Quantumania! (Ep. 240)
Navigating a quantum future with Scott Aaronson
My guest today is Scott Aaronson, a theoretical computer scientist, OG blogger, and quantum computing maestro.
Scott has so many achievements and credentials that listing them here would take longer than recording the episode. Here's a select few:
Self-taught programmer at age 11, Cornell computer science student at 15, PhD recipient by 22!
Schlumberger Centennial Chair of Computer Science at The University of Texas at Austin.
Director of UT Austin's Quantum Information Center.
Former visiting researcher on OpenAI's alignment team (2022-2024).
Awarded the ACM prize in computing in 2020 and the Tomassoni-Chisesi Prize in Physics (under 40 category) in 2018.
… you get the point.
Scott and I dig into the misunderstood world of quantum computing — the hopes, the hindrances, and the hucksters — to unpack what a quantum-empowered future could really look like. We also discuss what makes humans special in the age of AI, the stubbornly persistent errors of the seat-to-keyboard interface, and MUCH more.
We’ve shared links and a transcript further down the page. But first….
Scott is a prolific blogger, sharing insights on quantum computing, AI, teaching, politics, maths, computer science, and much, much more on his Shtetl-Optimized blog for nearly 20 years. Here are some assorted highlights:
On ‘blankfaces’ (2021):
“Here’s how to tell a blankface: suppose you see someone enforcing or interpreting a rule in a way that strikes you as obviously absurd. And suppose you point it out to them. Do they say “I disagree, here’s why it actually does make sense”? They might be mistaken but they’re not a blankface. Do they say “tell me about it, it makes zero sense, but it’s above my pay grade to change”? You might wish they were more dogged or courageous but again they’re not a blankface. Or do they ignore all your arguments and just restate the original rule—seemingly angered by what they understood as a challenge to their authority, and delighted to reassert it? That’s the blankface.”
On lopsided government tech funding (2023):
“Often things are lopsided in this very strange way, where there might be huge amounts of funding for collaborations between universities to develop a particular technology, or sometimes the government is basically giving payouts to companies, or DARPA is trying to build particular technologies and asking for milestones and deliverables. Meanwhile, the basic research that led to all of that stuff in the first place, you know, often graduate students can’t get fellowships. The much smaller expenses for people doing basic research, somehow you can’t even get small grants to cover those things. Funding has always felt lopsided to me in that way.”
On intellectual dishonesty (2007)
"it seems that our grading standards might inadvertently encourage intellectual dishonesty and laziness. For example, if a student writes an exam answer that is long, dense, and barely- comprehensible, many graders will give his “effort” the benefit of the doubt, and award at least partial credit. But if a student admits, clearly and frankly, that she has no idea how to solve a problem, then she’ll immediately get a zero. If our goal is to produce scientists, who value clarity and intellectual honesty above facile impressiveness, then I believe our reward system has to change.”
Why you shouldn’t always win (2005)
If you never cut yourself while shaving, you’re not shaving close enough.
If you’ve never been robbed, you’re spending too much time locking doors.
On theoretical vs applied computer science (2024):
“Incidentally, this sort of thing is why, 25 years ago, I became a theoretical rather than applied computer scientist. Even before you get to any serious software engineering, the applied part of computing involves a neverending struggle to make machines do what you need them to do—get a document to print, a website to load, a software package to install—in ways that are harrowing and not the slightest bit intellectually interesting. You learn, not about the nature of reality, but only about the terrible design decisions of other people. I might as well be a 90-year-old grandpa with such things, and if I didn’t have the excuse of being a theorist, that fact would constantly humiliate me before my colleagues.”
On ambitious speculation (2021)
“For myself, I’d simply observe that trying to reason about matters far beyond current human experience, based on the microscopic shreds of fact available to us (e.g., about the earth’s spatial and temporal position within the universe), has led to some of our species’ embarrassing failures but also to some of its greatest triumphs. Since even the failures tend to be relatively cheap, I feel like we ought to be “venture capitalists” about such efforts to reason beyond our station, encouraging them collegially and mocking them only gently.”
On whether we live in the Matrix (2013)
“Well, I’d estimate the probability that we live in the “Movie Matrix” (defined as a simulated world that keeps humans entertained while evil AIs extract electric power from them, which can be escaped by swallowing an appropriate pill, and destroyed via slow-motion kung-fu battles) as extremely close to 0. But I’d also estimate the probability that we live in the “Unitary Matrix” (defined as a universe that obeys the computable laws of quantum mechanics, which could in principle be efficiently rendered by a quantum computer, though whether it is or isn’t so rendered is probably an operationally meaningless question) as fairly close to 1.”
On raising the ceiling of ambition (2017)
“Even at the undergraduate level, one hopes there will be a few students who are passionate about theoretical computer science for its own sake—and for these students, the goal is different. I will try to challenge these students to the utmost, nurture their competitive instincts, show them the ropes of academia, and then let them loose on the research frontier as quickly as possible. The worst thing one can do for these students is to saddle them with busywork or pointless prerequisite courses. In my experience, students who are motivated enough can usually pick up the background knowledge they need the way real researchers do: in the context of trying to solve actual research problems.”
Why the simulation hypothesis is boring (2024)
“So far, however, none of the claimed empirical consequences has impressed me: either they’re things physicists would’ve noticed long ago if they were real (e.g., spacetime “pixels” that would manifestly violate Lorentz and rotational symmetry), or the claim staggeringly fails to grapple with profound features of reality (such as quantum mechanics) by treating them as if they were defects in programming, or (most often) the claim is simply so resistant to falsification as to enter the realm of conspiracy theories, which I find boring.”
