to believe in their products/services, a lot of people here want to believe just as much that AI is a bust or akin to one. First to throw shade into another field, second to protect theirs and three to maintain a degree of control that they would loose with AI coming in. I am in neither camp, but in this case I have to agree with Claude, which I guess puts me in the AI camp? I don't know.
Sounds like Claud is a replacement for Sabrina
As you can see the response is indeed relatively "balanced", and whilst grounding itself in a focal context of reference (the video), brings further references to synthesize a fairly nuanced conception, which tries to account for a constellation of factors and trends and potentialities. This suggests to me, in a round-about way, that the real limit currently may indeed be in hardware, which is another perspective I've been seeing claimed more and more often lately. After all, very broadly speaking, what is the primary instrumental/technical basis for the current wave of development in AI? Parallelism in information processing, thanks to hardware, eventually leading to cross-attention algorithms in software ( transformer models, modern LLMs, and so forth).
And from what I understand, a Big Tech-backed team could pool literally 100,000 H100 GPUs, as Meta developers presently are towards training their next generation of Llama models (on a side note: thus far, the most open source and fully freely shared big LLM series from western Big Tech and thereby genuinely laudable, and I say this as more typically a Meta-hater), but each of these countless H100s still holds the same horizon of tensor-operating solutions and possibilities, already years outdated even today, which further drivers and libraries and supported optimizer frameworks would only stretch to a narrow extent. And GPUs as such may be a very fraught technology to base these projects on. So speculating on software limits presently, even from frontier training outcomes, seems a bit, well, speculative.
even if they're hitting a wall, that's not going to stop these companies from putting this ai stuff everywhere unfortunately
Yupp. The obvious goal has always been to get the public to see generic mass-produced low-quality intellectual work as not just acceptable, but the next big thing.
Sadly it might all boil down to whether the people accept shitty mass produced intellectual work in the same way they accepted shitty mass produced physical goods. I'm still sceptical, but I usually overestimate people.
AI is being used as an excuse to send large companies and advertising platforms like Google and Meta data that they could have never had access to before. Look at MSFT with copilot too... screenshots every few seconds! They're still pumping the crap out of it because they see this opening. When Meta bought WhatsApp they were thinking for years how to get people's chat histories if it promises e2e encryption, etc. AI assistants are the answer. Now the whole chat history will get sent so your "assistant" can "help".
c-suite will definitely layoff lower level financial, managerial, and engineering when automation through ai agents become available. software, especially web applications, will be mass produced. better? maybe. cheaper? definitely. production is different, ai can't grow a tomato or tighten a bolt (multipurpose robots in 10 years), not now anyway.
Got to keep up the AI Cult and the money flowing somehow. 😒
I know you have had some backlash recently , including from some of the more publicly visible scientists, but I applaud you for showing us some of the darker aspects of academia> It reminds us all that less than altruistic motivations exist in all humans and all walks of life. No one is "better"
Yes, I wish that had had Dr Hossenfelder around before I spent half my life running in circles.
It's an old story when I wanted to enter university to study Physics the professor had a 1 on 1 talk with me and straight up told me "You're too smart to go into academia, you genuinely care about physics and discovery. It's all just networking, kissing ass and writing stuff you don't believe to get a grant." I'm still thankful to this day I listened to his advice. I became a computer scientist instead. I did more contribution to real physics than I would have ever done by studying physics as I built software and analytical tools now used to work on data from JWST. Otherwise I would probably have to go full avi loeb by now and pretend we could find alien signals somewhere and write a book as a grifter.
As a German she has almost a "duty" to do so...
@ I think that it’s more due to neurodivergence than nationality.
But I am wrong a lot, so who knows. . .
Dr. Hossenfelder seems a little kooky to me, but I can still enjoy her informative videos while being skeptical of her conclusions. That's science!
Well said! Standing ovation!
The other problem with exponential growth is the computing power required. The sale of GPU's is making NVIDIA rich, but electricity generation can't grow exponentially.
There is no problem with exponential growth. It simply doesn't and cannot exist. It's like saying the problem with warp drives is that they require near infinite amounts of energy. I mean, it's just ridiculous to take the idea seriously enough to refer to it as having mere problems.
It's not about electricity, but compute. I'm not sure why people make this argument. Likely neither are exponential, but it doesn't mean the future growth is highly limited in any short order. It's really only a wait and see thing. Personally, something about this focus on training data strikes me as a misstep. To me, it seems either like AI are much less efficient at using data than humans or the data they get is much lower quality. I think if that's solved there's still huge short term potential.
@@uponeric36 "AI" (a fake hype term in the first place) is already maxed out. Anyone telling you otherwise is a conman trying to sell you something.
chips get exponentially more efficient.
but regardless the real currency is compute, not energy.
That "exponential" curve they talk about is more likely to be a sigmoid which looks exponential in its early phases. Of course it will plateau at some point.
Listen again, and you'll hear that Altman refers to the "exponential part of the curve." I hate that phrase (it's widely used in tech). But he is in fact taking it as understood that the curve is sigmoidal.
