#143 Scott Aaronson: Revealing the Truth About Quantum Computing

34:00So let's say I have two qubits in the state zero, and now I Hadamard one of them. And now, so now one of them is in the state zero plus one. and now I do a controlled knot from that qubit to the second qubit. So now what I'm actually going to have is, well, in the zero branch, it's still just zero, zero. But in the one branch, I flip the second qubit. So now I've got zero, zero plus one, one, which is an entangled state, right? So, you know, by just putting together these gates in an appropriate pattern, I can create entangled states, you know, even very complicated entangled states, on hundreds or thousands of cubits.

34:41What I described to you so far is not quite a universal set, but you can throw in just about anything else. Like let's say a gate that rotates a cubit by some angle, like pi divided by eight or 22 and a half degrees, something like that. And then you get what we would call a universal set of quantum gates. And what universal means in this context is just that if you apply enough gates from that set, then you can affect any unitary transformation on any number of qubits to any desired precision. Okay. So it turns out that with just a small discrete set of one and two qubit gates that you already pass the threshold of universality.

35:40And that was something that people discovered more than 30 years ago. OK, and so now, you know, you could say that, you know, in quantum algorithms design, you know, what you're always doing is you're thinking about here is my collection of qubits. And now here are my allowed gates. And now which gates do I apply, you know, to which qubits when, you know, and we have a notation for for this called quantum circuit notation. You know, it looks almost like a musical score, you know, when you see it written. It's like each qubit is a line going from left to right, you know, and then the gates are like different musical notes that we write down, you know, over those lines.

36:19Okay. And then that, like, you know, you could say, you know, once they've written the quantum circuit, then that's when the quantum algorithm designer's job, you know, or the quantum programmer's job is done. And now this is when the quantum hardware engineer's job begins. Right. Because now what they need to do is take the sequence of gates and translate it into actual pulses that will, you know, act on these physical qubits, you know, in a way that has an effect of doing these gates. OK, now the good news here, you know, is that, you know, when we when we think about what what our qubits actually are in different quantum computing architectures.

37:04Well, like in the trapped ion approach, you know, each of our qubits is a usually it's a ytterbium nucleus. OK, so we have this like row of ytterbium nuclei and each one. And usually it encodes the qubit in its spin state. So like if it's spinning counterclockwise about some axis, then that represents zero. And if it's spinning clockwise about that axis, then that represents one. And it could also be an arbitrary superposition of the two. Okay, so now we have this row of nuclei suspended in a magnetic field. And now what people actually do in the trapped ion approach, you know, and this is not speculation anymore.

37:53This is all actual experiments that they're, you know, that they actually do now. They can pick up the individual nuclei and move them around using lasers. They just sort of grab onto a nucleus with a laser. They move it around. And now when two nuclei get close to each other, right, they have a natural electromagnetic interaction, right, that causes some, you know, two qubit unitary transformation, right? And now, you know, the hardware person's job is just to shape that, you know, natural unitary into something that looks like a CNOT gate, or, you know, whatever was the gate that the quantum algorithms, you know, designer asked for, right?

38:38Right. But, you know, but but but but but the good news is, like I said, you know, there are many, many different choices of one and two qubit gates that will be universal. Right. So so they'll like like, you know, they'll they'll they'll all be more or less as good as any other. Right. So. So so now, you know, what what they're doing is they're shaping a pulse sequence. So they're programming, you know, the, well, a classical computer, which is controlling a microcontroller, which is controlling, you know, the lasers, okay, to move around, you know, the different ions in a, you know, particular choreographed pattern, you know, that will have the effect of implementing this quantum circuit on the qubits, right?

39:27and then at the end there's a different thing that you have to do to measure the qubits you asked about measurements but before we get to the measurements so when you talk about gates I understand classical computing transistors create gates that are open or closed and that's the zero and the one and you build that up So in taking these nuclei, for example, manipulated by, what is the gate? I mean, is it the laser that's? I mean, the gate is a mathematical abstraction. A gate is the unitary transformation that is to be affected on this pair of cubits. The laser is a way of realizing that gate.

40:23That's what I mean. Okay. And then measurement. Yeah, measurement basically means you have to take your spin state and now it has to get recorded in something macroscopic, right? And typically, the way that measurement devices work, I mean, we have technology like electron microscopes that are able to see things at tiny scales, right? But typically there is a kind of amplification effect that goes on, right? Where like, you know, you'll have a single, you know, photon that let's say, you know, deflects in a different way depending on whether, you know, this nucleus is spinning clockwise or counterclockwise, right?

