Its kind of tragic, though glad you find it funny. She's on youtube because she's a *failed research scientist*. She failed because her colleagues all over the world found her ideas to be at best, half-baked.
I was not looking at it from the standpoint you are considering. My take was comparing someone who got a Nobel prize with someone who has a RUclips channel. No reasonable person would ever expect to get a Nobel prize no matter what their research area was. I’m sure, or at least fairly sure, that this year’s physics winners never expected to get a Nobel prize in physics. So I don’t think Sabina was looking at things quite the way you depicted. Just my opinion.
As a technician I have often found that I and other technicians occasionally run into technical system failures that we say "has teeth". That means after going through a troubleshooting process, many times, we can't seem to find the problem (s) to fix it. That is when we either set the device aside, and go work on another device and repair it, or have another technician come over and start the troubleshooting process. Either way, the problem is found pretty rapidly. Why? We in the initial troubleshooting process made an assumption that some component was good. Didn't verify it. A.I. systems, by narrowing down the possibilities might skip over something that fell outside of the narrowed scope of solutions available. That might be where a Quantum computer can be the refreshed eyes looking for the solution to a problem that eludes A.I.. That hybrid method of search might be best as it only gives the Quantum computer a limited number of problems, and lets A.I. find the solution quickly for problems that will fall within it's scope of applied rules.
Reminds me of a guy I knew in the air force. He was extremely intelligent... our career path was already full of intelligence, but this guy was miles beyond anyone. He checked out a piece of electronics that wasn't working. He found nothing causing it. So he checked each individual board, and found nothing wrong. He desoldered each individual component, checked them all out one by one, and still found nothing wrong. This occupied his time for over a week. While he was away, an airman came over and replaced a fuse in the lab bench. It worked perfectly. There was never anything wrong with the unit. The work bench power supply had a blown fuse. so the unit never had power. It's just one step away from not plugging it in. It's a common intelligence trap... the assumption that a complex problem must have a complex cause, or require a complex solution.
I conveyed this in a comment I made responding to the video, but I figure I mind as well just put it here instead because we're on the same track of thinking and I agree with your comment: I don't know. I feel like AI computing could replace some aspects of quantum computing, but I feel like it'd just offer another level on top of the bias that human researchers and scientific experiments already face. Neural networks are modeled after our brains, and work by placing weights to different nodes and on different edges, if an AI sees a solution resulting from the high weight of a specific node then what happens when it doesn't consider another node which is just as relevant. Moreover, what if it continues down the path of a specific edge trying to justify it's theory rather than consider it might be wrong? We already see this in ChatGPT and other large-scale neural networks where they can become so convinced in their interpretation of the input stream that they'll actually ignore things they should be doing or considering because they refuse to readjust their "thinking."
quantum problems (NPComplete) don't get solved by not skipping what humans would skip, but by trying every possible state at once which humans (or classical computers cannot do). And the problem you mentioned doesn't need quantum or AI, it just needs to work without assumptions in a completely classical way - simply check everything that is checkable, the possibilities don't go exponential... if they did it would not be solved in the way you say it does.
Veritasium recently explored similar ideas behind what contributes a human to be an expert in chess and other professions, the chess master is not memorizing each chess pieces position but rather chunking the configuration/rule that constraints them to their positions
The chess analogy is very useful. The chess grand master has a better set of neural network algorithms for chess than the novice. I stopped playing chess when I realised there was no luck involved and that likely there were plenty of people with a better network than I had! I prefer a little bit of chance to mess things up.
@@alwayscurious413 The question of whether human memory is related to neural networks touches on the very core of how we understand ourselves and how we build machines. Human memory seems to work in ways that are surprisingly similar to artificial neural networks, but with some critical differences. In humans, memory is distributed across the brain in networks of neurons. These neurons form connections that strengthen or weaken over time based on experiences, much like how weights in a neural network are adjusted during training. This process, called synaptic plasticity, allows humans to store complex patterns of information, associate ideas, and even "reconstruct" memories based on partial inputs, just like how AI models complete tasks. But here’s the twist. Unlike computers or artificial neural networks, human memory isn’t exact. It’s deeply tied to emotion, context, and even sensory inputs. For instance, a smell or a song can suddenly trigger a vivid memory, something machines can’t replicate yet. On the other hand, computers and traditional neural networks focus on efficiency and precision, storing data in a way that is either "on" or "off" and lacks the nuanced imperfections of human recall. To address the comparison to traditional computer memory... No, standard computer memory doesn’t operate like a neural network. Traditional storage is static and precise, like saving a document. Neural networks in AI, however, are more like human brains they’re dynamic and learn through experience, but they still lack the depth of our biological processes. For example, neural networks don’t “forget” the way humans do, and they can’t yet replicate the emotional weight we attach to memories.
Just take the conventional route until the first right, then follow the quantum path? AI and quantum computing are not in competition but are synergistic technologies. When combined, they can solve complex problems more efficiently and effectively than either technology could alone, enhancing each other's capabilities
I don’t know how to feel about the statement that they are synergistic. But one thing is for certain, the theoretical groundwork for the current Gen of AI/Deep Neural Networks has been laid the past ~50 years or so. Quantum Mechanics itself has not even existed that much longer. What I am trying to say is that AI just has a headstart right now. But the field seems to be narrowing as many low hanging fruits have been picked already. For QC, we don’t even know for certain how to use it yet and whether we can make good use of it at all. Hypothetically, the next breakthrough could come any day.
As a total outsider, I've also thought the same way recently. Now I wonder. I think @@SanguinaryOtter may be right about what he said. But the question arises that how much money and time should be put into it. There are too many possibilities. I think It can even happen that QC will be practically dropped in a decade or two, then later gaining new life; kind of like how electric vehicles did or how natural and smart architecture is now gaining new focus, after the realization that modern glass and steel electricity dependent structures don't have much for them except a phallic coolness factor. To provide two quite different time frames as examples. Edit: Question for those who know it. How different is too write algorithms for QC (if any) compared to today's standard? I imagine some difference is required when writing for them but I would like to have an educated answer.
Both self-limiting data and negatives prune analysis. Reduction of questions/operations are the key. Even the “tree” example is easier if reduced. Trees grow upwards and outwards. Leaves don’t grow from roots. There is a limit to the growth rate. Gravity/stability of the system are also self limiting. Environmental considerations can be applied: Altitude, wind, fire, lightning, walls/cliffs/ceilings, human pruning, animal damage, soil type, substrate issues & obstacles, considerations of time (always forward), etc. Now if there was a true fractal, that would be quantum computing.
Together with deep-learning, hyper-parallel algorithms are also emerging as a serious rival to quantum algorithms. Recently, problems seen as strong candidates for quantum advantage, like optimization and combinatorial tasks, have been solved faster using classical algorithms that leverage massive thread parallelism. That is a major problem for quantum firms since GPUs are becoming exponentially more numerous and powerful while quantum computers are struggling to do anything.
Not really, qc specific algo has some serious advantage compared to classical machines. More gpu doesn't mean classical system will beat quantum systems. Say specific problem like search using grover is always going to be faster on quantum systems with quadratic speedup as compared to classical.
yeah but this isn't due to the model of computation , but rather because gpu architectures have decades of work refining them. this is the equivalent of comparing two versions of an algorithm (sorting for ex) , one logarithmic and the other linear , and have the linear one benefit from hardware optimizations , like better cache access patterns , SIMD , SIMT etc. the linear algorithm is going to be faster , up to a certain point. as fancy as GPUs are , they are still classical computing machines, even if there are millions of threads.
Ultimately the issue is that eventually, quantum systems should have an edge, especially in terms of power efficiency. A lot of our advancements in conventional compute are coming through largely brute force methods, and we're rapidly approaching the miniaturization limit for conventional compute, while we're far from such for quantum compute. I believe that conventional compute will maintain an advantage for the next decade or two, however when we hit that miniaturization limit, any significant advancements in compute density will come from future advancements in quantum compute, as architecture improvements have historically been far less impactful than miniaturization in terms of efficiency gains by processor generations.
When I was a young teenager, my younger cousin and I had the nerdiest debate with each other. It came after I wondered aloud when in our lifetime we would see computers that could think. He was completely against the idea saying computers can only understand yes or no, ones and zeros, and fundamentally could not understand “maybe”. It occurs to me that AI is exactly that- a way to model on a deterministic system a way for it to learn what is probable. AI does understand maybe and maybe there is some link between how the rules that govern classical computing work and something fundamental we are missing in the rules of quantum mechanics. As for my cousin, he went on to become a physicist, so that’s his domain, but that argument we had long ago is what inspired me to become a computer scientist.
Great way to look at it! Classical computers may not be quantum in the strictest sense, but if they model it closely enough does it really matter? It shouldn't really matter how you get there as long as the output is useful and the speed is acceptable.
The big issue with science and philosophy today is that we don't take seriously the ways in which systems scale and produce new systems out of underlying regularities and irregularities. Humans are the product of a number of phase transitions from quantum > physical > chemical > biological > intelligent > sapient. Each higher transition constrains its environment to ensure its propagation. One could think of the Big Bang as the physical beginning its process of self-perpetuation out of the quantum, just as the biological began self-perpetuating out of the chemical, and sapience (aka memetics), began self-perpetuation out of intelligence. Each of those levels follows its own independent rule set, and yet also constrains & structures the underlying system. If I had to pick a field of study that would benefit humankind more than anything, it would be studying the structure of those phase transitions and discovering their patterns and limitations.
Quantum computers maybe able to quickly solve combinatorial optimization and problems with exponential complexity, but ironically building a quantum computer with enough error-free qubits itself becomes exponentially harder as scale increases.
Hii, just a minor note for Alphafold at 3:50, AF actually does not use physical rules, it generates model based on the similarity to the training dataset of solved structures... so structures closely related to something that has been solved before are more accurate that those distantly related.
@@lost4468yt yeah but those rules it finds are not physical rules, those are distances between evolutionary conserved residues in the protein structure, this means that is is very bad in prediction of unfamiliar proteins that are lacking evolution conservation in the training data .. .ergo, stuff that is interesting.. not even touching protein complexes or nucleoproteins
Exactly! The physics are NOT there. You can not fold a protein de novo using alpha fold 2 or 3, etc., due to a rule (e.g., hydrophobic effect) that was uncovered using QC. Evolutionary coupling and knowledge bases (alpha helices, beta pleated sheets, etc.,) from the protein data bank can give surprisingly good results when producing protein models. We’re a long way off from computing delta S of the free energy equation.
I still think the term "obsolete" is a bit of a stretch considering fields such as cybersecurity, specifically post-quantum encryption exist and offer challenges unique to quantum computing. Quantum computers will still have their uses in fields such as these where deep learning at the moment seems too different, and it's only a matter of time before new problems emerge that deep learning can't quite learn. Not to mention the large concerns about power draw that arise concerning deep learning and overall scalability. Supercomputers at the current moment are more appealing but I hope to one day see that change as we approach a plateau.
Sabine nailed it again. (And I bookmarked this contribution) Quantum chemistry has my interest as I did a minor once (a long while ago) in quantum chemistry. I was always fascinated by the idea that you could "calculate" some atoms and to a certain extent molecules. Just the idea of "in silico" chemistry is fascinating; it brings about new molecules and understanding of the workings of proteïnes and even vaccines (Covid e.g.). On the other hand I was able to look from the outside in to the emergence of quantum computing. It soon appeared from theoretical considerations (computer science and mathematics research) that quantum computing would be efficiënt only to a certain class of problems to solve, not so as a generally better approach to all classes of problems. As for the AI angle in this all I don't know, because "AI" is a container concept. Does this include molecular dynamics calculations (Karplus 2002) for instance? AI is a grab bag of (sometimes misunderstood) approaches and methods, and often "rediscovered" by other scientific disciplines.
Addendum. Researchers from Sandia Labs (Kallaugher and Parekh) found: " ... that quantum computers are unrivaled at solving an advanced math problem. Unusually, they proved quantum computers are not faster than regular computers; instead, they use far less memory."
Was this video implying that quantum computers are better at solving quantum mechanical problems than conventional computers? I don't see why that would be true. I was also under the impression that one of the advantages of QC was an increase in security of data transmission.
In that sense they're opposites. AI will confidently give you the wrong answer (rephrasing it every time, but never admitting that it just doesn't know), whereas quantum computing will tell you "probably, like, I'm 99.9999999% sure, but I could be wrong" even when you ask it to check that 1+1=2.
@@RFC3514 That is purely just a user-interface thing. Under the hood, AI gives answers in terms of a probability distribution of the likelihood of different possible outcomes. The user-interface is then designed so that the AI will just tell you whatever possibility has the highest likelihood _is_ the answer. You could just change the user-interface and have the AI spit out things in terms of probability if you want. Your claim about how quantum computers would answer 1+1=2 is also factually wrong as you can trivially write a quantum algorithm that outputs the answer to addition problems deterministically without any uncertainty.
@@amihartz - No, it's not just a UI problem (although some AI companies looking for funding would want you to believe that). Traceability is a *fundamental* problem with many models. Many of them literally can't tell you how or why they arrived at their "conclusion", and are truly convinced (like, close to 100% certainty) of complete falsehoods, while being _less_ convinced of some true facts. Simply showing the user a "confidence" value won't fix the problem.
@@RFC3514 How is that relevant to literally anything discussed here at all? The fact an AI cannot tell you how it arrived at its conclusion has zero relevance at all to anything here. A quantum computer also cannot tell you how it arrived at its conclusion. This is just something entirely outside of the realm of this discussion you are trying to arbitrarily inject into it with no explanation.
@@amihartz - I simply addressed your (incorrect) claim that "it's just a UI problem". Also, there's nothing "being discussed here". The OP made a joke, I played along with it, you posted something that simply isn't true, I corrected you, and you got mad.
I would say, the most useful applications of AI are going to be the ones that we needed the quantum advantage to solve, but not just because Quantum Computers are late, but because AI can not only "solve" but also "generate". For example, AlphaFold can predict protein folding, that's amazing. But just recently it allowed to make AlphaProteo, which can suggest brand new protein that bind to a specific site of an existing protein. The possibility to transform a "solver" into a "generator" is incredibly valuable. I don't know if an equivalent strategy exists for Quantum Computing?