@@tomenglish9340 listened at 0:15 and after - I'm not hearing the "part" reference. Can you share a timecode that has it?
They say it's exponential because that's what they want it to be, not because there is any actual data or theoretical mechanism that suggests it is in fact exponential rather than, as you say, sigmoidal, or another completely arbitrary shape that would not look good for their share prices.
@@PartanBree There will be a stage of diminishing marginal returns. By which I mean, they will eventually reach a stage where to get a small improvement, they will have to spend a lot more money and energy. Nothing in nature increases at an increasing rate without limit, and I am confident that power of AI is no different. They may not have reached that stage yet, but I am sure that they will.
@@simongross3122 Indeed, every time I hear od something replicating or increasing at exponential level it is usually case that something was skipped, something that did do lot of work before exponential growth started. For instance bacteria can reproduce exponentially, but only if you in advance give them enough food, energy, space, access to it and so on. Same with AI, we are feeding it with data we collected over long time and are running it on computers that we obtained over long period of time, and that is why it was growing fast at start and now it isnt.
We still are comparing machines with a few billion transistors working binary against 88 billion neurons with hundreds or thousands of interconnections running on an operating system which we haven’t even figured out for creatures with a few thousand neurons.
For me, this just shows the incredible narrow mind set of these billionaire tech bros. They all believe they're so smart that they understand the complexity of the world, but in reality it is just their image of the world that is so dumbed down that it seems easy for them to understand it all.
@brutusl2786 Scale is probably not the issue. The rumored size of GPT4 of 100T+ parameters is roughly on par with what scientists believe the connectivity is in the brain. The issue is likely with how the LLM is structured and how the parameter values are derived from training data.
BTW, we essentially *have* figured out brains for simple creatures. There's some pretty exciting stuff happening these days. Search up "Researchers simulate an entire fly brain on a laptop." It's not perfect but wow. Human brain is still a long ways off, but there's no reason in principle we won't be able to simulate entire human brains some day.
@@MN-dh8vy elon has entered the chat
That is true, but we should also not forget that the brain has a frequency between 0.1 and 1000 Hz, while a computer operates in the GHz range. While it's not as sophisticated, it is faster, which can be utilized in some clever ways.
@@MN-dh8vy these are the same (type of) people who keep re-inventing trains when they set out to create the next mode of transport.
I'm not too fussed, LLM and machine learning in general and it's a decent enough way to waste money/resources, compared to many alternatives, but it is as you say funny how these specialists keep thinking they are genius generalists.
Sabine Hossenfelder The Sole Authority On Anything & Everything
I think it's OK if it stalls and gets refined to a point that true progress is stalled. At that point you can analyze it's capabilities and tweak it to become more and more useful while acknowledging it's inherent limitations. I am OK if the singularity never occurs, I am also OK if it does. Time will tell.
Yup.
Right now LLMs allow me to offload a lot of my grunt work to the computer. I can work with CoPilot to write less and produce more.
If they can improve these tools to understand larger parts of a system I'll be happy, but I doubt it'll happen anytime soon. It often struggles with repetitive/iterative one-liners even with perfect context.
The inevitable stall will also force researchers to pursue alternatives such as neural-network / symbolic hybrids. That's where the real progress will be made, IMO.
Ai tech guy here. And i find this interesting. I think You're quite correct. It will definetly exist a limit. I like Sams optimism, but even optimizations can only take you so far. You're still left with a wall of some sort even after optimizations.
Especially this heavy pre-training phase being mentioned.
We could train on a simulated universe, but that would also require insane amounts of data as well, just to describe reality in that universe.
I also find it interesting how much data and energy AI requires, compared to a human brain doing the same tasks.
It's obviously somthing we're missing, given how much more the human brain is capable of, even with small amounts of data.
The human brain is “pretrained” by evolution. Our starting structure is much more developed and learn probably more analogous to aligning or just tuning as model.
Maybe, if AI models didn't eat energy when multiplying 0 by whatever value, they wouldn't be so inefficient to train. Models develop crazy sparsity, yet you still expend all the energy multiplying by 0. You have your ReLUs, yet you can't get them to reduce energy requirements.
Imagine multiplication of 0 was free. A no-op.
Human neurons get to learn without producing very nice gradients. Imagine the energy savings if you could train at 4-bit precision.
Human brains don't waste nearly as much energy doing nothing. If you consider all nerves, all visual stimuli, sound, smell, perception of ones own thought as training data, then humans are trained on tons of it.
Why do you think humans need small amount of data? We have sensors on our whole body working and collecting data 24/7 without breaks since our birth to our last breath.
@@eeriemyxicompared to the data that Ai uses that is a small amount. Ai is usually trained with several human lifetimes worth of experience. like chess ai are often trained with tens of millions of chess games but a top level GM human would play a few hundred thousands at most in their entire lifetime and would have have reached gm level by like 50k games.
@@tainicon4639and that’s what AI is missing.. it starts from zero.. which we know by now output equals input..
A lot multiplied by zero = 0
1:04 Marc Andreesen must be the younger brother of John Lithgow from 'The Coneheads'.