41:14And then that photon could trigger a cascade of additional photons, right? Which, you know, then ultimately you have a whole photo detector array and it becomes something that's macroscopically detectable, right? So, but, you know, you have to be very careful to only make the measurements when you're ready to, right? And in some sense, you know, a large part of the engineering problem of building a quantum computer is to prevent the environment from measuring your qubits before they are ready to be measured. Yeah. Right now, how advanced, how many qubits have they gotten to that you're able to write and read on?

41:59So, okay. So you can't just ask about how many qubits, right? You have to ask, what can you do with these qubits? How good are they? Because D-Wave will tell you, oh, yeah, we have thousands of qubits, as many thousands as you want. But then they're just trusting that people are not going to be asking, okay, but can I actually beat a classical computer with them? Or at any rate, that was the D-Wave of a decade ago. You know, it's, you know, I've heard that they've become more cautious since then. OK, but, you know, for me, you know, the single most important number to look at is actually the two qubit gate fidelity.

42:48OK, which means like with what degree of accuracy can I apply a single two qubit, you know, entangling gate of my choice? OK. And, you know, when I entered the field, you know, 20, 25 years ago, right, it would have been spectacular if you could do a single two qubit gate with 50 percent fidelity. Right. And, you know, and then, you know, maybe a decade ago, it became 90 percent fidelity and then it became, you know, 95 percent. And then, you know, with Google's quantum supremacy experiment in 2019, right, they had like a thousand gates roughly applied to 53 qubits, you know, and each one with 99 .5 percent fidelity.

43:35OK, now the more the most recent numbers from the past year that I'm hearing from, you know, both the superconducting and the trapped ion people is like ninety nine point eight percent or ninety nine point nine percent fidelity. And this is typically in systems with 50 or 60 qubits. And some people have scaled up even to, I think IBM has scaled up to hundreds of qubits, but not with fidelity that's quite that good. Right. And now just to help orient people, you know, where we are. Right. Like the key discovery in the mid 90s, you know, that made people feel like, OK, you know, building a quantum computer is merely a staggeringly hard engineering problem.

44:32Right. And it doesn't require any new fundamental physics. It was this discovery of quantum error correction. Right. And what quantum error correction effectively said was that in order to build a reliable quantum computer with as many qubits as you want and as many gates as you want, you know, you don't have to get the error all the way down to zero. Right. Like that. That was the fear. Right. Because if you had to get error all the way down to zero, that's just never going to happen. Right. It's just unrealistic. OK. But what people discovered in the 90s was that, no, you only have to get the error down to some very, very low, you know, non zero, but you know, but non zero level.

45:15And then there are these very clever quantum generalizations of error correcting codes that can take care of the rest for you and that can push your effective or encoded error rate all the way down to zero. OK, so, you know, so at that the original estimate from 1996 was, I think, that you needed like 99.9999 percent fidelity of two-cubit gates. So basically, like you could you could tolerate a one in a million chance of error. OK, you know, and that just seemed ridiculously far from from where anyone was technologically. Right. But it at least proved the principle that, you know, that there like there is only a finite amount of engineering that has to be done.

46:03Right. And not an infinite amount. And and, you know, now, you know, I tell you, you know, with the with the more recent error correcting codes that people have developed, it looks like you could build a scalable quantum computer with, you know, if you have two qubit gates with ninety nine point nine nine percent fidelity. Okay, so four nines, let's say. Okay, and if you remember, you know, where I told you, you know, we are right now, you know, I think the experimentalists are now closing in on 99.9 % fidelity. Okay, on three nines, right? So, you know, if you just plot that on a graph, like it looks like it's pretty on track, right?

46:45It looks like, yeah, you know, if this effort continues, then, you know, we should eventually, you know, get to the threshold where, you know, after which error correction becomes a net win. And now once you've crossed that threshold, then how many qubits you want is just a question of cost. Right. It's just just like with a classical computer. Right. It's you know, it's like, OK, if you want, you know, you know, two million qubits instead of one million, then, you know, then that's more engineering that you have to do. But, you know, you have figured out how to keep your qubits alive for for as long as you need them to be.

47:26And then there's no fundamental limitation on how many more you can add. OK, so that's so that that's the real goal to get to this fault tolerance threshold. right and you know the things that people can do now with 50 or 60 qubits you know they're they're very cool as demonstrations you know and they're important as proofs of principle right but none of this yet is has been error corrected right and and and and and that's the real goal right right and and none of it is i mean certainly the quantum supremacy experiment but none of it is beating a classical computer. Yeah, well, I mean, the quantum supremacy experiment, you know, arguably, you know, sort of is, right?