This is a question I asked but have gotten no response to. Isn't the title of this video a false comparison? Conventional and quantum computers are tools. AI is an objective. Isn't the video conflating AI and conventional computing? It's like saying I have a gasoline vehicle and and an electric vehicle. Will driving to work make my EV obsolete? It doesn't compute.
@@chilinitro5599 : Whether they are the best tool at the moment isn't my point. Comparing AI to quantum computers is a false comparison. Computers are tools. AI is a desired result of using those tools. Like comparing a hammer to hanging a picture.
What I find fascinating is how Reinforcement Learning (RL) is demoted as "too inefficient", yet it is used in Nobel Prizes winner Hassabis's AlphaFold and AlphaGo work, and Yann LeCunn often says to "get rid of it", but only use it if you are "fighting a ninja" or if your "plan does not work", it is just a "🍒on top of a cake".
Not a cherry, but a Strawberry (aka Q*). And nobody knowledgeable has questioned the utility of reinforcement post-training of transformers since OpenAI's release of o1-preview (on September 12, 2024, comparable to no day in my life but July 20, 1969). As for computational expense, recent experimental versions of Gemini have risen to the tops of leaderboards with just 32K-token context windows. This suggests that relatively small transformer networks will be used in test-time deliberation ("reasoning") to generate high-quality responses.
@@tomenglish9340 I agree, the method of "train, then infer/use; retrain, then infer/use" only goes so far and seems to work best for "autonomous nervous system-like" performances, even ones that require high levels of skill. But learning by doing - through experience and practice - and being able to recall and relate that when confronted by new contextual or situational "features" presented by edge cases seems essential (and should be faster and more efficient as well). I would call viewers' attention to a recent Know Time #97 RUclips video featuring a conversation about this very topic with Joscha Bach entitled "Artificial Intelligence, Consciousness & Computation."
I think of this as "not all functions are continuous". There are constraints everywhere and if your model is good at figuring them out, convoluted as they might be, you cut the search space significantly. ML has been used against cryptography before, and that's one of the places where you simply need to reduce the search space to something more tractable for brute forcing.
Yeah as a computer scientist, it doesn't even come down to concepts or ideas here. There is a physical difference in how classical computers compute, and they have physical limits. An actual functioning quantum computer will always be far more powerful than a classical computer with the same number of bits. The problem is the functioning part.
As a computer scientist, you should know then that the potential of classical computers hasn't remotely been tapped. The physical limits (two state transistors) are a transitory problem, are they not? Quantum computers will always have the problem of having to be suspended in a very specific state to even function properly. It is a practical limit which may never be overcome.
Yes, but this extra power might not be practical. For example, we could have most problems of industrial interest to admit decent approximate solutions on classical computers (this is where AI excels), then there's little need in quantum computers. Or, maybe scaling up QC will end up unfeasible: while theoretical considerations clearly show asymptotic quantum supremacy, the actual point where QC starts dominating everything classical (e.g. AI) on a financial scale might happen too late.
@Meditations2024 so classical limits are solvable, but quantum ones aren't, and your basis for that is? But no that gets into computer engineering and physics and I'm not a computer engineer nor a physicist.
Not needing all the possibilities reminds of compressive sensing when representing images (photographs). The permutations of pixels in an image space is unfathomably vast, but only a small subset of that space is actually useful which contains "natural" images, the rest is noise. Knowing this, it's possible to reconstruct images from very sparsely sampled random data, using clever minimisation algorithms.
I think a lot of people in the comments here fail to understand what the real advantage and applications of quantum computing are. Yes, some quantum algorithms only solve problems with *how* we are using classical computing, but the actually interesting ones are the ones which solve fundamental problems with classical computing, and they simply *can't* be modeled by AI- any accurate model is at most as efficient as the equivalent classical computing algorithm or less efficient.
so, if its impossible then its impossible? nothing is impossible, if you look back at 2000, and bring chatgpt to the world, the world would be shocked, because at the time, everyone thought LLM is as useless as a roach, but now it is not.
@@sans00xa good example is finding prime factors of a given number. The best classical algorithms are incredibly slow, which is something that makes it very useful for encryption. There are known quantum algorithms that can find prime factors by orders of magnitude faster. Get a working quantum computer and suddenly large amounts of modern computer security is trivially easy to bypass.
I believe the real question is how many *useful* algorithms are accelerated by quantum computing. As of now, I believe the only general purpose algo is really Shors algo. We already know how to do “post-quantum crypto”, so I think this is of historical interest only at this point. So, one thing we *know* QC is needed for is….simulating quantum systems. But not all quantum systems, in fact just the ones where entanglement is significant, ie not decohered. Which, by definition, is surprisingly few / small systems, because if large entangled systems were commonplace then QC would be much easier to make! Another way to look at it, for the specific case you gave in quantum chemistry: some aspects of molecular dynamics it’s really important to model the coherent delocalised effects - binding sites, transition states etc. But definitely *not* a 100kD protein. Because if it were, then making such a protein would be an excellent candidate as a functional unit of a QC. Instead, a protein of that size is well-modeled as a classical (stochastic) connected set of (coherent quantum) sub-regions. You’ve referred to that as “structural rules which AI can infer”, but the important bit isn’t the AI-ness, it’s the classical-ness. Therefore, the size of a relevant coherent system is somewhat bounded. Which in turn limits the scaling advantage that QC can possibly bring. I’m not saying there is *no* room for QC. Given that it’s currently really difficult to model quantum systems greater than about 150 atoms; but the fact that the largest useful natural coherent systems are probably no greater than few thousands atoms, kind of indicates a “slot” of usability in logarithmic complexity-space which is narrower than people have assumed.
There may be a silver lining in all this even if Quantum computing does not pan out. It helped scientists define new problems were thought to be only solvable on a Quantum computer. As you pointed out this could lead to innovative ways to solve these same problems on a classical computer. Even if Quantum computing does not become the breakthrough that was hoped for, it still has propelled computer science forward.
It reminds me of a seed company, I think Monsanto, created a number of crops using traditional breeding techniques, but based on their knowledge of GMOs, to create plants with very good properties but not directly genetically manipulation, so a sort of non-GMO plant with similar advantages as a GMO one
No one ever thought there were problems "only solvable" with a quantum computer. It's just some operations scale better. Just as different classical computers can scale in different ways...
I *think* I understand what you're saying @sabine but let's see... So if you simplify it down a bit to a simple 3 dimensional array that's 10 units on each dimension so there are 10x10x10 (10^3 "possibilities", way less that the 10^40 in your protein example but easier to visualise in my head). So we have arrX[0,0,0] to arrX[9,9,9] which we can visualise as a cube with 10 squares wide by 10 squares high by 10 squares deep, or perhaps better as a tree that has 10 branches at the base, and then 10 branches from every branch, and then another 10 branches below that... and the quantum people would go "ohhhh, all those branches and possibilities, you need a quantum computer to handle that"* but the "AI way" would be to learn from all the 10x10x10 arrays it's seen in nature and say, "yeah, well you can ignore everything on the [5,,] branch, that can't happen for whatever reason, so that reduces the space down by 100, and the [9,9,] branch is a dud, so that shaves off another 10, and you really don't want the [6,6,6] branch, so prune that one off too and so on, so that it turns out the answer is somewhere in the [2-4, 4-8, 0-9] space, so you're down to 3x5x10 = 150 possibilities, and at classical computer exaflop speeds vs quantum computer megaflop speeds this would be way faster to just brute force classically. And this works in the "real world" because real world physics (and chemistry etc) has rules... molecular bond like to be at certain angles, proteins can only fold in so many ways but not every way, and so on. So we can "prune off" the "cannot occur" branches from our tree, and AI is really good at finding these hidden rules, hidden symmetries etc.
One counterexample, relating the geometric search space that robots move and Alpha Go. The reason that Alpha Go was able to significantly reduce the space complexity is because of Monte Carlo Tree Search. That is, the game go can be modeled as a binary tree. Each move is independent of the other and the set of all moves is equivalent to a binary tree. One starts with the first move and following a tree structure, one ends with the last move of the game. Then, the complexity space is reduced using backpropagation from the last move to the first. This allows a set of moves to be compared with each other using real numbers in orders to find the best set of moves. This, however, can't be used when a robot moves. The best path that a robot moves is inherently a Euclidean distance and there is no geometric method to compare one set of robot geometric moves to another. The result is that it's computationally expensive and difficult for a robot to find how to move in an unknown and/or changing environment. This, also, is the reason that there are algorithms that can efficiently solve games and used in language and why there are no robots that one can buy to do your dishes or clean your bathroom. So, my point, I'd like to see the exact reasoning of why AI can replace quantum computing. One method that's used by classical computer doesn't necessarily mean that it can be used for all methods.
AI produces simulacra of things that look like the real thing. However if you substitute this for the real thing you eliminate the possibility of discovering new aspects of the real thing you have not yet discovered and are not present in the simulacrum. You may even discover new aspects that are only present in the simulacrum but not in the real thing.
Wasn't it John von Neumann who came up with the idea that any computer could be simulated in software? You wouldn't need a 'genuine' quantum computer if you could simulate one in code. Sure, it wouldn't run very fast, but it would be a faithful replica in terms of compute capability.
I think that's what the guy in the video was hinting at. For instance, a team at the university of Padova is trying this approach (which they call "Quantum inspired factorization") to try and break RSA encryption. Of course it scales worse than brute force right now, but it opens possibilities nonetheless.
So instead of using a classical algorithm that takes a million years to solve a hard problem, we take a quantum algorithm that can do it in a second, but then we run that in a simulator that will take a million years to simulate that second of quantum computing 👌
As the previous commenter says, this doesn't work. The advantage of quantum computers comes from the fact that it can do certain (very specific) operations in constant time that take exponential time to simulate on a classical computer. The point of Demis is more that you can optimize specific problems enough via machine learning to get fast results on classic hardware. This doesn't generalize though, you can't (barring some algorithmic breakthrough) simulate a general quantum computer faster than the theoretical limit.
@@Clayne151 it takes a long time now, but research may find a way to speed up the solution of the equations they use to model the problem. It could be a better algorithm (like the FFT sped up DFT) or a different (non quantum) type of analog computer.
Does this mean that if I ever want to find the shortest route to visit every town in Sweden to sell my door knobs, I will still need a quantum computer or can classical computers solve this conundrum for me now? It's kind of important to me.
ever heard of the Travelling Salesman Problem? There exist algorithms that do that with classical computers. I recently wrote code for a robot using this algorithm and it worked really well.
@@245trichlorophenate It's actually solved pretty efficiently on classical computers. Companies like Amazon can use GPS to get driving direction that route between all their stops before returning to the warehouse rather efficiently and quickly. Technically yes, a quantum computer would be in principle more efficient at solving it if they ever caught up to classical computers in terms of memory (a big "if"), but this difference would not be that significant unless you're like a giant enterprise like Amazon where slight improvements to routing can save you a lot of money. For the average person, it won't matter much.
Advancements on von Neumann architecture systems (like GPUs and TPUs and horizontal scaling) have made tasks like drug discovery faster and more efficient. Tools like AlphaFold and advanced AI models now rival quantum computing in these areas. If they become viable, quantum computers might enhance certain niche tasks, but they're no longer essential for breakthroughs in chemistry.
Classical computing has a two factor advantage over quantum computing First it has a head start of many decades. Second the prospect of new technology superseding it gives CP scientists a significant incentive to develop and innovative to leverage their technology. Plus while low temperature environments are beneficial to CP they work just fine at ambient so there is that.
Ah yeah, the air taxi, and the research she didn't even look into. Amazon would use them directly, followed by Uber, and there's many more, when infrastructure is solved. Germany will wait another 70 years. When way over half of Europe has it, Germany might think about legalizing the regulations.
The final decision will be financial...which system can deliver business outputs at a cheaper price. And for now classical computers have a global network of production and distribution.
Protein folding is a classical optimization problem that suffers from the “curse of dimensionality” and from having many local extrema. Folding itself is a dynamic process but we are interested in the final static structure of the protein in the equilibrium state. It can be solved by the “virtual work” principle that has been around since the 17th century if we have the computing resources. Using NN based AI to solve optimization problems smacks of a certain inelegance although it can narrow the space in which we would search for the exact locations for the equilibrium configurations. It is similar to using some heuristics in which NN acts like an “expert system” to invoke another term from the history of AI. Expert systems are basically a set of “rule of thumb”s; it is like sticking to a traditional family recipe when cooking and it contains a good amount of stubbornness.
My bet is on AI. Quantum computing has not produced anything meaningful despite the billions invested, whereas AI, on the other hand, continues to amaze us.
@@didiervandendaele4036 Where are you going after you die? What happens next? Have you ever thought about that? Repent today and give your life to Jesus Christ to obtain eternal salvation. Tomorrow may be too late my brethen😢. Hebrews 9:27 says "And as it is appointed unto man once to die, but after that the judgement
The Science article tackles the problem of ground state preparation, which is a static state always difficult for a quantum computer, and possibly a suboptimal problem for a quantum computer to solve, while bypassing addressing the difficulty of dynamical evolution of states which variational algorithms are incapable, but has a major advantage in quantum computing.
Problem solving with AI is like trying to find a better search algorithm. While solving a problem with quantum computing is more like brute forcing the solution. Both paths have their merits and drawbacks. And in the end, the two might need each other.
It all comes down to the "P vs NP" classification problem in mathematics. Many NP Hard and NP Complete problems can be solved. We can never know if we have the most optimal solution, as that would require checking every possible combination, but we can demonstrate that we have "a solution". I worked on this as an undergrad and managed to solve problems that were considered insolvable just 10 years prior.
I am for breakthroughs in any technology. There are many paths we have gone down that have not proved fruitful. To me, Quantum computing has been stalled for quite some time. I am starting to wonder that If it wasn't for government FOMO funding, would anyone still be pursuing Quantum computing so aggressively?
I would argue that path of least resistance is not always correct way to do things. Easy example is gasoline vs electric car. Just because we dont have a good ways to use quantum computing right now and its expensive doesnt mean it wont be the missing piece down the line.