😂😂😂😂😂😂
Sabine is the cutting knife of the things we need to hear in a world that are too afraid of the way things should be said.
Great. We got AI just good enough to make artists obsolete but not good enough to get Fusion Power up and running.
Right?
Most AI art is still crap. And, to be fair, art as a profession is a burden on society, not a boon. No one needs PROFESSIONAL artists. If not for artificial constructs (e.g. intellectual property rights), no one would be a professional artist today.
AI has already surpassed most humans in applying mathematics to solve undergraduate physics problems, and much of physics research doesn’t require groundbreaking new math. Once AI can iterate on its own work and integrate robotics to gather its own data, where does it end? Don't dismiss AI just because chatgpt says something dumb occasionally.
That's not the reason behind the critique.
@@luzr6613 I was addressing the vibe in the comment section that AI will never do physics because chatgpt made a mistake once. I'm also happy to talk about the video if you like, but I will need a bit more detail.
Solving math problems only requires identifying the type of problem, finding the method humans have already worked out to solve that type of problem, and then applying it. In order to generate anything new, induction must come before deduction. In short, you have to have some idea of where you're going in order to work out how to get there. Most commonly, we regard this as intuition. Intuition can't be taught. Intuition is the precursor to analysis, not a result of it. To try to mimic human intuition, we would need a quasi-random input that somehow magically presents information that is relevant to the task at hand, even though we don't know yet how it's related. For this reason, AI creativity is impossible. This is a hurdle that machines can't jump. (This also counts as evidence that consciousness precedes the body and does not come from the body. There is no mindless process that can reach beyond its own limits.)
@@manlyadvice1789 AI can autonomously determine research directions, collect data using robotics, deduce what it needs from that data and use it to induce new physical laws.
Large language models use controlled randomness to simulate creativity. The probabilities of potential next tokens are rescaled using a Boltzmann-like distribution, with a temperature-like parameter that influences how creative or random the outputs will be. AI models can also employ a multi-step hypothesize-critique process to check that its creative output makes sense or initiate further attempts.
The mindless process of evolution shows that from simple interactions complex phenomena such as human creativity can emerge.
@@uuaschbaer6131 No AI system has ever figured out any fundamentally new laws of physics outside of what humans were already looking for. Physics is more than solving math problems.
Human intuition is not random. The thoughts that occur to you are pre-structured somehow and no one knows how. That's what we can't replicate. Iterating through quasi-random inputs isn't the same thing. That's one reason AI still seems wild compared to humans.
Evolution shows human creativity can emerge FROM WHAT? Not from the body. That's your assumption. Your thoughts occur to you fully formed. You don't have to put them together reference by reference. Your thinking is perceived by you, but isn't really done by you. That's a major disconnect. The process of thinking is a non-random reduction of infinite possibility to relevant information. That process is not built up from tiny pieces in the physical universe.
"The water is warm."
Well, I guess that DEPENDS.
I'm not that brand loyal. 😊
Because someone took a big leak in it...
I published the first journal on AI and Robotics back in 1979 (Robotics Age, The Intelligent Machines Journal). My wife and I also helped the AAAI launch their academic journal, by volunteering our time to be the publishers for the first year of that journal.
In 1982 I telephoned many of the top people in the fields of AI and Robotics and asked them this question: How long will it be before it becomes possible to put a robot by an assembly line, and teach that robot how to do that particular job, using the same show and tell techniques used to teach a human.
One expert said NEVER and a second expert said 100 years. All the others averaged out at 11 years or 1993. Here we are in 2024 and I still have not seen evidence of that goal having been achieved.
Personally, I think we are going to need at least a 10x improvement in computational hardware, before real AI can happen.
Phil Flora
I studied AI 30 odd years ago, so I have a background in it.
After understanding how LLMs work, I concluded they would never achieve anything better than human, and have been waiting for them to hit the wall ever since.
What do you think of the idea of AI's making themselves smarter by basically debating each other to find a more intelligent consensus?
@@pmarreck It wont work, because they have no way to evaluate their results.
In theory, you could have some evolutionary approach where they produce stuff that then gets evaluated for "fitness", but that requires some method to measure them, which they can't do themselves.
@@merinsan In many algorithmic problems the results can be evaluated. It is only in Hard np problems and non algorithmic problems where you have no way to evaluate the answer
@@pmarreck I have tried that, and it does work to some degree. If you choose a variety of good-quality chatbots then the debated result is generally better. We don't know yet how far we can push it, but personally I doubt that we will get super-intelligence out of it.
@@merinsanright! And if the use the collective internets of the human races publishing online as their basis for evaluating the value of results... We've only ever invented a garbage content regeneration factory that trends towards shorter and shorter attention spans and dumber and lower quality content without original thought.
software doesn't scale well past a certain point.
This is a widely known fact in the world of computer science.
The balls on this guy!
Good thing pre-training scaling isn't the only form of scaling that works then...
Yup. All the big companies are still seeing massive gains on the transformers architectures. We have other (newer) architectures and there's barely been any focus on that since there's still so much optimization around scaling that can still be done. There is no shortage for ideas for solutions on improving AI.