48:11You know, not for a useful problem, you know, but at least for some contrived benchmark, you know, and, you know, that came out of work that my student and I did in 2011, where we said, you know, hey, you know, it's not, You know, people are talking about like, what can you do that's useful with a non-error corrected quantum computer? But, you know, maybe we should just first try to pin down, can you do anything at all that is classically hard, right? Even if it's not useful. And we came up with a proposal for how you might do that, which was called boson sampling, which was adapted to optical quantum computing.

48:54Right. And then, you know, the optics people got very interested and started doing very small scale experiments with, you know, five or six photons. Right. At that point, it's still trivial for a classical computer to simulate. But then then what happened was that Google in 2014, I think, hired John Martinez, who was one of the leading superconducting experimentalists. And Martinez actually said, like, let's go for this. Let's just do a quantum supremacy demonstration, you know, with with with, you know, I think, you know, they were hoping for 60 or 70 qubits. Okay. And, and, you know, this wouldn't exactly be boson sampling because, you know, their, their, their, their, their hardware was not designed for it.

49:45But they said, you know, let's do something like boson sampling, but, but adapted to our hardware. So, you know, so we talked to them at the time and we helped, helped them figure out what that experiment would be, you know, and then we sort of adapted the theory of boson sampling to the kind of hardware that they had. And then in 2019, they actually announced results that with 53 qubits and about 20 layers of gates, about 1 ,000 gates in all, they solved this sampling task. At the time, they estimated that maybe it would take 10 ,000 years for a classical computer to do the same thing. And then that number got quoted all over the press.

50:27Unfortunately, you have to be really careful with these numbers, right? Because then what happened was that the classical computer scientists came along and said, no, we can actually spoof this a lot faster with a classical computer. And now they can, I would say, if you're willing to spend enough money to just have enough parallel classical processors, then you can actually simulate the Google experiment faster than the experiment itself. Right. So it all just comes down to like energy and money. Right. You know, and by those metrics, I would say, you know, on these quantum supremacy benchmarks, the quantum computers are still winning, but only by, you know, a couple of orders of magnitude.

51:15but not by a whole lot. So there's an urgent need for better quantum supremacy experiments that will reestablish a clear speed up. I think, you know, given the way that the technology has advanced just within the last four years or so, you know, that should definitely be possible. But, you know, in the meantime, a lot of the companies have said, you know, they don't really care about quantum supremacy anymore. They just want to go straight for error correction. Right. Right. And so error correction is kind of the... That's the holy grail. The foundation or the frontier across which we'll be into practical quantum computing.

51:54Yeah. I mean, again, some people hope that we will get lucky and be able to do practical things even before fault tolerance. Okay, there's a term people use for this, NISC, coined by John Preskell, noisy intermediate scale quantum computing. You know, I mean, by analogy, right, before the transistor was invented, you know, we had only much noisier building blocks for classical computers, like vacuum tubes or electromechanical relays. But, you know, those were still they still worked well enough for people to do various useful things with them, among which was was winning World War Two. Right. So, you know, so you could hope that maybe the same will happen here, that even before we get truly error corrected qubits, you know, we'll have qubits that last just long enough that we can eke out some advantage.

52:52And I think, you know, if we're going to do that, then the best hope by far is that we'll do that for some sorts of quantum simulation problems, such as simulating the properties of various materials. Okay. And, and, and, and, and there's a much better hope of doing something that is scientifically interesting, you know, and, and, and new and classically hard than there is also doing something that's commercially useful, which is kind of a higher bar to clear. And people are working on that. But the point is, there is no guarantee. Nature will have to be kind to us for us to get clear quantum computing speed ups before error correction.

53:42After error correction, then we're much, much more confident that we could get such speed ups. Yeah. Well, we're almost up to an hour. Yeah. But but I do. Can you just give a synopsis of your safety work at OpenAI? You did a very good job on the collective intelligence webinar. All right. Well, so OpenAI approached me a year and a half ago because, you know, several people there were fans of my blog. I think. And they said, you know, would you take a leave of absence and, you know, work with, you know, work for us on what can computational complexity theory do for AI safety? You know, and my first reaction was, you know, why do you want me?

54:29I'm a quantum computing person. Right. You know, we know what, you know, I mean, I've been, you know, and already at the time I was, you know, I had seen GPT-3, you know, I was bowled over by it. I knew that it was this unbelievable advance from the previous state of the art in chatbots. But I was like, first of all, the progress in machine learning has been driven almost entirely empirically. For the most part, it hasn't depended on theory. It's depended on having lots of data and lots of computing power. Right. And, you know, and then and then tweaking the algorithms until you find the ones that just empirically turn out to work well.