Artificial neural networks are arbitrarily large or small systems of linear equations. This can be used to approximate functions. You just have to find a suitable approximation.
Now LLM's are being distilled to run on 1Gb of ram, and shrinking. I never thought such large training models could produce tiny, meaningful algorithms. Natural computation has potential, but neural networks are evolving right now!
That's just a normal feature of neural networks compared to deterministic algorithms: You can cut down existing layers to reduce size and increase speed at the cost of accuracy. The thing about LLMs though is that accuracy isn't very important so it's not that big a deal. Performing computations or autopilots like that where precision and accuracy actually matter would be a much bigger deal. We know that can be done but we haven't figured out how to reproduce it (housefly brain vs two flight computers on a Harrier).
→ DKN You don’t change the “weights” but reduce the number of bits allocated to each weight. So if the original weights were 32 bits long then reduce it to 8 bits and do the computation that is optimized for the lower precision. Maybe you can “drop” layers but that is something that I have not read about, or has not been explained how it would work. If one reduces the precision of the weights, it takes up less memory. I would be interested to know if layers can be dropped and a reference to that work would be appreciated. Always interested in learning new.
AI might actually help to make quantum computing controllable by reducing the search space. And help stabilise the system. Because even if the AI follows rules and therefore reduces complexity there might still be rules we don’t know about and computation is needed to find them. This is where quantum computing might then have an edge.
I don’t agree with his take. A true quantum computer of scale, would simply have capabilities beyond any classical computer or LLM or general ai. This just seems like a re forming of the same old question that been around since the start. Can we scale the quantum computer and handle noise/ error checking efficiently. Nothing really new there, not exactly a hot take.
Quantum computer and AI Large Model are two completely different technologies. It is just that they try to solve the same kind of problem. AI through pattern matching and latent space can compress high dimensional data leading to faster solving, with some caveats such as it's a probabilistic approach, while quantum computing explore the whole problem space at once. There is tons of applications of quantum computing which cannot be replaced by ML, such as prime factorization problems. Because there is no patterns to discover and no way to compress the problem. So yeah, I don't think AI is competing with quantum computer, at least not on 100% of the problems we are trying to solve.
If quantum computers can do the unthinkable, how do you program for such 'miracles'? And how could anyone verify the results, except by using another Q-machine? It's like praying to God, getting an answer, then having to ask another God if the answer is correct.
@@pirobot668beta By clever use of quantum gate, which "pump" the probability for a particle to be in an specific output state. You "force" entangled particles to behave in such a way that when you do the measurement and interact, they "probably" give a coherent result. You then measure multiple thousand time, and often problems are not NP to check if the result is valid, so you can assert that the output is valid with a normal computer. Easier said than done, plus it's very complex and I myself do not master the subject. You can read the Shor's algorithm for example to see step by step how it works.
I didn't think that there was any question that Turing or von Neuman machines could simulate quantum processes. It was simply that quantum computing could do that and other calculations (particularly P and NP) much faster
@@randomYTwatcher Nope. Even the part we do know about how human brains work is VASTLY more complex than how puny little ChatGPT does things. Be a little more proud of the cells between your two ears.
I’m pretty sure I’ve been saying the same thing as this guy but using words like relative and sharing. The tree is the foundation or a sharing dimension with the relative, the differences expand the relative like space and time. Space is the stuff and time is the difference, if the differences share then new dimensions can be revealed. Unconditional freedom is part of this but it’s still relative emergence but still connected to infinity. Honestly, the dimensions/space/stuff isn’t limited because it’s all relative, there are no separations just interference patterns. We listen and share and things that matter matter, know what i mean? The parts we share with are what reflects or connects, it’s why we feel empathy, the difference is relative even if unfound by the self.
I am a quantum chemist and this is gibberish. The class of computer problema that can be solved efficiently or QBP is much bigger than the class of problems that can be efficiently solved by any classical Turing machine, or P. Obviously all interest in quantum computing stems from their capacity to solve QBP problems that aren't in P, such as elliptic curve factorization and certain hard optimization problems. In the specific case of quantum chemistry, most interesting problems we cannot tackle well, like light-matter interactio, also happen to be computationally hard in the QBP sense. For the class of problems in P we already have plenty computationally efficient algorithms developed over decades of research, that I doubt AI will improve much upon.
Had an inkling this was nonsense (having been following both techs for my entire 3 decade IT career) and always felt both are overhyped! Fantastic explanation; thank you!
It's not gibberish. the graph @2:58 is the key to understand as she mentioned. Where the intersection is (or will be) - we just don't know yet. If we never reach the intersection then classical/AI computations will win out, otherwise QC.
That's not what the argument was about. They are saying, since AI makes damn good approximations and runs in O(n) where n is the number of weights in the NN, even when dealing with NP-hard problems, is quantum computing worth all the money? It's obviously going to have its situations where it is faster and produces a perfect output, but in other situations where it could have an advantage on normal computing you could just do a good ML algo with 80-90% accuracy and run it on one of the conventional computers (which run much much much faster than quantum computers) and call it a day, you would save a ton of money
That might be true, but the point is there are problems that are (for all we know) not in P, which in practice can still be solved relatively efficiently on a classical computer (a typical example for this is the SAT-problem). Problems like protein folding and go are most likely also not in P, but AlphaFold and AlphaGo have performed remarkably well on these problems. AI systems are different to most classical algorithms in the sense that they are probabilistic and in general not guaranteed to find the correct answer, but it is still possible that they can find the correct answer efficiently for most cases (even when the problems are not in P). By definition problems in P require an algorithm which always returns a solution in polynomial time, which is not what AI systems do, so there is no reason why AI could not perform well on problems that are not in P. While I doubt that AI systems will ever be able to fully simulate a quantum computer on a conventional computer, there are most likely problems that classical computers were not able to solve efficiently, but that become solvable efficiently enough in practice because of AI. If there are enough problems like this, then quantum computing becomes less and less attractive/useful, even though theoretically it is probably more capable than a classical computer.
I'm not a quantum computer nerd, but this "classical" workaround with AI only works on things we already have examples of, reducing them to higher order "first principles" (how chemical bonds works in folding for instance), which leaves a still juicy area of exploration: A quantum computer would be needed to explore from firster principles (quantum mechanics) the paths nature didn't take. Since nature is based on randomness, and building on what works with evolution, our current world may have circumvented amazing things in the space of possibilities.
Exactly this. Classical cannot model quantum because classical will never violate the CHSH inequality. AI may make computers faster, but it will never make them quantum!
@@t23c56 Well, you don't need to be quantum to generate random result. AI could certainly compete with quantum computers on that field! And does it matters if you call it hallucination or decoherence? What matters is that it gave this result way faster than classical computer! 🤣
@@igorstasenko9183Sorry can't really understand what you're trying to say here, but I'll try. Pure random number generation is not a use case of AI is that's what you're getting at. Perfect random number generation actually does require quantum mechanics, however in practice a complex enough classical system can easily produce random numbers. I remember seeing a wall of lava lamps somewhere with a camera pointed at it and they used that to generate random numbers. Totally agree that the result is usually all that matters, but there are still chemical and physical applications which do require the actual process itself to have been quantum. 🤣
One key aspect of generative adversarial learning, as well as any form of semi-supervised learning, so for all of the most successful approaches in machine learning tso far, is that the task fundamentally needs to be somewhat easy to verify as correct or incorrect. For GAN you need a very detailed dataset to begin with, and for any form of supervision you need a method to periodically check if the result is at all correct. Reinforcement learning is just straight up not applicable as you need an efficient computational way to calculate the reward function, or how correct the solution is. For chemistry and material sciences we know the properties of thousands of molecules and combound materials. But is that the case for quantum computations. I guess we could use LHC data for training, but thats probably quite noisy.
Quantum computing: Simulates all solutions at once to find the correct/best one. AI computing: Derivative guessing that "looks correct". What could possibly go wrong...
It would have to mean that like with protein folding it's verifiable after the answer is found. Once you have a better structure you can clearly see that it's better than your previous record. In areas where 'improvement' is more important that 'proven to be the most optimal possible' these areas don't need quantum computing that gives perfect simulation. It only needs answers that are significantly better than the answers we had before.
One case I can see is getting actual data to train AI. Since, I guess we do not have much data for let us say folding of amino acids, quantum computer can be used to stimulate probable data point (starting from the physical rules we know and generating possible data points), which can be used to train AI
A classical system to model quantum computing uses 100 MW per hour as electricity; new probabilistic approach into computing in quantum field using noise as a good thing could surpass everything
Two things. One is that coming up with better algorithms is important, and a well-known fact in computer science. One learns that in any graduate level course in analysis of algorithms. A "good" algorithm can be orders of magnitude better than a "poor" one. The other is what you said about AI. It learns from the data. In other words, it is probabilistic rather than deterministic. Add to that an understanding of the rules and you basically trim the tree of possibilities, as you point out. For example, I once wrote a program for my wife for a graduate course in mechanical engineering. Some of her classmates used a brute force approach on a large super-minicomputer (this was in the late 1980s). They actually ran out of virtual memory in a 32-bit computer. With a better algorithm taking into account the constraints of the problem, I got it to run on a 16-bit IBM PC with, wait for it, 2MB of memory. It ran in seconds. The algorithm is king!
To me, this is the stuff AI is supposed to be used for--identifying patterns to help scientists do better science--not create pictures or write things. Note I don't say "to do science;" there will always be the need for humans to review whatever AI notices and points out. It is a tool to help people find underlying patterns, not put creatives out of jobs.
Quantum computing is not the problem of quantum chemistry. QC is required only for very detailed analysis of spectroscopy and reactivity. Ahmed Zewail showed that more than 20 years ago for laser beam reactions. The same abstract that you show in the video mentions density functional theory, a semi-empirical approach that has been quite proficient at predicting all kinds of spectra and patterns of reactivity. As for the physics of protein folding, the work was done by David Baker, who himself is one of the winners of the award you mentioned.
Two very crude remarks (what else?) 1) The fact that Nobel prices in Physics and Chemistry are now attributed to computer scientists may mean that either Science is dead, or a computer science Nobel price should be created. 2) This question on IA seems to be a nice pretext to give up quantum computing without admitting that it will never work since, putting it simply, superimposed states are a theoretical artifact and do not really exist.
Hi Sabine, I am a big fan of yours and I value your opinions and insights very much. I recently came across an interview with a physicist/mathematician/philosopher from Harvard (Jacob Barandes) who claims to have come up with a new set of postulates for QM that is based on stochastic processes (that does not involve wavefunctions, etc.). RUclips video title: There’s No Wave Function? | Jacob Barandes. By the sounds of it, his new set of postulates supposedly re-creates the usual QM postulates, etc. However, his new postulates takes the "mysticism" out of QM (like wavefunction collapse, measurement ambiguity, etc.). I'm a little skeptical about his claims. I have a feeling that his new postulates simply swaps one set of mysticism for another (while still retaining the same kind of ambiguities and weirdness). It would really mean A LOT to me if you were able to take a look into this and let us know what you think. Please and thank you.
In the whimsical land of Far Far Away, Shrek and Donkey discovered a connection that transcended friendship. Amidst their adventures, a deeper bond formed, defying societal norms. Their unconventional love story unfolded quietly, a tale of acceptance and understanding. Far from the conventional fairy tales, Shrek and Donkey navigated their feelings in a world that had yet to grasp the diversity of love. In the end, it wasn't the castle or the dragon that defined their happiness, but the genuine connection they found in each other, proving that love knows no boundaries, even in a swampy fairy tale realm.
One AI program a guy used to develop a wild alien looking rocket motor. It was designed for maximum laminar flow and efficiency. Mixing of air and fuel like a carburetor. Copper based alloy seems to withstand the heat well. Thank you.
AI/Machine learning also deals with (discrete) probability distributions, I'm curious if there will be an opportunity in the future for AI to be applied to quantum computing in some way to reduce the probabilistic search space that quantum computers deal with
I've been looking recently into quantum-inspired machine learning, tensor networks, MPS etc. and it really seems that the fact that quantum systems "present to us" a highly-dimensional, tensor interface doesn't mean they are highly-dimensional "all the way". Internally they can be equivalent just to a bunch of matrix multiplications.
Most quantum algorithms currently use quantum+ classical optimizer like in VQE. AI like AlphaFold requires substantial computing power utilizing HPCs with multiple GPUs and TPUs in large scale data centres. So overall computational and energy costs could be large .it might be cheaper for startups to run quantum algorithms than AI on HPCs
Many times, I have posted on RUclips that there are two different quantum states in the atom: the real and the hidden quantum states, also known as hidden variables. I even posted that it is impossible to understand quantum mechanics if classical mechanics is totally dismissed, because the hidden quantum state is ruled by the classical mechanics. Physicists commonly mess up the properties of quantum mechanics with that of classical mechanics or vice versa. But this time, I would like to share how I understand the inferences behind the atomic mystery in different way. Every complicated atom having two or more electrons behaves like an inductance-capacitance oscillator or LC-oscillator for short. Quantum mechanics can explain the hardware part of the LC-oscillator, but classical mechanics can explain the software part of the same, which is the energy content of the same, that can be changed continuously. So, if we hold the values of the inductance and capacitance of the said oscillator constant, the frequency of the LC-oscillator remains the same no matter how much energy we put in into the same. Going back to atom, similarly even if the electron in question doesn't jumps, the hidden quantum state remains vulnerable and susceptible to the influence of the outside world, which means that the energy content of the hidden variables can still be changed by the outside observer without quantum jump involved in the process. This is the same phenomenon that brought so much trouble to Erwin Schrodinger that eventually led physicist to introduce a scenario also known as the "Cat in the Box" that can be both dead and alive at the same time that remains a paradox up to this point of time. But based on the inferences, actually there are many cats, but one of them is a pregnant cat, the rest are just unborn kittens. The pregnant cat can be described by quantum mechanics; and the unborn young cats can be described by classical mechanics. However, the very act of opening the box forces the pregnant cat giving birth to a kitten. This phenomenon is analogous to breaking the circuit of the said LC-oscillator where the breaking point generates a spark that signifies the kitten, the young cat. So, all these explanations suggest that the existence of hidden variables or hidden quantum state or classical changes is mainly responsible for holding all the information we need to know about some of the things we already know and understand like the binding energy between two or more atoms combined together to form a molecule, the spread of spectral lines, superposition, entanglement including chemistry and biology, just to name a few.