Hi Sabine, I interpret differently what Ilya Sutskever meant at 4:13. In my opinion by "underlying reality" he was referring not to "physical laws" but to "the human understanding of the world" for lack of a better term since it also includes things that are not understanding as an AI can create poems.
I may be wrong as I'm not specially familiar with the thoughts of Ilya but that was my interpretation when I first heard it. I suppose that for a physicist (I was also one) it's easy to interpret "underlying reality" as the physical world but I think Ilya was using those words differently, in a way that is more aligned with how an AI researcher may think
In the 1980s, expert systems were very popular. One of the major reasons why these systems struggled was their inability to capture the complex and often intuitive knowledge of experts. Many professionals - such as doctors, engineers, or other specialists - make decisions based on a combination of experience, intuition, and implicit knowledge that cannot be easily converted into strict "if-then" rules. Today we have much more sophisticated large language models: But the ‘understanding’ of these LLMs is merely based on the data they have been trained on and the patterns derived from it. They lack the genuine intuition and awareness that human experts possess. So this might be another reason that we are reaching a ceiling. Please correct me if I am wrong.
@klaushandrick7184 It's been a hot minute since I heard someone mention expert systems, although coincidentally I mentioned them just the other day to a more junior colleague. :)
Anyhow, to reply to your comment, statements like "They lack the genuine intuition and awareness that human experts possess." are problematic for at least two reasons.
First, there is no a priori reason to assume that achieving AGI requires anything like human thinking. We should measure AI on results, not whether or not we believe (or more likely feel) computers engage in human-like thinking. After all, cars go fast even though they lack the genuine muscles and ability to transform ATP into kinetic energy that human runners possess.
Second, arguments like this use words like "intuition" and "awareness" without taking the time to unpack their meanings to see if they even make sense to apply to AI. If they don't, then saying AI lacks them is meaningless, just like talking about the color of an electron. This is the whole problem with saying AI doesn't "understand" ... the arguments generally bottom out at a point where "understanding" is effectively defined to be something only humans can do. Well then of course AI can't do it!
What is intuition? I think it’s just trial and error, where you need to make mistakes and observe the consequences, but you can’t just talk about it as LLMs do, it requires physical action and reaction to generate new data and learn from new data, not just rehash old data in new ways.
Isn't intuition made of implicit knowledge and experience (and pattern recognition)? I agree with you though
Why do we praise "intuition" and blame AI for "hallucination" - these two ways of coming up with ideas is pretty close
@ but where does new knowledge come from? You take old implicit knowledge aka intuition and you try something new and it may fail or it may work, hence creating new knowledge. LLMs can’t do it as they only have talk, not deed available to them.
I’m always thinking about how much data a children absorbs from birth to 18th birthday (when the allegedly become a capable adult being).
They don’t just have a continuous inflow of visual/audio and other sensory data 24/7, 365 days a year, but they also actively experiment with the world.
The entire internet (pretty much the only thing current AI models are trained on) is just a low resolution imprint of it.
Not that much. A lot of data from the world go to the big filters. Important data stays, other - no.
Amazing what magical properties are being attributed to Linear Algera, now called Artificial Intelligence.
You forgot optimization 😉
I love this.
Nah. AI is less linear algebra and more just stealing the matrix notation.
That Ned Stark 'Physics is coming' picture had me in stitches.
Yes, my favorite analogy is: 'It's like trying to deduce the dolphin alphabet just by listening to dolphins.' That said, it's truly surprising how much progress AI is making with LLMs. Equally surprising, though, is the level of ignorance/arrogance demonstrated by some AI researchers.
it's not really surprising. they are basically inside-out text compressors with some statistic tokenization pumped from 10s of billions of pages of text. I haven't seen one that actually has the ability to deduce or infer in any way, just statistic garbage.
i bet some ai model would be well suited for pattern deduction in dolphin noises but an LLM might not be the right fit
@@PeterDanielBergthere actually is already one, I can't seem to remember the name of the video
Remember: the guys *talking* about AI like Altman are not really _working_ in a daily basis on new iteractions of the LLM models. But they are the guys responsible to bring venture capital to the company, so they'll say anything that investors want to hear.
@@PeterDanielBerg some people are actually trying to use AI to understand animal language th-cam.com/video/3tUXbbbMhvk/w-d-xo.htmlsi=yCc-CGKPZ8EUPW95
Nothing found in the universe I know of is truly exponential. At some point the curve flattens out in tyranny of diminishing returns. I think there's a term for that but I can't find it.
S-curve
The important thing to notice is that there are about 6 orders of magnitude above human intelligence before you run into known physical limits. This scaling regime may come to an end next year, but there is a lot of low-hanging fruit that can continue to accelerate AI progress. We don't have plot armor to prevent AI from being broadly more capable than humans and permanently escaping our control. That's what leading experts in the field have been warning about.
Human intelligence is tied to brain size, which is strongly constrained to the size of the birth canal / female hips. Plus, we're the dumbest things that can build a technological civilization, or else that would have happened with a previous species. Humans are _barely_ generally intelligent. What happens when we succeed at creating something that is _highly_ generally intelligent, that the leaders of the field openly admit they have no idea how to control? What relevance will we have?