55:17Right. And, you know, no one no one really understands, you know, at a at a at a deep mathematical level why any why any of this works. Right. They just you know, they just train the models and then they try them out. And, you know, and it turns out that when when they're big enough, they do work. Right. So, you know, and then and then furthermore, you know, none of it has anything to do with quantum computing, you know, except that, OK, you know, machine learning and quantum computing both involve extremely high dimensional vectors. You know, so, you know, but but but but but beyond that, you know, they're they're just they're just very different fields.

55:58But, you know, they they made a case to me that, well, no, you know, like like the problem going forward is not just going to be how do we get this to work, but how do we ensure that it's safe? Right. And that does seem like a problem where where we have to go back to first principles, where we have to think about it theoretically. You know, and they gave examples of, you know, some of the big results that we know in computational complexity about how like a weak verifier can check the behavior of a super powerful but untrustworthy prover. And they said, well, you know, that's basically the situation that we're in.

56:36Right. You know, the super powerful prover is the AI and, you know, the limited verifier is humanity. Right. and maybe you can port these results over. And I said, well, it sounds like at any rate I'd be able to find something interesting to work on. But so I'm teaching like quantum computing courses. I'm supervising my students. I have a whole, so maybe some future year I get involved. And they said, well, trust us. this is going to be a really big year for AI. You want to get involved this year. I said, you know, and, and, but, you know, they were nice enough to let me do it from, from Austin where my family is and just, you know, travel to San Francisco periodically.

57:26So, you know, they made it very hard to turn down. And so I, so, so I did, I started working for them and, you know, I found, you know, so it's now been a little over a year since I started. and, you know, I found various projects to work on. Maybe the easiest one to explain is a water market. Okay. So, you know, like I've not been able to, you know, predict where all of this is going, you know, even like, you know, decades in the future as like people constantly ask me to do now. But I'm proud that at least I was able to see about three months into the future, right? So before ChatGPT was released, I just had this moment of terror when I thought, oh, my God, every student in the world is going to be using this to cheat on their homework, aren't they?

58:21Or at least is going to be tempted to. Right. And, you know, and every troll is going to, you know, have this amazing tool for spamming every forum on the Internet. Right. Or impersonating people or, you know, you know, what about propaganda campaigns or fraud? And for all these different categories of misuse, they would all be much harder if only we had a way to detect which text was generated by large language models and which was not. So then I realized that this provenance or attribution problem was going to become really central. Right. And, you know, there are a bunch of different approaches that you could imagine to it.

59:13But one one particular approach that I've worked on is called water marking. OK, and this is where we slightly change the way that our language model, you know, such as GPT operates. So language models are inherently probabilistic. Right. You know, they're always like, you know, you have this transformer, you know, neural network. That's always taking as input, you know, the context of like the previous however many words. And then it's generating a probability distribution over the next word. Right. And, you know, now now like in normal operation, all that would happen next is that we sample according to that distribution.

59:58Right. Sometimes the distribution will be very, very centered on one possibility. Like if I say the ball rolls down the, you know, then GPT is nearly certain that the next word is hill, right? But, you know, in other cases, it's balanced between several possibilities, right? Or, you know, even dozens of them. Okay, so with watermarking, what we do is we choose the next word, the next token, you know, in a way that is secretly deterministic, okay, but which looks random, right? It looks like we're doing it according to the distribution that GPT said to sample from. You know, casual user can't tell the difference at all.

1:00:44But we're actually choosing the next word deterministically in a way that is biasing a score that we can calculate later, given only the text. Right. So we're sort of we're doing it in a way that systematically favors certain combinations of words over others. right, you know, combinations that would be like otherwise arbitrary and, you know, meaningless. Okay. And then, you know, if later some teacher gets a term paper and they suspect that it was written by, you know, using chat GPT, they would be able to submit it to a detection tool. And that detection tool would calculate that score, you know, using the key of the pseudorandom function, you know, which it would have.

1:01:30And, and then, you know, if that score is like really, really unusually large, then you would be, you know, depending on how many, on how long the document was, you could be statistically almost certain that yes, GPT must have been involved in writing this. Right. And would that, would that work if only portions of the document were? Yeah, that's an excellent question. So I do have an algorithm also to do that, to take a long document, portions of which were, you know, we believe were written by GPT, and then identify which were the most likely regions to have been GPT generated. Okay, now, none of these approaches are foolproof.