About a decade ago I read a paper about how a team was able to use insights from quantum systems to develop an algorithm for classical computers that beat even the quantum alternatives. Since then I have been wondering what else we could learn from quantum systems that could make classical ones better.
Did I say something recently about chemistry? It is worth studying as its own subject - also quantum chemistry which models only electrons (Born-Oppenheimer approximation) as electrostatic potential clouds consisting of occupied molecular/atomic orbitals (only using probability density as any QM system).
Energy is key. QC are supposed to perform computing using Q principles directly, which means efficiently in terms of energy. That is especially true to compute Q things using QC. Try that with conventional machines eventually running AI. Don't forget your wattmeter.
you need more silicon than there are atoms in the observable universe to accurately simulate 50 logical qubits on a classical computer, many use cases for quantum computers need in excess of that, no matter how smart an AI is, if its built out of silicon semi conductors, it is physically impossible run shors algorithm on large numbers. The funny thing about Quantum computers is that there are quantum algorithms which can improve LLMs, so its not like the two have to be mutually exclusive.
As an aside regarding classical modelling of quantum systems, it is important to remember that nature herself has no agency and doesn’t perform maths or make decisions etc, it just does. So unless we are living inside a simulation (which can have agency), nature’s interactions from quantum to cosmological are seamless. This is the fundamental error in the wave-particle duality hypothesis - nature doesn’t know of either scenario - it just is and does irrespective of our own context. Classical versus quantum is our dichotomy but it certainly cannot be natures.
We mathematicians have been using this phenomenon for some decades now. We kinda don't care whether P = NP anymore because in real life systems follow sets of rules and you can exploit those rules to manufacture a heuristic that will give a solution in nearly polinomial time. Neural Networks (read "AI") are merely one of those heuristics.
I’ve been intrigued with the idea of melding quantum computers, or I suppose very advanced and efficient AI, and evolution design wherein you specify the parameters of a final design - what you want it to do - and within the physical constraints of what is being modelled the system sort of brute forces many designs that satisfy the desired parameters. Some type of evaluation is done on each output, prototype testing in a virtual environment for example and the worst designs are rejected. Over time the number of designs narrows as they evolve to better satisfy the desired constraints. I guess the idea is that if you have some way of efficiently calculating a very large number of options and evaluating them the best designs will “float to the top”. Obviously like with Alphafold you only need to compute things that are within what’s possible so maybe quantum methods aren’t needed just super efficient ways of modelling
AI absolutely can act as a means to focus searches to enable good results to be found using traditional means, in the way mentioned for chemistry. BUT ALSO, AI approaches may provide the necessary traction to allow quantum computing to become properly realized. I wouldn't bet on this being an either/or situation.
Of course they can. Intelligence is essentially the ability to make predictions based on memory, which involves drawing from an abstract model of reality. While a computer operates using literal zeros and ones at its core, the information it processes exists on a higher, abstract level. This abstract realm is what we associate with thought-it’s the relationships and patterns between the zeros and ones that give them meaning. These patterns form the context and structure, shaped by their sequencing in time and their relationship to the model they represent. In this way, classical computers can indeed model anything a quantum computer can, because the abstraction transcends the physical differences in how they operate.
The point in Quantum computing, is not only that the randomness is innate. That you do not need to use some pseudo-random number generator code. Yes, it is one advantage, but the main advantage in Quantum computing is that you can form a algorithm to solve a problem like a multi exit obstacle track for electrons and then measure how many electrons come out of each exit. Thus you can solve every problem which can be translated to such a track immediately as if it was run in parallel on many cores in a classical computer. So yes, a classical computer can emulate it - but in a classical computer you will still need many cores running in parallel to do it.. That's the whole point. The question of why this is important demands an example - well, one glaring example is code breaking. In code breaking you frequently need to find rules that will diminish the option space to be more manageable, and AI can help here - for sure. But in many cases, the remaining option space is still huge. And so, with only classical computing to break a code, you will still need hundreds of years, even with AI. Enter quantum computing - quantum computing can enable one to produce an 'obstacle course' such that in one case the code is getting one step closer to being decoded without the need to employ a parallel fleet of GPUs working for years on it. Suddenly the problem becomes trivial. So yes, in this specific example, AI may still help (as in most if not all examples..). But AI cannot solve code breaking alone. Here quantum computing is of the essence. And there are many other problems in which QC may be necessary. However, yes, many of these problems are marginal, in specific fields, like code breaking.. I reluctantly have to admit that much of the fan-fair for QC was about things that AI had recently demonstrated it can handle perfectly well on classical computers (such as folding of proteins). It doesn't mean "AI Could Make Quantum Computing Obsolete". It could mean AI could make Quantum Computing into a marginal field of research specific to special cases such as code breaking where it cannot be done without it, but otherwise it would be redundant.
Solving the Schrödinger or Dirac equation for quantum chemical computations numerically is essentially non polynomial in system size and accuracy. If you are out for analyzing subtle chemical patterns of structure and reactivity you feel this limit even today. There will be a certain class of chemical problems remaing unsolved by classical methods, especially when a deep understanding is aimed. Not sure if quantum computing still couldn't be the answer to that.
I shared this with a friend who works in exactly this area - and these were her comments: TL;DR: Sabine mischaracterizes the problem ---- She brings up some points which are kind of right but arrives there through wrong interpretations. My entire job is studying the crossover point between quantum and classical for chemistry. In general yes, if classical simulations continue to improve such that the crossover point between classical and quantum is never practical, then quantum computing will not be useful. Most of her arguments hinge on the ”space of possibilities” being why quantum computers might be advantageous, and that is incorrect, so a lot of her arguments (like comparison to Alphafold) fall flat. The main challenge for chemistry isn’t the expressibility of a quantum or classical model, but how efficiently you can navigate through that possibility space. I didn’t see that talk by Demis, but from the clip shown it sounds like he is referring to complexity class assumptions that BPP (classical) is contained by BQP (quantum), or NP being contained by QMA. This definitely would be deep and crazy if that hierarchy were inverted or it collapsed. In general we look for quantum advantage in specific problems where we believe it is NP hard classically but BQP quantumly. This won't be true for many chemistry problems, so we target specific chemical systems which are known to be difficult classically, such as strongly correlated systems. Its like how you would never target quantum advantage to do addition, because that is super easy classically. Instead, factoring into primes is hard classically but has been proven to be in BQP using Shor's algorithm. So In that case it is interesting to study the crossover point between quantum/classical. So what she said about quantum not being useful for all of chemistry is right, but that was never the goal. It looks like the excerpt with Carleo Guiseppe quotes was leading to the question with wether training a foundation model on systems that are easy to simulate classically can generalize to more complex systems. This is an active research area many people are optimistic about (I'm also involved in this), but nothing big is out yet. I don't know much about Sabine but this made me not trust her videos on content I don't know about 😅
Current computers are still very sequential in what they do but the trend is towards making them more parallel. This trend in some ways makes an additional threat to quantum computing. In parallel systems stopping to decide what to do next is a burden but if you can design the system such that the need to decide if rare, then it tends to consider more cases in parallel much like quantum stuff considers many cases at the same time.
Sabine, I think you are confusing two different problems here. The most of experts thought that computer or AI could never master a complex game such as Go. The rationale was that the total possible move that can be made in Go is simply mindboggling (10^180) and calculating that many possible moves would be impossible. This rationale was flawed because we humans don't calculate all the possible move combinations either. Rather we learn and adopt the game based on the patterns and pattern relations. What Denis Hassabis is talking about is related to this. Even generative AI works on this principle as well (attention mechanism and transformer architecture for example.) However, there is a more fundamental problem that needs to be solved. As Richard Feyman said in his 1981 speech, there is a fundamental mismatch of a deterministic machine trying to simulate a probabilistic nature. And he is right. The current AIs are called the Boltzman machines for a reason and this mismatch will only get worse as AI advances. For one, there is an enormous resource and energy requirement need to do the AI work and this has to do with this mismatch. I am not familiar with the works in Quantum computing but you can be certain that probabilistic machines will rise sooner or later. If it is not Q-bits, there will be P-bits, O-bits, or whatever but they will come since the current mismatch will stop the rise of AGI.
I think, that instead of fumbling around pure states of qubits, more promising branch of quantum computing is ising model optimisation. Which doesn't really give so much damn if something lost coherence. Only most stable state will persist anyway. And could specifically solve hard problems of optimisation, which are necessary for AI. Also, quantum computers could very successfully screw about lattice problems. Yes, there you need quite more stable qubits. And a lot of them. But lattices are quite a powerful things to screw with a lot of cases of logical tasks. For example, it's trivial for LMB solver to find decomposition of binary string into xor of several periodic signals.
Interesting. Thanks for talking about this, Sabine. I'd like to hear IBM's and PSI Quantum's take on this, too, because they're investing a lot of $ in developing quantum computing, although both with different roadmaps. Perhaps combining AI and quantum computing could lead to more powerful computations?
The only serious application of quantum computing, arguably the one it's being developed for is prime factorization - being able to do it quickly would break (some of) asymmetric encryption schemes used on the web and this would allow to decrypt past online communication (as long as it passed through nodes you controlled and you stored it in your data center). I don't see AI helping with that unless it shows that prime factorization can be done in polynomial time (of reasonably low degree) or that P = NP.
A while ago: tenant models with selective goal ranges can approximate quantum systems- this still has a nesting problem with float errors that cannot be resolved. This is baseline approximation error.
like alphafold once the inflection point of translating atomic/molecular structure into human scale physical properties at any given temperature/pressure is reached...
I would love for you to weigh in to Jacob Barande's non-markovian non-divisible classic stochastic approach to quantum mechanics. Please make it digestible to common folk 😊
2:53 your lines should not be starting at the same point. "Conventional" computing just sped up existing logic that was accomplished using electromechanicals before the vacuum tube, then miniaturized it with the integrated circuit decades later. Quantum today is like computing before the IC - quantum still needs its great scaling breakthrough allowing logic to be constructed as true VLSI. Current quantum processors really looks like where CORE memory was in the 1960's - the concept for how to use it at scale seems to be getting worked out, but the very edge of what can reasonably be manufactured is ~1k bits (or qbits).
I think you find that heuristics and classical computing are good enough to predict qualitative folding structure but not good enough to predict shapes, energy levels or binding strengths to the quantitative accuracy needed for drug development, biosystem or large-scale nanosystem design without testing.
This good science for me, because there are archievable goals to get. With KI based material-science maybe we find something very useful, like supraconduct on roomtemp on day. With this then, we can go further and so on. This is more important to focus on than any other "old unsolved" problem.
Born too early for quantum computing
Born too late for quantum computing
Born just in time for quantum computing
First, prove that you were born.
@@RFC3514 Not even they can be certain.
Schrödinger's computing for sure
@@lifeinanutshell7147 Literally both 1s and 0s at the same time.
nerd humor! +1
Quantum computing will neither be a success or a failure, it will forever be suspended in a state of uncertainty.
By definition.
Heisenberg Computing?
Yes, we know where quantum computing stands, but how fast is it going?
😅
It's more and more becoming a scam, similar to nuclear fusion reactors.
"On the other hand I'm on RUclips...". Hilarious.. Love your humor and your channel.
Its kind of tragic, though glad you find it funny. She's on youtube because she's a *failed research scientist*. She failed because her colleagues all over the world found her ideas to be at best, half-baked.
I was not looking at it from the standpoint you are considering. My take was comparing someone who got a Nobel prize with someone who has a RUclips channel. No reasonable person would ever expect to get a Nobel prize no matter what their research area was. I’m sure, or at least fairly sure, that this year’s physics winners never expected to get a Nobel prize in physics. So I don’t think Sabina was looking at things quite the way you depicted. Just my opinion.
@@datamatters8 Nobel Prize winners in the hard sciences usually expect it, largely because they're informed of their nomination months ahead of time.
@@yttrxstein4192 bitchy much?
Dry humour, because I don't laugh until I cry, but boy do I laugh, and laugh and.................
The principal competition between AI and Quantum Computing is for speculative funding.
The real correct answer, always boils down to money.
The difference here is that AI is already, very clearly, getting results. See AlphaFold for example.
So true!
Why is no-one speculatively funding you. :(
They should start selling Quantum AI NFTs.
As a technician I have often found that I and other technicians occasionally run into technical system failures that we say "has teeth". That means after going through a troubleshooting process, many times, we can't seem to find the problem (s) to fix it. That is when we either set the device aside, and go work on another device and repair it, or have another technician come over and start the troubleshooting process. Either way, the problem is found pretty rapidly. Why? We in the initial troubleshooting process made an assumption that some component was good. Didn't verify it. A.I. systems, by narrowing down the possibilities might skip over something that fell outside of the narrowed scope of solutions available. That might be where a Quantum computer can be the refreshed eyes looking for the solution to a problem that eludes A.I.. That hybrid method of search might be best as it only gives the Quantum computer a limited number of problems, and lets A.I. find the solution quickly for problems that will fall within it's scope of applied rules.
Reminds me of a guy I knew in the air force. He was extremely intelligent... our career path was already full of intelligence, but this guy was miles beyond anyone. He checked out a piece of electronics that wasn't working. He found nothing causing it. So he checked each individual board, and found nothing wrong. He desoldered each individual component, checked them all out one by one, and still found nothing wrong. This occupied his time for over a week.
While he was away, an airman came over and replaced a fuse in the lab bench. It worked perfectly. There was never anything wrong with the unit. The work bench power supply had a blown fuse. so the unit never had power. It's just one step away from not plugging it in. It's a common intelligence trap... the assumption that a complex problem must have a complex cause, or require a complex solution.