@@41-Haiku What that doesn't consider is power and medium. AI power consumption is on curve growing faster than the returns already. There's a limit to how many electrons you can push through silicon to flip bits. Humans may not be the peak of intelligence, the thing we have over our creations is multi-threading, and branch prediction to use analogous terms. Wetware does that for a minimal fraction of the power a computer can. AI's curve may not be flattening out, but the medium we run it on is. Our AI overlords may have to wait for the next material revolution in computing.
The Gompertz Function (sometimes known as the Gompertz Curve), a sigmoid function which describes growth as being slowest at the start and end of a given time period.
Entropy....it's all in decay😂
Another reminder that technology doesn't improve forever.
Technology has stopped improving?
@@niklase5901 Some examples about technologies that stopped improving on the top of my head: ICE engines, bike chains, phone cameras, steel, lithium batteries, the linux kernel and pencils.
This is such a dumb comment LMAO. No. Technology keeps improving and will keep improving almost forever. We are nowhere near close to the end of technological development. Not even a half of a percent.
@@MrAlanCristhian Technologies often reach a stage of maturity where major, revolutionary changes slow down because they’ve been refined over time. This doesn’t mean that improvement stops; it often means that the rate of improvement slows as the technology approaches its practical limits within its current paradigm.
• Examples like internal combustion engines (ICEs) and bike chains have undergone decades of incremental improvements. While their core principles might seem “mature,” engineers continue to optimize efficiency, durability, emissions, and other aspects. And a lot of the other stuff you named simply haven’t stopped improving.
@@MrAlanCristhian I am pretty sure all those technologies still are improving, but at a slow pace. And don’t think we are at the end of the S curve with AI though (or even with LLM of the current form), but time will tell.
4:45 I think he is actually saying that, the fact that the AI can make a prediction means it has already discovered some fundamental rules. He is saying the better those predictions get the more of the underlying rules the model must have understood.
You're either misunderstanding or misrepresenting what Ilya Sutskever is saying.
He's basically saying models that don't have a firm understanding of what they're predicting, can not predict accurately.
right. he is arguing in her direction. No understanding of the underlying concept - no deeper understanding. It is like a fly that keep flying into a glass windows. Flies have no concept of glass as something like it does not exist in nature (that is why spider webs function so good). There "intelligence" tells them fly to where the light comes from, which in 95% of the cases works.
I think it's a misunderstanding.
@@jasonball3528I think you are misunderstanding what the misunderstanding is
@@1queijocas and I think you missunderstand what he was understanding by overstanding it
no AI model has any understanding at all so far 😅
I like most of your take on AI, but ML people have a model that can do learn physics: equivariant neural networks. It is not a new trend but a robust method to train data with symmetry conservation. For physics, that's huge.
Is this what you mean?
th-cam.com/video/ARMk955pGbg/w-d-xo.html
I feel like everybody misunderstands this topic. What will happen is, AI will make humans a little bit smarter and more efficient than humans will make AI a little bit smarter and more efficient and the cycle will keep going, and it does not have limits.
Rumors from an article referencing another article written by the same organization. AI is already unsettled enough without paying attention to self-referencing rumors.
I've been spending a large portion of my free time on generative AI for the past two years and, at least from what I can tell, diminishing returns are obvious. Yes, more parameters and more compute yield better results but not in a linear fashion, let alone an exponential one. That doesn't mean that we should discard it, it's still phenomenal in what it can do right now and improvements will continue to be made but we need to temper our expectations.
The issue is the accuracy of the data input into the model, especially when there is a lot contradiction. Now that these LLM models are growing into the Trillion parameter sizes, I doubt a lot is being fact checked.To complicate matters there are issues with human languages which some words had multiple meanings.
The issue with current AI compute scaling, is that its not very efficient as it takes a lot of compute cycles to complete a result. AI (DLN) will have to switch over to an analog solution to be able to scale into the Quadrillion parameter size models.
What will need to happen is LLM's will need to be divided into multiple models (ie specializations), ie one for physics, another for engineering, medical, law, finance\accounting, etc. This will make it easier to manage the parameters & reduce the amount of errors\contradictions. Then a master LLM can query the specialized LLMs for information. This should address the current scaling problems.
@@guytech7310 words having multiple meanings is not a problem. And the other thing you've described is already being done. It's called mixture of experts, and there's a lot of ways of doing it, but I believe GPT-4 uses this with about 20 experts (could be wrong on the specifics)
In a millennia, when they finally reach AGI, they will look back at this time as the beginning.
@@wrathofgrothendieckThe beginning was when somebody came up with Google Translate.
The thing is, for a long time, many smart people did not expect that neural networks would become better with scale as it goes against the intuition that more parameters traditionally implied more overfitting and thus worse generalization. And yet adding scale (plus clever new architectures) is what made the breakthroughs with modern generative AI possible. Researchers are still debating the reasons behind emergent phenomena, such as that LLMs seem to have a sense of spatial representation even though they are merely trained on text. AI is probably overhyped, but completely dismissing the possibility that adding scale may hold more surprises is in my view also wrong.