1:02:14Okay, you know, one can think of attacks, you know, that would get around them, some of which are, you know, like, imagine a student who asks GPT to write their term paper, but in French, and then they put it into Google Translate, right? That would remove any watermark that we currently know how to add. right and so you know how do you watermark at the level of the ideas you know at the semantic level like at a way in a way that would survive you know all these sorts of you know rephrasing or sort of you know you know mere changes to the to you know local changes to uh syntax and things like that i think that's a very very hard question but that's that's fascinating uh and and you're optimistic about the...

1:02:59Well, I mean, I'm optimistic that we can deploy things, you know, that will, you know, that will help at least somewhat, and especially that we'll be able to learn more by deploying them, right? You know, I'm optimistic that AI safety, you know, after decades of like armchair speculation about it is finally an empirical subject. And, you know, I think it It might be very difficult to stay one step ahead of all of the misuses that people come up with. Right. It might be a cat and mouse game. But, you know, in other cases, you know, people have just bitten the bullet. Right. I mean, I mean, you know, we we have a giant thriving software industry, you know, despite, you know, the ease of pirating software.

1:03:49Right. We have, you know, search engines like Google, you know, despite all of the attempts to to game their results. Right. So, you know, so it might just be that, you know, it will take a large effort to stay ahead of the misuses. But, you know, at least in the in the near to medium term. Yeah, I'm reasonably optimistic that we can make progress. We can break down the problems into sub problems and tackle them. Now, once you have AI that is just better than humans at everything, right? Or, you know, let's say AI that, you know, is to humans as we are to orangutans, right? Then, you know, I don't know what kind of world that leads to, right?

1:04:36I don't know how to ensure that that world is safe. But, you know, my defense is I don't think anyone else knows either. Yeah, yeah. And I have one last question. All right. I've been talking to people about the pros and cons of open sourcing generative AI. Yes. And there is an argument that it's great because, you know, it expands the use cases and you get this network effect of people improving. And then the other side who says, well, really, it's such an asset, intensive game that open source really doesn't have the resources to compete. But what I'm interested in is even at the lower with smaller models, that once they're open source, I mean, things like this watermarking is not going to, that will work on a proprietary model.

1:05:44But once it's open source, it's anybody's game. Is that right? No, you're absolutely right. And that's true, not just for watermarking, but for pretty much any safety mitigation that you could possibly think of. Right. Once the model is open source, then people can do what they want with it. And, you know, and empirically, like, you know, you know, if you if you fine tune your model by like this RLHF reinforcement learning with human feedback, you know, to be, you know, to only give, you know, safe, you know, helpful, inoffensive answers. It takes about two days for people to take an open source model and remove that fine tuning.

1:06:24You know, the model will spell, you know, whatever racist, invective or bomb making instructions you want it to. Right. And so so that so that that that's the trade off with open sourcing models. Right. And so I think it fundamentally depends on how powerful the model is. Right. Stuff that's at the level of GPT-2, you know, has already been open sourced. Right. You know, and stuff at the level of GPT-3, you know, probably, you know, you know, is in the process. Right. Or, you know, will, you know, either is or soon will be. Right. GPT-4, you know, you know, I feel that, you know, OpenAI made a very defensible decision to not open source that.

1:07:05Right. Because there is probably some mayhem that people could already do with it if they had access to the weights. I mean, even just aside from the obvious commercial reasons why OpenAI doesn't want to open source it, I do think there's also a safety issue there. And eventually, you may get AIs that are powerful enough that open sourcing them is kind of like open sourcing thermonuclear weapons. right? It's, you know, it's just a thing that you don't want to do. This episode is sponsored by Crusoe Cloud. High-performance cloud computing or low environmental impact, Crusoe Cloud was built because the innovations of the future need both.

1:07:51Crusoe Cloud is a scalable, clean, high-performance cloud optimized for AI and HPC workloads and powered by wasted, stranded, or clean energy. Crusoe offers virtualized compute and storage solutions for a range of applications, including generative AI, computational biology, and rendering. Visit crusoecloud.com, that's C-R-U-S-O-E-C-L-O-U-D dot C-O-M, to see what climate-aligned computing can do for your business. That's it for this week's episode. I want to thank Scott for his time. If you want to read a transcript of this conversation, you can find one on our website, IonAI. That's E-Y-E hyphen O-N dot A-I.

1:08:47And as always, the singularity may not be near, But AI is about to change your world, so pay attention.