I conveyed this in a comment I made responding to the video, but I figure I mind as well just put it here instead because we're on the same track of thinking and I agree with your comment:
I don't know. I feel like AI computing could replace some aspects of quantum computing, but I feel like it'd just offer another level on top of the bias that human researchers and scientific experiments already face. Neural networks are modeled after our brains, and work by placing weights to different nodes and on different edges, if an AI sees a solution resulting from the high weight of a specific node then what happens when it doesn't consider another node which is just as relevant. Moreover, what if it continues down the path of a specific edge trying to justify it's theory rather than consider it might be wrong? We already see this in ChatGPT and other large-scale neural networks where they can become so convinced in their interpretation of the input stream that they'll actually ignore things they should be doing or considering because they refuse to readjust their "thinking."
quantum problems (NPComplete) don't get solved by not skipping what humans would skip, but by trying every possible state at once which humans (or classical computers cannot do).
And the problem you mentioned doesn't need quantum or AI, it just needs to work without assumptions in a completely classical way - simply check everything that is checkable, the possibilities don't go exponential... if they did it would not be solved in the way you say it does.
The problem is quantum can still do that in almost the same time as that hybrid…
@@jdlechdoesn’t apply cuz unlike your airman quantum computing is like 100s of airmen and they will arrive at the solution at the same time as one
Veritasium recently explored similar ideas behind what contributes a human to be an expert in chess and other professions, the chess master is not memorizing each chess pieces position but rather chunking the configuration/rule that constraints them to their positions
The chess analogy is very useful. The chess grand master has a better set of neural network algorithms for chess than the novice. I stopped playing chess when I realised there was no luck involved and that likely there were plenty of people with a better network than I had! I prefer a little bit of chance to mess things up.
chess is merely about memorizing. trust me (i'm grandmaster)
@@mintea-i5o - it’s interesting to wonder whether human memory is neural network related too! In computers it wasn’t / isn’t.
@@alwayscurious413 The question of whether human memory is related to neural networks touches on the very core of how we understand ourselves and how we build machines. Human memory seems to work in ways that are surprisingly similar to artificial neural networks, but with some critical differences.
In humans, memory is distributed across the brain in networks of neurons. These neurons form connections that strengthen or weaken over time based on experiences, much like how weights in a neural network are adjusted during training. This process, called synaptic plasticity, allows humans to store complex patterns of information, associate ideas, and even "reconstruct" memories based on partial inputs, just like how AI models complete tasks.
But here’s the twist. Unlike computers or artificial neural networks, human memory isn’t exact. It’s deeply tied to emotion, context, and even sensory inputs. For instance, a smell or a song can suddenly trigger a vivid memory, something machines can’t replicate yet. On the other hand, computers and traditional neural networks focus on efficiency and precision, storing data in a way that is either "on" or "off" and lacks the nuanced imperfections of human recall.
To address the comparison to traditional computer memory... No, standard computer memory doesn’t operate like a neural network. Traditional storage is static and precise, like saving a document. Neural networks in AI, however, are more like human brains they’re dynamic and learn through experience, but they still lack the depth of our biological processes. For example, neural networks don’t “forget” the way humans do, and they can’t yet replicate the emotional weight we attach to memories.
Just take the conventional route until the first right, then follow the quantum path?
AI and quantum computing are not in competition but are synergistic technologies. When combined, they can solve complex problems more efficiently and effectively than either technology could alone, enhancing each other's capabilities
Yes, that's how I have been thinking about it. Now I am wondering!
If ur total amount of money can be put into invest either one is the same, they are competing
I don’t know how to feel about the statement that they are synergistic. But one thing is for certain, the theoretical groundwork for the current Gen of AI/Deep Neural Networks has been laid the past ~50 years or so. Quantum Mechanics itself has not even existed that much longer.
What I am trying to say is that AI just has a headstart right now. But the field seems to be narrowing as many low hanging fruits have been picked already. For QC, we don’t even know for certain how to use it yet and whether we can make good use of it at all. Hypothetically, the next breakthrough could come any day.
As a total outsider, I've also thought the same way recently. Now I wonder.
I think @@SanguinaryOtter may be right about what he said. But the question arises that how much money and time should be put into it. There are too many possibilities.
I think It can even happen that QC will be practically dropped in a decade or two, then later gaining new life; kind of like how electric vehicles did or how natural and smart architecture is now gaining new focus, after the realization that modern glass and steel electricity dependent structures don't have much for them except a phallic coolness factor. To provide two quite different time frames as examples.
Edit: Question for those who know it. How different is too write algorithms for QC (if any) compared to today's standard? I imagine some difference is required when writing for them but I would like to have an educated answer.
For investors, they _are_ in competition. The question is which bubble grows faster and which will be the first to burst.
Both self-limiting data and negatives prune analysis. Reduction of questions/operations are the key.
Even the “tree” example is easier if reduced. Trees grow upwards and outwards. Leaves don’t grow from roots. There is a limit to the growth rate. Gravity/stability of the system are also self limiting. Environmental considerations can be applied: Altitude, wind, fire, lightning, walls/cliffs/ceilings, human pruning, animal damage, soil type, substrate issues & obstacles, considerations of time (always forward), etc.
Now if there was a true fractal, that would be quantum computing.
Each probability can be placed in its own statistical set.
2+2=4 (most if the time), unless one or both 2s are negative, then it would be zero or -4.
Together with deep-learning, hyper-parallel algorithms are also emerging as a serious rival to quantum algorithms. Recently, problems seen as strong candidates for quantum advantage, like optimization and combinatorial tasks, have been solved faster using classical algorithms that leverage massive thread parallelism. That is a major problem for quantum firms since GPUs are becoming exponentially more numerous and powerful while quantum computers are struggling to do anything.
Interesting, thanks for sharing!
Not really, qc specific algo has some serious advantage compared to classical machines. More gpu doesn't mean classical system will beat quantum systems. Say specific problem like search using grover is always going to be faster on quantum systems with quadratic speedup as compared to classical.
yeah but this isn't due to the model of computation , but rather because gpu architectures have decades of work refining them.
this is the equivalent of comparing two versions of an algorithm (sorting for ex) , one logarithmic and the other linear , and have the linear one benefit from hardware optimizations , like better cache access patterns , SIMD , SIMT etc.
the linear algorithm is going to be faster , up to a certain point. as fancy as GPUs are , they are still classical computing machines, even if there are millions of threads.
Ultimately the issue is that eventually, quantum systems should have an edge, especially in terms of power efficiency. A lot of our advancements in conventional compute are coming through largely brute force methods, and we're rapidly approaching the miniaturization limit for conventional compute, while we're far from such for quantum compute. I believe that conventional compute will maintain an advantage for the next decade or two, however when we hit that miniaturization limit, any significant advancements in compute density will come from future advancements in quantum compute, as architecture improvements have historically been far less impactful than miniaturization in terms of efficiency gains by processor generations.
@@SCComega Agreed we can't just keep on making things faster and smaller indefinitely without facing off with laws of quantum mechanics
It's like saying nothing will outrun a Horse when cars first came out in the late 1800's.
The physics of a car were sound
@KieranLeCam this is an engineering problem. The physics pan out.
@@OBGynKenobi they're incomplete imo, until we reconcile GR with QM
When I was a young teenager, my younger cousin and I had the nerdiest debate with each other. It came after I wondered aloud when in our lifetime we would see computers that could think. He was completely against the idea saying computers can only understand yes or no, ones and zeros, and fundamentally could not understand “maybe”. It occurs to me that AI is exactly that- a way to model on a deterministic system a way for it to learn what is probable. AI does understand maybe and maybe there is some link between how the rules that govern classical computing work and something fundamental we are missing in the rules of quantum mechanics. As for my cousin, he went on to become a physicist, so that’s his domain, but that argument we had long ago is what inspired me to become a computer scientist.
I like this, thanks for sharing😊
Great way to look at it! Classical computers may not be quantum in the strictest sense, but if they model it closely enough does it really matter? It shouldn't really matter how you get there as long as the output is useful and the speed is acceptable.
Maybe 🤔 😏
The big issue with science and philosophy today is that we don't take seriously the ways in which systems scale and produce new systems out of underlying regularities and irregularities. Humans are the product of a number of phase transitions from quantum > physical > chemical > biological > intelligent > sapient.
Each higher transition constrains its environment to ensure its propagation. One could think of the Big Bang as the physical beginning its process of self-perpetuation out of the quantum, just as the biological began self-perpetuating out of the chemical, and sapience (aka memetics), began self-perpetuation out of intelligence.
Each of those levels follows its own independent rule set, and yet also constrains & structures the underlying system. If I had to pick a field of study that would benefit humankind more than anything, it would be studying the structure of those phase transitions and discovering their patterns and limitations.
Each of your brain cells are not conscious on their own right.
Quantum computers maybe able to quickly solve combinatorial optimization and problems with exponential complexity, but ironically building a quantum computer with enough error-free qubits itself becomes exponentially harder as scale increases.
Willow says hi
Hii, just a minor note for Alphafold at 3:50, AF actually does not use physical rules, it generates model based on the similarity to the training dataset of solved structures... so structures closely related to something that has been solved before are more accurate that those distantly related.
How do you think it comes up with a way to model the data? By coming up with rules it finds in the data. It's not some basic interpolation.
@lost4468yt how do you know it’s rules are correct. That’s the big problem with ai it’s very good at looking correct instead of being correct
@@lost4468yt yeah but those rules it finds are not physical rules, those are distances between evolutionary conserved residues in the protein structure, this means that is is very bad in prediction of unfamiliar proteins that are lacking evolution conservation in the training data .. .ergo, stuff that is interesting.. not even touching protein complexes or nucleoproteins
@@thorondor666
Exactly! The physics are NOT there. You can not fold a protein de novo using alpha fold 2 or 3, etc., due to a rule (e.g., hydrophobic effect) that was uncovered using QC. Evolutionary coupling and knowledge bases (alpha helices, beta pleated sheets, etc.,) from the protein data bank can give surprisingly good results when producing protein models. We’re a long way off from computing delta S of the free energy equation.
I still think the term "obsolete" is a bit of a stretch considering fields such as cybersecurity, specifically post-quantum encryption exist and offer challenges unique to quantum computing. Quantum computers will still have their uses in fields such as these where deep learning at the moment seems too different, and it's only a matter of time before new problems emerge that deep learning can't quite learn.
Not to mention the large concerns about power draw that arise concerning deep learning and overall scalability. Supercomputers at the current moment are more appealing but I hope to one day see that change as we approach a plateau.
Sabine nailed it again. (And I bookmarked this contribution) Quantum chemistry has my interest as I did a minor once (a long while ago) in quantum chemistry. I was always fascinated by the idea that you could "calculate" some atoms and to a certain extent molecules. Just the idea of "in silico" chemistry is fascinating; it brings about new molecules and understanding of the workings of proteïnes and even vaccines (Covid e.g.). On the other hand I was able to look from the outside in to the emergence of quantum computing. It soon appeared from theoretical considerations (computer science and mathematics research) that quantum computing would be efficiënt only to a certain class of problems to solve, not so as a generally better approach to all classes of problems.
As for the AI angle in this all I don't know, because "AI" is a container concept. Does this include molecular dynamics calculations (Karplus 2002) for instance? AI is a grab bag of (sometimes misunderstood) approaches and methods, and often "rediscovered" by other scientific disciplines.
Addendum. Researchers from Sandia Labs (Kallaugher and Parekh) found: " ... that quantum computers are unrivaled at solving an advanced math problem. Unusually, they proved quantum computers are not faster than regular computers; instead, they use far less memory."
Was this video implying that quantum computers are better at solving quantum mechanical problems than conventional computers? I don't see why that would be true. I was also under the impression that one of the advantages of QC was an increase in security of data transmission.
I just want to know which one can most persuasively arrive at the wrong answer.
In that sense they're opposites. AI will confidently give you the wrong answer (rephrasing it every time, but never admitting that it just doesn't know), whereas quantum computing will tell you "probably, like, I'm 99.9999999% sure, but I could be wrong" even when you ask it to check that 1+1=2.
@@RFC3514 That is purely just a user-interface thing. Under the hood, AI gives answers in terms of a probability distribution of the likelihood of different possible outcomes. The user-interface is then designed so that the AI will just tell you whatever possibility has the highest likelihood _is_ the answer. You could just change the user-interface and have the AI spit out things in terms of probability if you want.
Your claim about how quantum computers would answer 1+1=2 is also factually wrong as you can trivially write a quantum algorithm that outputs the answer to addition problems deterministically without any uncertainty.
@@amihartz - No, it's not just a UI problem (although some AI companies looking for funding would want you to believe that). Traceability is a *fundamental* problem with many models. Many of them literally can't tell you how or why they arrived at their "conclusion", and are truly convinced (like, close to 100% certainty) of complete falsehoods, while being _less_ convinced of some true facts. Simply showing the user a "confidence" value won't fix the problem.
@@RFC3514 How is that relevant to literally anything discussed here at all? The fact an AI cannot tell you how it arrived at its conclusion has zero relevance at all to anything here. A quantum computer also cannot tell you how it arrived at its conclusion. This is just something entirely outside of the realm of this discussion you are trying to arbitrarily inject into it with no explanation.
@@amihartz - I simply addressed your (incorrect) claim that "it's just a UI problem".
Also, there's nothing "being discussed here". The OP made a joke, I played along with it, you posted something that simply isn't true, I corrected you, and you got mad.
I would say, the most useful applications of AI are going to be the ones that we needed the quantum advantage to solve, but not just because Quantum Computers are late, but because AI can not only "solve" but also "generate".
For example, AlphaFold can predict protein folding, that's amazing. But just recently it allowed to make AlphaProteo, which can suggest brand new protein that bind to a specific site of an existing protein. The possibility to transform a "solver" into a "generator" is incredibly valuable.
I don't know if an equivalent strategy exists for Quantum Computing?
This is a question I asked but have gotten no response to. Isn't the title of this video a false comparison? Conventional and quantum computers are tools. AI is an objective. Isn't the video conflating AI and conventional computing? It's like saying I have a gasoline vehicle and and an electric vehicle. Will driving to work make my EV obsolete? It doesn't compute.
Quantum computers are inherently probabilistic, hence are well suited to generative models.
@@chilinitro5599 : Whether they are the best tool at the moment isn't my point. Comparing AI to quantum computers is a false comparison. Computers are tools. AI is a desired result of using those tools. Like comparing a hammer to hanging a picture.