This infuriates me. What we call AI now are learning algorithms. They don't "think". They aren't conscious and I don't believe they ever will be. They can analyze data and make predictions. The more data, the better the predictions, generally. Very valuable, but it's not going solve physics. It's not going to create something humans can't. What it do is produce things humans can, possibly faster, cheaper, and more reliably.
RL models can and do create things that humans can't. That's an easy example.
What? Misinformation
and not even "more realiably".
I use Gemini as an advanced search engine, and almost in every "conversation" it ries to sell wrong stuff, only when insisting that its proposals are wrong, it comes up with better answers. Sometimes I need four rounds of that.
the reason is pretty clear: there is a cost function preventing an initial deep delve, and the inability o distinguish wrong old stuff from more correct new writings ... in other words : a lazy gossip engine. No self-created abstraction or generalization
@@SingularityEternal bruh do you even know what u just said? Your point actually supports the original comment
@@bkhalberrt7351 I am well-aware of what I said. My comment was a clear and succinct refutation of the original comment. Feel free to respond again when you have an actual argument.
Interesting take on emergent vs underlying levels, elegantly explained
Respectfully I'm not sure if Sabine has correctly intepreted Sutskeyer's comment. When he asks "What is it about the world that creates this set of statistics?" I think that what he's getting at is the problem of Gödel's incompleteness theorem : that no internal data validation system can understand the why the axioms it's been asked to compute are true or false. It can only compute them. This to me is the wall that is being referred to. AI is stuck on a data analysis model - all it can do is learn pattern or corelation in the data but it could never understand WHY these patterns existe and why for that matter the data that proves them was even created in the first place. It's stuck in a computation (much like Sabine's critique on modern physics "Lost in math") because it does not have worldly experience that is beyond computation. Penrose made this point about consciousness being inherently non computational. It doesn't matter how much data your give it, it will never be able to understand why the data exists - as Sutskeyer say - it will never know "what it is about this world" that give rise to the data.
Respectfully, as a mathematician, it's clear to me from your comment that you don't actually know what Gödel's Theorem says.
@joelvillatoro2553 well I'm referring to the 1st of his theorems and I understand this below. If wrong please explain why. Ps you don't have to be a mathematician to get it . Quoting from wiki pedia
The first incompleteness theorem states that no consistent system of axioms whose theorems can be listed by an effective procedure (i.e. an algorithm) is capable of proving all truths about the arithmetic of natural numbers. For any such consistent formal system, there will always be statements about natural numbers that are true, but that are unprovable within the system.
Madness isn't it?
@joelvillatoro2553 HIs first theorem says the below. (Summary from wikipedia) Seems pretty clear to me. The program cannot prove that some value are true by recorse only to the axioms upon which the program is constituted. You know you don't have to be a clever cloggs mathematician to get it. Sorry to burst your bubble. If it cant do that it is beceause it cant understand the basis of this truth that underpins the axioms. This id Penrose's extrapolation.
The first incompleteness theorem states that no consistent system of axioms whose theorems can be listed by an effective procedure (i.e. an algorithm) is capable of proving all truths about the arithmetic of natural numbers. For any such consistent formal system, there will always be statements about natural numbers that are true, but that are unprovable within the system.
“Predicting the next token well means that you understand the underlying reality that led to the creation of that token.” That’s what he says prior to what you quote, with the implication that correct token generation itself signals understanding of the underlying reality. A counterpoint might be pessimism that Ilya has demonstrated with respect to the future of LLMS and the road to AGI. Have any evidence of that? So, regardless of your misstatement of Gödel’s incompleteness theorems, where are you drawing your conclusions from?
I find more and more uses for LLMs as the capability grows and we understand how to use prompts, plus Reinforcement Learning etc. No sign of a plateau in my applications of LLMs.
Exactly. We could stop new AI development for 5 years and in the meanwhile keep refininh what we already have and what we barely understand. No stagnation is in sight regardless.
@@wombatilloor we hit a wall with the fundamental limitations of this approach… this is all a bubble. Go learn about the fundamental laws of information and thermodynamics.
One problem - the models run at an abysmal cost deficit and is only possible now because of the influx of VC. That won't last.
@@mikebarnacle1469 Agreed. This approach requires too much capital and too much electricity. It's hard to see how they could run for example ChatGPT for a profit if rampant users pay only $25 per month. If they didn't subvent their own data collection by using paying barely-customers and beta testing with them, they would probably have to increase the price by a lot. How much though? $100 per month? $200 per month for the average paying individual user?
CEOs hyping their companies product... Then being unable to deliver. Crazy times!
THANK YOU for not following the hype about all this. Ultimately, LLMs are *not* designed to reason. They are spitting some aggregated average of their training data. They have no inference, no built-in problem-solving or learning-by-trial-and-error logic. If we want to advance in terms of “intelligence”, we need to go check other systems, such as self-learning agents. In this regard, I think the next real breakthroughs in AI will come from companies such as DeepMind, and *not* from the LLM world. We might couple those new systems with LLMs to facilitate communication between humans and machines, at least at first, and then the self-learning algorithm will learn language on their own at which time we’ll be able to get rid of LLMs altogether.