@@lrvogt1257good explanation, I agree. I meant to reply to the original comment about solve vs generate. Quantum can certainly generate.
Good show. It's nice to know you can agree with someone. 😊
What I find fascinating is how Reinforcement Learning (RL) is demoted as "too inefficient", yet it is used in Nobel Prizes winner Hassabis's AlphaFold and AlphaGo work, and Yann LeCunn often says to "get rid of it", but only use it if you are "fighting a ninja" or if your "plan does not work", it is just a "🍒on top of a cake".
RL is the way to AGI. RL made ChatGPT possible. DeepMind is on the right track and has been for a long time.
Not a cherry, but a Strawberry (aka Q*). And nobody knowledgeable has questioned the utility of reinforcement post-training of transformers since OpenAI's release of o1-preview (on September 12, 2024, comparable to no day in my life but July 20, 1969). As for computational expense, recent experimental versions of Gemini have risen to the tops of leaderboards with just 32K-token context windows. This suggests that relatively small transformer networks will be used in test-time deliberation ("reasoning") to generate high-quality responses.
@@tomenglish9340 I agree, the method of "train, then infer/use; retrain, then infer/use" only goes so far and seems to work best for "autonomous nervous system-like" performances, even ones that require high levels of skill. But learning by doing - through experience and practice - and being able to recall and relate that when confronted by new contextual or situational "features" presented by edge cases seems essential (and should be faster and more efficient as well). I would call viewers' attention to a recent Know Time #97 RUclips video featuring a conversation about this very topic with Joscha Bach entitled "Artificial Intelligence, Consciousness & Computation."
LeCunn is an idiot.
I think of this as "not all functions are continuous". There are constraints everywhere and if your model is good at figuring them out, convoluted as they might be, you cut the search space significantly. ML has been used against cryptography before, and that's one of the places where you simply need to reduce the search space to something more tractable for brute forcing.
Yeah as a computer scientist, it doesn't even come down to concepts or ideas here. There is a physical difference in how classical computers compute, and they have physical limits. An actual functioning quantum computer will always be far more powerful than a classical computer with the same number of bits. The problem is the functioning part.
@@justmillenialthings So much this. The AI hype and all the utter ignorance contained it annoys me out of my fucking mind.
I thought quantum computers could be much faster than an ordinary conputer. Because it can occupy all states simulaneously.
As a computer scientist, you should know then that the potential of classical computers hasn't remotely been tapped. The physical limits (two state transistors) are a transitory problem, are they not?
Quantum computers will always have the problem of having to be suspended in a very specific state to even function properly. It is a practical limit which may never be overcome.
Yes, but this extra power might not be practical. For example, we could have most problems of industrial interest to admit decent approximate solutions on classical computers (this is where AI excels), then there's little need in quantum computers. Or, maybe scaling up QC will end up unfeasible: while theoretical considerations clearly show asymptotic quantum supremacy, the actual point where QC starts dominating everything classical (e.g. AI) on a financial scale might happen too late.
@Meditations2024 so classical limits are solvable, but quantum ones aren't, and your basis for that is? But no that gets into computer engineering and physics and I'm not a computer engineer nor a physicist.
Not needing all the possibilities reminds of compressive sensing when representing images (photographs). The permutations of pixels in an image space is unfathomably vast, but only a small subset of that space is actually useful which contains "natural" images, the rest is noise. Knowing this, it's possible to reconstruct images from very sparsely sampled random data, using clever minimisation algorithms.
I think a lot of people in the comments here fail to understand what the real advantage and applications of quantum computing are. Yes, some quantum algorithms only solve problems with *how* we are using classical computing, but the actually interesting ones are the ones which solve fundamental problems with classical computing, and they simply *can't* be modeled by AI- any accurate model is at most as efficient as the equivalent classical computing algorithm or less efficient.
Can you give examples of problems you are talking about?
What fundamental problems
so, if its impossible then its impossible?
nothing is impossible, if you look back at 2000, and bring chatgpt to the world, the world would be shocked, because at the time, everyone thought LLM is as useless as a roach, but now it is not.
You mentioned only in Quantum Computing, If you do search for Quantum Sensing and Quantum Communication it’s also an highly evolved field
@@sans00xa good example is finding prime factors of a given number. The best classical algorithms are incredibly slow, which is something that makes it very useful for encryption. There are known quantum algorithms that can find prime factors by orders of magnitude faster. Get a working quantum computer and suddenly large amounts of modern computer security is trivially easy to bypass.
I believe the real question is how many *useful* algorithms are accelerated by quantum computing. As of now, I believe the only general purpose algo is really Shors algo. We already know how to do “post-quantum crypto”, so I think this is of historical interest only at this point.
So, one thing we *know* QC is needed for is….simulating quantum systems. But not all quantum systems, in fact just the ones where entanglement is significant, ie not decohered. Which, by definition, is surprisingly few / small systems, because if large entangled systems were commonplace then QC would be much easier to make!
Another way to look at it, for the specific case you gave in quantum chemistry: some aspects of molecular dynamics it’s really important to model the coherent delocalised effects - binding sites, transition states etc. But definitely *not* a 100kD protein. Because if it were, then making such a protein would be an excellent candidate as a functional unit of a QC. Instead, a protein of that size is well-modeled as a classical (stochastic) connected set of (coherent quantum) sub-regions. You’ve referred to that as “structural rules which AI can infer”, but the important bit isn’t the AI-ness, it’s the classical-ness.
Therefore, the size of a relevant coherent system is somewhat bounded. Which in turn limits the scaling advantage that QC can possibly bring.
I’m not saying there is *no* room for QC. Given that it’s currently really difficult to model quantum systems greater than about 150 atoms; but the fact that the largest useful natural coherent systems are probably no greater than few thousands atoms, kind of indicates a “slot” of usability in logarithmic complexity-space which is narrower than people have assumed.
There may be a silver lining in all this even if Quantum computing does not pan out. It helped scientists define new problems were thought to be only solvable on a Quantum computer. As you pointed out this could lead to innovative ways to solve these same problems on a classical computer. Even if Quantum computing does not become the breakthrough that was hoped for, it still has propelled computer science forward.
It reminds me of a seed company, I think Monsanto, created a number of crops using traditional breeding techniques, but based on their knowledge of GMOs, to create plants with very good properties but not directly genetically manipulation, so a sort of non-GMO plant with similar advantages as a GMO one
No one ever thought there were problems "only solvable" with a quantum computer. It's just some operations scale better. Just as different classical computers can scale in different ways...
@@lost4468yt Being precise is admirable. For all practical purposes didn't we just say the same thing?
I *think* I understand what you're saying @sabine but let's see...
So if you simplify it down a bit to a simple 3 dimensional array that's 10 units on each dimension so there are 10x10x10 (10^3 "possibilities", way less that the 10^40 in your protein example but easier to visualise in my head).
So we have arrX[0,0,0] to arrX[9,9,9] which we can visualise as a cube with 10 squares wide by 10 squares high by 10 squares deep, or perhaps better as a tree that has 10 branches at the base, and then 10 branches from every branch, and then another 10 branches below that... and the quantum people would go "ohhhh, all those branches and possibilities, you need a quantum computer to handle that"* but the "AI way" would be to learn from all the 10x10x10 arrays it's seen in nature and say, "yeah, well you can ignore everything on the [5,,] branch, that can't happen for whatever reason, so that reduces the space down by 100, and the [9,9,] branch is a dud, so that shaves off another 10, and you really don't want the [6,6,6] branch, so prune that one off too and so on, so that it turns out the answer is somewhere in the [2-4, 4-8, 0-9] space, so you're down to 3x5x10 = 150 possibilities, and at classical computer exaflop speeds vs quantum computer megaflop speeds this would be way faster to just brute force classically.
And this works in the "real world" because real world physics (and chemistry etc) has rules... molecular bond like to be at certain angles, proteins can only fold in so many ways but not every way, and so on. So we can "prune off" the "cannot occur" branches from our tree, and AI is really good at finding these hidden rules, hidden symmetries etc.
One counterexample, relating the geometric search space that robots move and Alpha Go. The reason that Alpha Go was able to significantly reduce the space complexity is because of Monte Carlo Tree Search. That is, the game go can be modeled as a binary tree. Each move is independent of the other and the set of all moves is equivalent to a binary tree. One starts with the first move and following a tree structure, one ends with the last move of the game. Then, the complexity space is reduced using backpropagation from the last move to the first. This allows a set of moves to be compared with each other using real numbers in orders to find the best set of moves.
This, however, can't be used when a robot moves. The best path that a robot moves is inherently a Euclidean distance and there is no geometric method to compare one set of robot geometric moves to another. The result is that it's computationally expensive and difficult for a robot to find how to move in an unknown and/or changing environment. This, also, is the reason that there are algorithms that can efficiently solve games and used in language and why there are no robots that one can buy to do your dishes or clean your bathroom.
So, my point, I'd like to see the exact reasoning of why AI can replace quantum computing. One method that's used by classical computer doesn't necessarily mean that it can be used for all methods.
AI produces simulacra of things that look like the real thing. However if you substitute this for the real thing you eliminate the possibility of discovering new aspects of the real thing you have not yet discovered and are not present in the simulacrum. You may even discover new aspects that are only present in the simulacrum but not in the real thing.
Excellent! I think the limitation of rules on possibilities also applies to other proposed solutions, like the Many Worlds Theory, etc.
Wasn't it John von Neumann who came up with the idea that any computer could be simulated in software?
You wouldn't need a 'genuine' quantum computer if you could simulate one in code.
Sure, it wouldn't run very fast, but it would be a faithful replica in terms of compute capability.
I think that's what the guy in the video was hinting at. For instance, a team at the university of Padova is trying this approach (which they call "Quantum inspired factorization") to try and break RSA encryption. Of course it scales worse than brute force right now, but it opens possibilities nonetheless.
So instead of using a classical algorithm that takes a million years to solve a hard problem, we take a quantum algorithm that can do it in a second, but then we run that in a simulator that will take a million years to simulate that second of quantum computing 👌
As the previous commenter says, this doesn't work.
The advantage of quantum computers comes from the fact that it can do certain (very specific) operations in constant time that take exponential time to simulate on a classical computer.
The point of Demis is more that you can optimize specific problems enough via machine learning to get fast results on classic hardware.
This doesn't generalize though, you can't (barring some algorithmic breakthrough) simulate a general quantum computer faster than the theoretical limit.
@@Clayne151 it takes a long time now, but research may find a way to speed up the solution of the equations they use to model the problem. It could be a better algorithm (like the FFT sped up DFT) or a different (non quantum) type of analog computer.
@@Clayne151 At least it would be stable...
Does this mean that if I ever want to find the shortest route to visit every town in Sweden to sell my door knobs, I will still need a quantum computer or can classical computers solve this conundrum for me now? It's kind of important to me.
Either way you have to hurry up, because superstring theory (or self-driving cars, or 3D TVs) could make door knobs obsolete.
ever heard of the Travelling Salesman Problem? There exist algorithms that do that with classical computers. I recently wrote code for a robot using this algorithm and it worked really well.
@@alexanderokak5112 Yeah, super inefficient on classical computers.
@@245trichlorophenate It's actually solved pretty efficiently on classical computers. Companies like Amazon can use GPS to get driving direction that route between all their stops before returning to the warehouse rather efficiently and quickly. Technically yes, a quantum computer would be in principle more efficient at solving it if they ever caught up to classical computers in terms of memory (a big "if"), but this difference would not be that significant unless you're like a giant enterprise like Amazon where slight improvements to routing can save you a lot of money. For the average person, it won't matter much.
@@alexanderokak5112 - If you need to ask that first question, I think your sarcasm detector might be broken.
Advancements on von Neumann architecture systems (like GPUs and TPUs and horizontal scaling) have made tasks like drug discovery faster and more efficient. Tools like AlphaFold and advanced AI models now rival quantum computing in these areas. If they become viable, quantum computers might enhance certain niche tasks, but they're no longer essential for breakthroughs in chemistry.
What tasks beside cryptography?
Classical computing has a two factor advantage over quantum computing
First it has a head start of many decades.
Second the prospect of new technology superseding it gives CP scientists a significant incentive to develop and innovative to leverage their technology.
Plus while low temperature environments are beneficial to CP they work just fine at ambient so there is that.
A strange parallel with air taxi research that Sabine have covered recently.
Ah yeah, the air taxi, and the research she didn't even look into. Amazon would use them directly, followed by Uber, and there's many more, when infrastructure is solved. Germany will wait another 70 years. When way over half of Europe has it, Germany might think about legalizing the regulations.
The final decision will be financial...which system can deliver business outputs at a cheaper price.
And for now classical computers have a global network of production and distribution.
Protein folding is a classical optimization problem that suffers from the “curse of dimensionality” and from having many local extrema. Folding itself is a dynamic process but we are interested in the final static structure of the protein in the equilibrium state. It can be solved by the “virtual work” principle that has been around since the 17th century if we have the computing resources. Using NN based AI to solve optimization problems smacks of a certain inelegance although it can narrow the space in which we would search for the exact locations for the equilibrium configurations. It is similar to using some heuristics in which NN acts like an “expert system” to invoke another term from the history of AI. Expert systems are basically a set of “rule of thumb”s; it is like sticking to a traditional family recipe when cooking and it contains a good amount of stubbornness.
My bet is on AI. Quantum computing has not produced anything meaningful despite the billions invested, whereas AI, on the other hand, continues to amaze us.
Quantum Computing is like Dark Matter in Physics an illusion ! 😂
Ehhhhhhh, OTOH amazing laymen isn't that difficult. There's a whole industry devoted to doing just that: marketing.
@@didiervandendaele4036
Where are you going after you die?
What happens next? Have you ever thought about that?
Repent today and give your life to Jesus Christ to obtain eternal salvation. Tomorrow may be too late my brethen😢.
Hebrews 9:27 says "And as it is appointed unto man once to die, but after that the judgement
"AI" hasn't produced anything meaningful either.
Yeah I was amazed by the "ai" that allows people to shop without the checkout. Oh but that turned out to be actual people making the decisions, oops!