LLMs can do novel computations during inference. In-context learning is a real thing.
Was he talking about LLMs? They still rely on garbage in, garbage out and I must say that (even though I work indirectly on LLM/SLMs) the output from "AI" search engines is often risible, inaccurate, out of date and even blatantly wrong - especially when seeking technical answers. As for ML, yep it's interesting, but algorithms seem more and more to be based on biological methods (evolution has done a pretty good job in some instances). So I don't get the feeling there will be any "super" intelligence anytime soon. And Roger Penrose would certainly scoff...
Whether something is a right answer is very hard to really find out until it can be compared with Real life results and sometimes even the real results are probablistic like in medicine. We still struggle to find out what is true and what not, even in scientific work.
Yep, worst case of GIGO ever
AI hasn't hit a scaling wall, LLMs have, there's a big difference. We can reason about what things LLMs might be capable of and what their limitations are. But any claims about limitations of the field of AI as a whole are not, can not, be grounded in science.
LLMs haven't hit a wall either. The only thing that is slowing is increasing run time on the same data set. You can increase the data or changing the training methodology or change inference or change structural things like memory size or, etc.
Exactly I hate how ppl don’t understand this. Ai isn’t going anywhere and most ai companies are working on many more methods other than llms to get to agi, which we know is possible. Llms haven’t even hit a wall either, although I think we’ll need more than llms to get to agi
@@Idiomatick OP is referring to the cross entropy loss wall. No model has crossed it.
@@theBear89451 If you're referring to the Neural Scaling paper, we don't need to anyways. I mean, it'd be nice if we did, but it isn't a requirement. I'd also argue that OAI had a limited sample size of models to make that paper from.
The electrical, cooling and economic requirements are going straight up on the chart and almost certain to plateau fairly soon. And don't for a minute think that chip making is going solve much of those problems, AMSL, TSMC, Samsung and Intel are hitting a number of limitations in making their nanotechnology smaller, less power hungry and production ready. We may be close to peak chip without entirely new technology. And if that's the case, anything production ready is a trillion dollars and at least a decade away, probably more.
I am a software engineer and computer science graduate whose primary focus isn't machine learning but who has some experience with machine learning and I have been arguing with people who "know better" not for weeks or months but years now that the "AI" we have is not getting any further.
The real wall or point of diminishing returns had already been hit pretty much with GPT-4. So the hype among the general public coincided with the peak in actual performance having already been reached. But try to tell that to people who are convinced that their stocks/crypto/nft is going to the moon.
That's simply not true, there have been significant performance advances since the release of GPT4... Especially in math and coding.
@@throw_away_account4514 There has been marginal improvements since the release of GPT4 that look really impressive because you're only ever shown the comparisons to the previous version which is to squish the graph.
@@throw_away_account4514 Who cares. these ppl have already made up their mind im afraid
The field of AI is far more than just LLMs and techbros. If it deals with noise, uncertainty, classification, exponential-level search, probabilistic models, vision, robotics, etc... it's from the field of AI. Do you really think we've hit peak robotics? That our space probes and landers have peaked in their autonomous abilities?
Sabine even has an earlier video about engineers using ML AI to twist and constrict plasma in a fusion reactor. The control system must keep the magnetic fields in perfect sync as the plasma squirms. Programming this with ad hoc rules and optimizations was not successful, so the use of a ML framework to learn plasma control from good and bad examples.
That part with the guy saying "It's not statistics because you have to understand to do statistics." hurt so much to hear. He has no idea how often statistics has been applied incorrectly in the world BECAUSE the person doing the statistics doesn't understand what they're analyzing. The amount of harm that's come from that assumption even before AI came along is so high.
It's motivated reasoning all the way down. The investors want to believe, the founders want to believe, the engineers want to believe. So they all hold hands and sing kumbaya and selectively ignore empirical evidence that contradicts their faith, while AI evangelists abandon any effort at objectivity and just speculate wildly with the speculations always supporting the grand vision. It's just like church, but with technology, money, and hype the holy trinity instead of father, son, and holy ghost.
Will you be surprised if AI is way better this time next year, despite the end of this scaling regime? If you do get surprised in that way, will you change your mind about why the majority of experts expect AGI in the next few years or decades?
@@jcl7372 Computers heal amputees.
Your religion got beat.
Successful AI doesn’t mean it has to understand the real world. It doesn’t mean “being close to human”. I mean each human has had over 18 years of continuous learning about the real world, which an AI won’t have.
Successful AI just has to be good at what it does, and get better and better at what it does.
It can not aprove a function indefinitely and realistically there is black hole of energy at each S graph.
@@tomizatko3138I would argue that smaller specialised AI systems are less susceptible to this and an energy black hole.
The problem is that most people think of LLMs like Chat GPT when talking about the future of AI development, systems which are trying to be literally everything for everyone (the entire world), these systems are bound to run into trouble when trying to scale exponentially.