The Science article tackles the problem of ground state preparation, which is a static state always difficult for a quantum computer, and possibly a suboptimal problem for a quantum computer to solve, while bypassing addressing the difficulty of dynamical evolution of states which variational algorithms are incapable, but has a major advantage in quantum computing.
Problem solving with AI is like trying to find a better search algorithm. While solving a problem with quantum computing is more like brute forcing the solution.
Both paths have their merits and drawbacks. And in the end, the two might need each other.
you have no clue how either work. please stop saying this stupid shit with such confidence
It all comes down to the "P vs NP" classification problem in mathematics. Many NP Hard and NP Complete problems can be solved. We can never know if we have the most optimal solution, as that would require checking every possible combination, but we can demonstrate that we have "a solution". I worked on this as an undergrad and managed to solve problems that were considered insolvable just 10 years prior.
I am for breakthroughs in any technology. There are many paths we have gone down that have not proved fruitful. To me, Quantum computing has been stalled for quite some time. I am starting to wonder that If it wasn't for government FOMO funding, would anyone still be pursuing Quantum computing so aggressively?
I would argue that path of least resistance is not always correct way to do things. Easy example is gasoline vs electric car. Just because we dont have a good ways to use quantum computing right now and its expensive doesnt mean it wont be the missing piece down the line.
Artificial neural networks are arbitrarily large or small systems of linear equations. This can be used to approximate functions. You just have to find a suitable approximation.
Now LLM's are being distilled to run on 1Gb of ram, and shrinking. I never thought such large training models could produce tiny, meaningful algorithms. Natural computation has potential, but neural networks are evolving right now!
That's just a normal feature of neural networks compared to deterministic algorithms: You can cut down existing layers to reduce size and increase speed at the cost of accuracy. The thing about LLMs though is that accuracy isn't very important so it's not that big a deal. Performing computations or autopilots like that where precision and accuracy actually matter would be a much bigger deal. We know that can be done but we haven't figured out how to reproduce it (housefly brain vs two flight computers on a Harrier).
@@DKNguyen3.1415
“Cutting down layers” vs “Reducing the number of bits allocated to any particular weight and computation” ‽
@@Mentaculus42 actually cutting down layers was my understanding since changing weights wouldn't reduce size or increase speed.
→ DKN
You don’t change the “weights” but reduce the number of bits allocated to each weight. So if the original weights were 32 bits long then reduce it to 8 bits and do the computation that is optimized for the lower precision. Maybe you can “drop” layers but that is something that I have not read about, or has not been explained how it would work. If one reduces the precision of the weights, it takes up less memory. I would be interested to know if layers can be dropped and a reference to that work would be appreciated. Always interested in learning new.
@Mentaculus42 i have to find it again.
AI might actually help to make quantum computing controllable by reducing the search space. And help stabilise the system. Because even if the AI follows rules and therefore reduces complexity there might still be rules we don’t know about and computation is needed to find them. This is where quantum computing might then have an edge.
I don’t agree with his take. A true quantum computer of scale, would simply have capabilities beyond any classical computer or LLM or general ai. This just seems like a re forming of the same old question that been around since the start. Can we scale the quantum computer and handle noise/ error checking efficiently. Nothing really new there, not exactly a hot take.
Quantum computer and AI Large Model are two completely different technologies. It is just that they try to solve the same kind of problem. AI through pattern matching and latent space can compress high dimensional data leading to faster solving, with some caveats such as it's a probabilistic approach, while quantum computing explore the whole problem space at once.
There is tons of applications of quantum computing which cannot be replaced by ML, such as prime factorization problems. Because there is no patterns to discover and no way to compress the problem.
So yeah, I don't think AI is competing with quantum computer, at least not on 100% of the problems we are trying to solve.
If quantum computers can do the unthinkable, how do you program for such 'miracles'?
And how could anyone verify the results, except by using another Q-machine?
It's like praying to God, getting an answer, then having to ask another God if the answer is correct.
@@pirobot668beta By clever use of quantum gate, which "pump" the probability for a particle to be in an specific output state.
You "force" entangled particles to behave in such a way that when you do the measurement and interact, they "probably" give a coherent result.
You then measure multiple thousand time, and often problems are not NP to check if the result is valid, so you can assert that the output is valid with a normal computer.
Easier said than done, plus it's very complex and I myself do not master the subject.
You can read the Shor's algorithm for example to see step by step how it works.
I didn't think that there was any question that Turing or von Neuman machines could simulate quantum processes. It was simply that quantum computing could do that and other calculations (particularly P and NP) much faster
Current AI is just an auto-correct on sterioids. It had already hit the dead end with no new training data.
guess what you are
They started that way, but the engineers are continuing to tinker with them in new ways so interesting emergent behaviors are probable.
That's true of large language models, but that's not the kind of AI referred to here.
@@randomYTwatcher Nope. Even the part we do know about how human brains work is VASTLY more complex than how puny little ChatGPT does things. Be a little more proud of the cells between your two ears.
@@Terigena Well you go ahead and tell me then, what architecture would you propose to model quantum algorithms with AI?
I’m pretty sure I’ve been saying the same thing as this guy but using words like relative and sharing. The tree is the foundation or a sharing dimension with the relative, the differences expand the relative like space and time. Space is the stuff and time is the difference, if the differences share then new dimensions can be revealed. Unconditional freedom is part of this but it’s still relative emergence but still connected to infinity. Honestly, the dimensions/space/stuff isn’t limited because it’s all relative, there are no separations just interference patterns. We listen and share and things that matter matter, know what i mean? The parts we share with are what reflects or connects, it’s why we feel empathy, the difference is relative even if unfound by the self.
I am a quantum chemist and this is gibberish.
The class of computer problema that can be solved efficiently or QBP is much bigger than the class of problems that can be efficiently solved by any classical Turing machine, or P.
Obviously all interest in quantum computing stems from their capacity to solve QBP problems that aren't in P, such as elliptic curve factorization and certain hard optimization problems.
In the specific case of quantum chemistry, most interesting problems we cannot tackle well, like light-matter interactio, also happen to be computationally hard in the QBP sense. For the class of problems in P we already have plenty computationally efficient algorithms developed over decades of research, that I doubt AI will improve much upon.
Had an inkling this was nonsense (having been following both techs for my entire 3 decade IT career) and always felt both are overhyped! Fantastic explanation; thank you!
I'm pretty sure you didn't understand the video. Like, at all.
It's not gibberish. the graph @2:58 is the key to understand as she mentioned. Where the intersection is (or will be) - we just don't know yet.
If we never reach the intersection then classical/AI computations will win out, otherwise QC.
That's not what the argument was about.
They are saying, since AI makes damn good approximations and runs in O(n) where n is the number of weights in the NN, even when dealing with NP-hard problems, is quantum computing worth all the money?
It's obviously going to have its situations where it is faster and produces a perfect output, but in other situations where it could have an advantage on normal computing you could just do a good ML algo with 80-90% accuracy and run it on one of the conventional computers (which run much much much faster than quantum computers) and call it a day, you would save a ton of money
That might be true, but the point is there are problems that are (for all we know) not in P, which in practice can still be solved relatively efficiently on a classical computer (a typical example for this is the SAT-problem). Problems like protein folding and go are most likely also not in P, but AlphaFold and AlphaGo have performed remarkably well on these problems.
AI systems are different to most classical algorithms in the sense that they are probabilistic and in general not guaranteed to find the correct answer, but it is still possible that they can find the correct answer efficiently for most cases (even when the problems are not in P). By definition problems in P require an algorithm which always returns a solution in polynomial time, which is not what AI systems do, so there is no reason why AI could not perform well on problems that are not in P.
While I doubt that AI systems will ever be able to fully simulate a quantum computer on a conventional computer, there are most likely problems that classical computers were not able to solve efficiently, but that become solvable efficiently enough in practice because of AI. If there are enough problems like this, then quantum computing becomes less and less attractive/useful, even though theoretically it is probably more capable than a classical computer.
I'm not a quantum computer nerd, but this "classical" workaround with AI only works on things we already have examples of, reducing them to higher order "first principles" (how chemical bonds works in folding for instance), which leaves a still juicy area of exploration: A quantum computer would be needed to explore from firster principles (quantum mechanics) the paths nature didn't take. Since nature is based on randomness, and building on what works with evolution, our current world may have circumvented amazing things in the space of possibilities.
You mean there IS some hidden variable in QM that an AI could figure out?!
If hallucinations count as 'figuring out'
@@whataboutthis10 Your hallucination wasn't that bad.
Exactly this. Classical cannot model quantum because classical will never violate the CHSH inequality. AI may make computers faster, but it will never make them quantum!
@@t23c56 Well, you don't need to be quantum to generate random result. AI could certainly compete with quantum computers on that field!
And does it matters if you call it hallucination or decoherence? What matters is that it gave this result way faster than classical computer!
🤣
@@igorstasenko9183Sorry can't really understand what you're trying to say here, but I'll try. Pure random number generation is not a use case of AI is that's what you're getting at. Perfect random number generation actually does require quantum mechanics, however in practice a complex enough classical system can easily produce random numbers. I remember seeing a wall of lava lamps somewhere with a camera pointed at it and they used that to generate random numbers.
Totally agree that the result is usually all that matters, but there are still chemical and physical applications which do require the actual process itself to have been quantum.
🤣
One key aspect of generative adversarial learning, as well as any form of semi-supervised learning, so for all of the most successful approaches in machine learning tso far, is that the task fundamentally needs to be somewhat easy to verify as correct or incorrect.
For GAN you need a very detailed dataset to begin with, and for any form of supervision you need a method to periodically check if the result is at all correct.
Reinforcement learning is just straight up not applicable as you need an efficient computational way to calculate the reward function, or how correct the solution is.
For chemistry and material sciences we know the properties of thousands of molecules and combound materials. But is that the case for quantum computations.
I guess we could use LHC data for training, but thats probably quite noisy.
Quantum computing: Simulates all solutions at once to find the correct/best one.
AI computing: Derivative guessing that "looks correct".
What could possibly go wrong...
It would have to mean that like with protein folding it's verifiable after the answer is found.
Once you have a better structure you can clearly see that it's better than your previous record.
In areas where 'improvement' is more important that 'proven to be the most optimal possible' these areas don't need quantum computing that gives perfect simulation.
It only needs answers that are significantly better than the answers we had before.
Lol
thats not how quantum computing works at all
@@gubunki That's the point. Everyone misunderstands current AI also.
One case I can see is getting actual data to train AI. Since, I guess we do not have much data for let us say folding of amino acids, quantum computer can be used to stimulate probable data point (starting from the physical rules we know and generating possible data points), which can be used to train AI
A classical system to model quantum computing uses 100 MW per hour as electricity; new probabilistic approach into computing in quantum field using noise as a good thing could surpass everything
Two things. One is that coming up with better algorithms is important, and a well-known fact in computer science. One learns that in any graduate level course in analysis of algorithms. A "good" algorithm can be orders of magnitude better than a "poor" one. The other is what you said about AI. It learns from the data. In other words, it is probabilistic rather than deterministic. Add to that an understanding of the rules and you basically trim the tree of possibilities, as you point out.
For example, I once wrote a program for my wife for a graduate course in mechanical engineering. Some of her classmates used a brute force approach on a large super-minicomputer (this was in the late 1980s). They actually ran out of virtual memory in a 32-bit computer. With a better algorithm taking into account the constraints of the problem, I got it to run on a 16-bit IBM PC with, wait for it, 2MB of memory. It ran in seconds.
The algorithm is king!
To me, this is the stuff AI is supposed to be used for--identifying patterns to help scientists do better science--not create pictures or write things. Note I don't say "to do science;" there will always be the need for humans to review whatever AI notices and points out. It is a tool to help people find underlying patterns, not put creatives out of jobs.
Quantum computing is not the problem of quantum chemistry. QC is required only for very detailed analysis of spectroscopy and reactivity. Ahmed Zewail showed that more than 20 years ago for laser beam reactions. The same abstract that you show in the video mentions density functional theory, a semi-empirical approach that has been quite proficient at predicting all kinds of spectra and patterns of reactivity. As for the physics of protein folding, the work was done by David Baker, who himself is one of the winners of the award you mentioned.
That just means Quantum computing isn't viable in the first place.
Exactly, which is why it's such a big claim!
so quantum computing will be used only for encryption and game engines?
Two very crude remarks (what else?)
1) The fact that Nobel prices in Physics and Chemistry are now attributed to computer scientists may mean that either Science is dead, or a computer science Nobel price should be created.
2) This question on IA seems to be a nice pretext to give up quantum computing without admitting that it will never work since, putting it simply, superimposed states are a theoretical artifact and do not really exist.
Hi Sabine,
I am a big fan of yours and I value your opinions and insights very much. I recently came across an interview with a physicist/mathematician/philosopher from Harvard (Jacob Barandes) who claims to have come up with a new set of postulates for QM that is based on stochastic processes (that does not involve wavefunctions, etc.). RUclips video title: There’s No Wave Function? | Jacob Barandes.
By the sounds of it, his new set of postulates supposedly re-creates the usual QM postulates, etc. However, his new postulates takes the "mysticism" out of QM (like wavefunction collapse, measurement ambiguity, etc.).
I'm a little skeptical about his claims. I have a feeling that his new postulates simply swaps one set of mysticism for another (while still retaining the same kind of ambiguities and weirdness).
It would really mean A LOT to me if you were able to take a look into this and let us know what you think. Please and thank you.
I watched it too, nothing fundamental new in it
In the whimsical land of Far Far Away, Shrek and Donkey discovered a connection that transcended friendship. Amidst their adventures, a deeper bond formed, defying societal norms. Their unconventional love story unfolded quietly, a tale of acceptance and understanding. Far from the conventional fairy tales, Shrek and Donkey navigated their feelings in a world that had yet to grasp the diversity of love. In the end, it wasn't the castle or the dragon that defined their happiness, but the genuine connection they found in each other, proving that love knows no boundaries, even in a swampy fairy tale realm.
One AI program a guy used to develop a wild alien looking rocket motor. It was designed for maximum laminar flow and efficiency. Mixing of air and fuel like a carburetor. Copper based alloy seems to withstand the heat well. Thank you.
I strongly believe that Turing machines can do what quantum computers can do but it will take far more clock cycles.