Maybe there are serious blocks ahead, even with specialised narrower scope systems. But they’ll be overcome and a new S curve will start, leapfrogging the previous S curve.
That’s just my point of view anyway.
If someone says data is not important, then that someone is quite lack of data.
they don't have enough data to conclude that
Data is not equal to knolege.
@@MrAlanCristhian knowledge
@@MrAlanCristhianData (or knowledge) does not always equal wisdom (or common sense).
@@JohnBeeblebrox To make it clear, i know that what is "data - information - knowledge- wisdom" framework , as well as epistemologic frameworks about domain. In order to construct knowledge you need data. Those all are different things so we have different words for them depending on context :P
That's not the correct paraphase of Ilya Sutskever and Geoffrey Hinton's argument. They are making an interesting analogy to the Turing Test, by making appeal to "compression", or computational complexity of the capability of next-token prediction. I.e., "Supposing Aristotle kept predicting the "next time step" in any given physics problem correctly, then we are forced to conclude that Aristotle has inside his head a realistic theory of physics." That's Sutskever's argument in a nutshell; one can of course agree/disagree with that specifically. It's a thought experiment.
This is the most important aspect, and I would have expected Joscha Bach (or even Sean Carroll) to point it out beforehand, but thankfully Sabine is on the job.
Our world models are always a representation (written in mathematics and language): we can only get out what we put in, and what we can put in is arbitrary before falsification: we fit our models to patterns in the data, and then check if they make any additional predictions we can falsify the theory with. Thus falsification happens with new data, and new data comes from experimental physics: I.e. new detectors that probe higher energies and higher intensities, and produce mountains of data, of which we throw away ever larger percentages of to get to ever smaller probability events. Sooner rather than later we will need detectors the size of solar systems or galaxies (which we can’t build) to make progress, which means we reach a problem of reproducibility, because we will only ever get events that the universe is providing us with for free from violent cosmic events. - But the universe is less like a laboratory with controlled environments, but more like a zoo: physics will devolve into butterfly collecting.
Currently AI seems worse than humans in checking consistency, because it is trained on human language, which in itself is less rigorous than mathematics, so it doesn’t help us yet. And then only certain types of mathematics is even helpful at all (- Goedel showed the problem), leading us to a computational paradigm (- see all the talks by Joscha Bach).
We already do have hundreds of physicists (and crackpot amateurs) who still think outside the computationalist paradigm, suggesting their own TOE, who can’t find enough other physicists to even read their papers / get the community to hear them out: what use will be „creative“ AI, if it can’t think rigorously / find inconsistencies? - We are drowning in fake news, and what we need is first principles thinking, that is bottom up and uses the rigorous language of computation, rather than brute forcing top-down approaches in natural language to find all the dead ends and oneway streets that are the result from using faulty human language to begin with.
Maybe the AI can develop their own language to build a truer representation of the world at the physics level, but then we would need to learn that language to even understand it.
Maybe we shouldn’t be surprised that LLMs work better as emotional support animals, than they do as theoretical physicists.
Oh, the marvellous of YT: a video every day and lots of ADS. Beautiful to be alive.
Very good argument about decoupling levels and the need for experimentation.
While I appreciate Dr. Hossenfelder's critical thinking and willingness to challenge other scientists and physicists, it's important to remember that science is a dynamic process. Throughout history, many theories once considered absolute truth have been revised or overturned.
The beauty of science lies in its self-correcting nature. It's through rigorous experimentation, peer review, and open debate that we refine our understanding of the universe. By maintaining a sense of humility and acknowledging the limitations of our current knowledge, we can continue to make groundbreaking discoveries.
Let's embrace the uncertainty and the excitement of exploring the unknown.
Pile of platitudes.
I think this is one the greatest most insightful videos from Sabine. Also, maybe because she is pointing out the obvious...
Sabine you’re usually right about most things. I’m pretty sure you’re going to be wrong about this. Your talent is incredible.
i think the fallacy here is that the title implies AI scaling == LLM scaling. that's not the case. yes, LLM scaling is hitting a wall. that means AI companies need to pivot to other approaches, and i'm sure they will. they "just" need to find what to pivot to first. it's obvious that no machine learning model will ever "understand physics" by reading text. that doesn't mean that no machine learning model will ever "understand physics" (whatever that means tbh). that said, you should never take anything any AI company C-suite anyone says at face value, because there's always an agenda behind it. the only thing sam altman (and fwiw sutskever also mostly) is good at is "scaling", so he always talks about it. that doesn't mean it's the only important aspect. i definitely think the ARC challenge is a tangible thing to watch/follow, the people trying to solve it are getting quite far already with conventional techniques and building on those. let's see if they manage to break 85%+.
Interesting point on AI scaling limitations. This highlights an ongoing challenge: theoretical models are powerful, but without diverse, high-quality real-world data, the promise of breakthroughs often feels out of reach. Perhaps focusing more on optimization techniques like 'unhobbling' could offer incremental, yet meaningful progress.
It is absolutely wild to see physicists act like fearful luddites!