I mean yeah duh, that's theoretically already known. Many more clock cycles is often the barrier to practicality
Dear Sabine, It would be really exciting to see a video about Ising machines or annealers as a competitors as a specific analog approach.
AI/Machine learning also deals with (discrete) probability distributions, I'm curious if there will be an opportunity in the future for AI to be applied to quantum computing in some way to reduce the probabilistic search space that quantum computers deal with
I've been looking recently into quantum-inspired machine learning, tensor networks, MPS etc. and it really seems that the fact that quantum systems "present to us" a highly-dimensional, tensor interface doesn't mean they are highly-dimensional "all the way". Internally they can be equivalent just to a bunch of matrix multiplications.
Most quantum algorithms currently use quantum+ classical optimizer like in VQE. AI like AlphaFold requires substantial computing power utilizing HPCs with multiple GPUs and TPUs in large scale data centres. So overall computational and energy costs could be large .it might be cheaper for startups to run quantum algorithms than AI on HPCs
Many times, I have posted on RUclips that there are two different quantum states in the atom: the real and the hidden quantum states, also known as hidden variables. I even posted that it is impossible to understand quantum mechanics if classical mechanics is totally dismissed, because the hidden quantum state is ruled by the classical mechanics. Physicists commonly mess up the properties of quantum mechanics with that of classical mechanics or vice versa. But this time, I would like to share how I understand the inferences behind the atomic mystery in different way. Every complicated atom having two or more electrons behaves like an inductance-capacitance oscillator or LC-oscillator for short. Quantum mechanics can explain the hardware part of the LC-oscillator, but classical mechanics can explain the software part of the same, which is the energy content of the same, that can be changed continuously. So, if we hold the values of the inductance and capacitance of the said oscillator constant, the frequency of the LC-oscillator remains the same no matter how much energy we put in into the same. Going back to atom, similarly even if the electron in question doesn't jumps, the hidden quantum state remains vulnerable and susceptible to the influence of the outside world, which means that the energy content of the hidden variables can still be changed by the outside observer without quantum jump involved in the process. This is the same phenomenon that brought so much trouble to Erwin Schrodinger that eventually led physicist to introduce a scenario also known as the "Cat in the Box" that can be both dead and alive at the same time that remains a paradox up to this point of time. But based on the inferences, actually there are many cats, but one of them is a pregnant cat, the rest are just unborn kittens. The pregnant cat can be described by quantum mechanics; and the unborn young cats can be described by classical mechanics. However, the very act of opening the box forces the pregnant cat giving birth to a kitten. This phenomenon is analogous to breaking the circuit of the said LC-oscillator where the breaking point generates a spark that signifies the kitten, the young cat. So, all these explanations suggest that the existence of hidden variables or hidden quantum state or classical changes is mainly responsible for holding all the information we need to know about some of the things we already know and understand like the binding energy between two or more atoms combined together to form a molecule, the spread of spectral lines, superposition, entanglement including chemistry and biology, just to name a few.
About a decade ago I read a paper about how a team was able to use insights from quantum systems to develop an algorithm for classical computers that beat even the quantum alternatives. Since then I have been wondering what else we could learn from quantum systems that could make classical ones better.
Did I say something recently about chemistry? It is worth studying as its own subject - also quantum chemistry which models only electrons (Born-Oppenheimer approximation) as electrostatic potential clouds consisting of occupied molecular/atomic orbitals (only using probability density as any QM system).
Energy is key. QC are supposed to perform computing using Q principles directly, which means efficiently in terms of energy. That is especially true to compute Q things using QC. Try that with conventional machines eventually running AI. Don't forget your wattmeter.
you need more silicon than there are atoms in the observable universe to accurately simulate 50 logical qubits on a classical computer, many use cases for quantum computers need in excess of that, no matter how smart an AI is, if its built out of silicon semi conductors, it is physically impossible run shors algorithm on large numbers. The funny thing about Quantum computers is that there are quantum algorithms which can improve LLMs, so its not like the two have to be mutually exclusive.
As an aside regarding classical modelling of quantum systems, it is important to remember that nature herself has no agency and doesn’t perform maths or make decisions etc, it just does. So unless we are living inside a simulation (which can have agency), nature’s interactions from quantum to cosmological are seamless. This is the fundamental error in the wave-particle duality hypothesis - nature doesn’t know of either scenario - it just is and does irrespective of our own context. Classical versus quantum is our dichotomy but it certainly cannot be natures.
It will be interesting to see which will be used for what and how they might be melded. thanks doc
We mathematicians have been using this phenomenon for some decades now. We kinda don't care whether P = NP anymore because in real life systems follow sets of rules and you can exploit those rules to manufacture a heuristic that will give a solution in nearly polinomial time. Neural Networks (read "AI") are merely one of those heuristics.
I’ve been intrigued with the idea of melding quantum computers, or I suppose very advanced and efficient AI, and evolution design wherein you specify the parameters of a final design - what you want it to do - and within the physical constraints of what is being modelled the system sort of brute forces many designs that satisfy the desired parameters. Some type of evaluation is done on each output, prototype testing in a virtual environment for example and the worst designs are rejected. Over time the number of designs narrows as they evolve to better satisfy the desired constraints. I guess the idea is that if you have some way of efficiently calculating a very large number of options and evaluating them the best designs will “float to the top”. Obviously like with Alphafold you only need to compute things that are within what’s possible so maybe quantum methods aren’t needed just super efficient ways of modelling
AI absolutely can act as a means to focus searches to enable good results to be found using traditional means, in the way mentioned for chemistry. BUT ALSO, AI approaches may provide the necessary traction to allow quantum computing to become properly realized. I wouldn't bet on this being an either/or situation.
Of course they can. Intelligence is essentially the ability to make predictions based on memory, which involves drawing from an abstract model of reality. While a computer operates using literal zeros and ones at its core, the information it processes exists on a higher, abstract level. This abstract realm is what we associate with thought-it’s the relationships and patterns between the zeros and ones that give them meaning. These patterns form the context and structure, shaped by their sequencing in time and their relationship to the model they represent. In this way, classical computers can indeed model anything a quantum computer can, because the abstraction transcends the physical differences in how they operate.
The point in Quantum computing, is not only that the randomness is innate. That you do not need to use some pseudo-random number generator code. Yes, it is one advantage, but the main advantage in Quantum computing is that you can form a algorithm to solve a problem like a multi exit obstacle track for electrons and then measure how many electrons come out of each exit. Thus you can solve every problem which can be translated to such a track immediately as if it was run in parallel on many cores in a classical computer. So yes, a classical computer can emulate it - but in a classical computer you will still need many cores running in parallel to do it.. That's the whole point. The question of why this is important demands an example - well, one glaring example is code breaking. In code breaking you frequently need to find rules that will diminish the option space to be more manageable, and AI can help here - for sure. But in many cases, the remaining option space is still huge. And so, with only classical computing to break a code, you will still need hundreds of years, even with AI. Enter quantum computing - quantum computing can enable one to produce an 'obstacle course' such that in one case the code is getting one step closer to being decoded without the need to employ a parallel fleet of GPUs working for years on it. Suddenly the problem becomes trivial. So yes, in this specific example, AI may still help (as in most if not all examples..). But AI cannot solve code breaking alone. Here quantum computing is of the essence. And there are many other problems in which QC may be necessary. However, yes, many of these problems are marginal, in specific fields, like code breaking.. I reluctantly have to admit that much of the fan-fair for QC was about things that AI had recently demonstrated it can handle perfectly well on classical computers (such as folding of proteins). It doesn't mean "AI Could Make Quantum Computing Obsolete". It could mean AI could make Quantum Computing into a marginal field of research specific to special cases such as code breaking where it cannot be done without it, but otherwise it would be redundant.
Solving the Schrödinger or Dirac equation for quantum chemical computations numerically is essentially non polynomial in system size and accuracy. If you are out for analyzing subtle chemical patterns of structure and reactivity you feel this limit even today. There will be a certain class of chemical problems remaing unsolved by classical methods, especially when a deep understanding is aimed. Not sure if quantum computing still couldn't be the answer to that.
This is so old school, Love, Greg I would just emulate it after you you wore its hat. And this works perfectly
Weights in ANNs might represent kinetic energies of quantum particles.
Hi @sabine the graph is wrong, the time hould be on the X axis, and Operations on Y axis,but very nice video about AI and Quantum Computing
I shared this with a friend who works in exactly this area - and these were her comments:
TL;DR: Sabine mischaracterizes the problem
----
She brings up some points which are kind of right but arrives there through wrong interpretations.
My entire job is studying the crossover point between quantum and classical for chemistry. In general yes, if classical simulations continue to improve such that the crossover point between classical and quantum is never practical, then quantum computing will not be useful.
Most of her arguments hinge on the ”space of possibilities” being why quantum computers might be advantageous, and that is incorrect, so a lot of her arguments (like comparison to Alphafold) fall flat. The main challenge for chemistry isn’t the expressibility of a quantum or classical model, but how efficiently you can navigate through that possibility space.
I didn’t see that talk by Demis, but from the clip shown it sounds like he is referring to complexity class assumptions that BPP (classical) is contained by BQP (quantum), or NP being contained by QMA. This definitely would be deep and crazy if that hierarchy were inverted or it collapsed.
In general we look for quantum advantage in specific problems where we believe it is NP hard classically but BQP quantumly. This won't be true for many chemistry problems, so we target specific chemical systems which are known to be difficult classically, such as strongly correlated systems. Its like how you would never target quantum advantage to do addition, because that is super easy classically. Instead, factoring into primes is hard classically but has been proven to be in BQP using Shor's algorithm. So In that case it is interesting to study the crossover point between quantum/classical. So what she said about quantum not being useful for all of chemistry is right, but that was never the goal.
It looks like the excerpt with Carleo Guiseppe quotes was leading to the question with wether training a foundation model on systems that are easy to simulate classically can generalize to more complex systems. This is an active research area many people are optimistic about (I'm also involved in this), but nothing big is out yet.
I don't know much about Sabine but this made me not trust her videos on content I don't know about 😅
This is the risk of experts addressing topics outside their own field.
Current computers are still very sequential in what they do but the trend is towards making them more parallel. This trend in some ways makes an additional threat to quantum computing. In parallel systems stopping to decide what to do next is a burden but if you can design the system such that the need to decide if rare, then it tends to consider more cases in parallel much like quantum stuff considers many cases at the same time.
What about the Maximum Goodness Theorem? A countable finite has a maximum.
Sabine, I think you are confusing two different problems here. The most of experts thought that computer or AI could never master a complex game such as Go. The rationale was that the total possible move that can be made in Go is simply mindboggling (10^180) and calculating that many possible moves would be impossible. This rationale was flawed because we humans don't calculate all the possible move combinations either. Rather we learn and adopt the game based on the patterns and pattern relations. What Denis Hassabis is talking about is related to this. Even generative AI works on this principle as well (attention mechanism and transformer architecture for example.)
However, there is a more fundamental problem that needs to be solved. As Richard Feyman said in his 1981 speech, there is a fundamental mismatch of a deterministic machine trying to simulate a probabilistic nature. And he is right. The current AIs are called the Boltzman machines for a reason and this mismatch will only get worse as AI advances. For one, there is an enormous resource and energy requirement need to do the AI work and this has to do with this mismatch. I am not familiar with the works in Quantum computing but you can be certain that probabilistic machines will rise sooner or later. If it is not Q-bits, there will be P-bits, O-bits, or whatever but they will come since the current mismatch will stop the rise of AGI.
I think, that instead of fumbling around pure states of qubits, more promising branch of quantum computing is ising model optimisation.
Which doesn't really give so much damn if something lost coherence. Only most stable state will persist anyway.
And could specifically solve hard problems of optimisation, which are necessary for AI.
Also, quantum computers could very successfully screw about lattice problems. Yes, there you need quite more stable qubits. And a lot of them.
But lattices are quite a powerful things to screw with a lot of cases of logical tasks. For example, it's trivial for LMB solver to find decomposition of binary string into xor of several periodic signals.
Interesting. Thanks for talking about this, Sabine. I'd like to hear IBM's and PSI Quantum's take on this, too, because they're investing a lot of $ in developing quantum computing, although both with different roadmaps. Perhaps combining AI and quantum computing could lead to more powerful computations?
The only serious application of quantum computing, arguably the one it's being developed for is prime factorization - being able to do it quickly would break (some of) asymmetric encryption schemes used on the web and this would allow to decrypt past online communication (as long as it passed through nodes you controlled and you stored it in your data center).
I don't see AI helping with that unless it shows that prime factorization can be done in polynomial time (of reasonably low degree) or that P = NP.
Sabine, you should do a piece on Module Learning With Error and other PQC algos.
As an expert in discrete optimization, I'd like to mention so many things here!
A while ago: tenant models with selective goal ranges can approximate quantum systems- this still has a nesting problem with float errors that cannot be resolved. This is baseline approximation error.
Problems*
Problems*
like alphafold once the inflection point of translating atomic/molecular structure into human scale physical properties at any given temperature/pressure is reached...
Sabine makes science sensible.
I would love for you to weigh in to Jacob Barande's non-markovian non-divisible classic stochastic approach to quantum mechanics.
Please make it digestible to common folk 😊
2:53 your lines should not be starting at the same point. "Conventional" computing just sped up existing logic that was accomplished using electromechanicals before the vacuum tube, then miniaturized it with the integrated circuit decades later. Quantum today is like computing before the IC - quantum still needs its great scaling breakthrough allowing logic to be constructed as true VLSI.
Current quantum processors really looks like where CORE memory was in the 1960's - the concept for how to use it at scale seems to be getting worked out, but the very edge of what can reasonably be manufactured is ~1k bits (or qbits).
I think you find that heuristics and classical computing are good enough to predict qualitative folding structure but not good enough to predict shapes, energy levels or binding strengths to the quantitative accuracy needed for drug development, biosystem or large-scale nanosystem design without testing.
Outstanding - eloquent and clear
This good science for me, because there are archievable goals to get. With KI based material-science maybe we find something very useful, like supraconduct on roomtemp on day. With this then, we can go further and so on. This is more important to focus on than any other "old unsolved" problem.
Thank you, Sabine.