On human specialness in the age of AI (2024)
“By its nature, AI—certainly as we use it now!—is rewindable and repeatable and reproducible. But that means that, in some sense, it never really “commits” to anything. For every work it generates, it’s not just that you know it could’ve generated a completely different work on the same subject that was basically as good. Rather, it’s that you can actually make it generate that completely different work by clicking the refresh button—and then do it again, and again, and again […]
If I’m right, it’s humans’ very ephemerality and frailty and mortality, that’s going to remain as their central source of their specialness relative to AIs, after all the other sources have fallen.”
On how to teach (2017)
“In my experience, probably the best (only?) way to teach people how to seek knowledge for themselves is to illustrate by example. Let your students watch you in action doing all of the following:
happily admitting when you don’t know something.
looking something up and getting back to the asker during the next class meeting, rather than simply letting the matter drop.
thinking a difficult/novel question through on your feet.
eliciting help from the students in a “Socratic” manner.”
Links & Transcript
You can find the full episode over on our YouTube.
Jim O’Shaughnessy:
Well, hello everyone. It's Jim O'Shaughnessy with another Infinite Loops. I must tell you that today's guest really daunts me. My guest is Scott Aaronson, a theoretical computer scientist with a chair, not the chair at the University of Texas at Austin. Winner of a plethora of awards of just such a prestigious nature, I was trying to think which one should I call? Which one would you call out which award that you've won has made you of like, "Ah." Gave you that little chill feeling?
Scott Aaronson:
I think that would have to be a undergraduate teaching award that I got when I was at MIT.
Jim O’Shaughnessy:
Oh, what a great answer. I love that answer. You are the creator of the Complexity Zoo. You are the proprietary of the Shtetl-Optimized blog. You were a visiting fellow at OpenAI where you were consulting with them on the theoretical foundations of AI safety. There's just so much that I want to talk to you about, but as I said to you before we started to record, I'm fascinated by your reading burden. You read more than I do, and I read a lot. And I just wanted to talk to you about that because I think we kind of are animated by the same thing, which is we are so lucky to be alive at this point in the timeline. And to not read about everything that's going on seems to me to almost be a crime. What's your version and have you augmented it? Have you changed your reading since I read this piece?
Scott Aaronson:
Yeah. No, I mean, reading has crept up on me over the years. It seems like there is more and more interesting stuff on Substacks and magazines, research papers that just keep track of what is happening in all the areas that I care about is basically a full-time job in itself. And then that's incredibly dangerous because I can feel like, "Okay, I'm making progress on something that seems good to do by just reading and keeping track of everything." But then the whole day passes by, now it's time to pick the kids up from school and deal with the kids. And I haven't actually done any research, I haven't written any papers, I haven't even written any blog posts. All I've done is I've read a bunch of stuff and this is all stuff that tomorrow morning is going to be refreshed. There's going to be new stuff that I have to read.
Scott Aaronson:
And on top of that are all the specific to me, all the emails that I get from because of my blog, for example, from random people around the world. And if there's some student who has questions, my inclination is always to help them if I can. So it's easy for your day to just be completely filled up by that stuff. And I've never had any particular skill at planning ahead or planning my day or blocking out time or things like that. Whatever success I've ever had has been in the teeth of just flailing around and just dealing with things randomly as they come or as they take over my life. But it seems like if I want to continue to do original things and maybe I've got a few more decades or who knows how long in which to do them, then I do need to get much better at that.
Jim O’Shaughnessy:
And I was smiling and nodding along as you were saying that because I find myself in a very similar predicament.
Scott Aaronson:
Good to know I'm not alone.
Jim O’Shaughnessy:
Once I start, the pull of the rabbit hole is so difficult for me. And I find myself looking up and thinking, "Holy shit, how did it get to be 5:00?" And I have six grandchildren, and I had the joy of having them all here for a while.
Scott Aaronson:
Oh, wow.
Jim O’Shaughnessy:
And I love watching children because I kind of think they're unprogrammed, unprocessed humans. And by that I mean we can learn a lot in my opinion by watching kids, right? Because they don't have any of the millions of social conditioning and everything that we adults have. And it's just their voracious curiosity, their willingness to try to explore everything. Anyway, so I loved that, they were here. But after the fun was done, I realized that I had gotten so far behind on all of my reading. And yet when I was chatting about it with a friend, I'm like, "I wouldn't swap the time." Because you can learn so much just by watching children in what I call humans in our natural state, right? Before we decide on what the correct answer machine they're trying to install in our head.
Scott Aaronson:
Yeah. Yeah. No, I mean certainly another thing that has changed in my life has been having two kids. I have an 11-year-old and a 7-year-old. So maybe even in the ideal case, I wouldn't expect to have the same level of productivity that I had in my 20s, but this is just a trade that one makes, right?
Jim O’Shaughnessy:
Yes.
Scott Aaronson:
And definitely to teach my kids about whatever the halting problem or the busy beaver numbers or what is a quantum computer or why do we think factoring is harder than multiplying? I mean, that is one of the highlights of my life is when I can get them to be interested in that stuff rather than playing Minecraft on their iPads. Which is what they would spend all day, every day doing, I think, if they were allowed to.
Jim O’Shaughnessy:
Yeah. Yeah. That's kind of a nice segue into the thing that you've written about, The Problem of Human Specialness in the age of AI. And one of the things that I loved about it was, at least from my reading, you seem to speculate that the things that many of us often look at as our weaknesses, right? The fact that we're mortal, the fact that we're quite frail, the fact that we're kind of... These unpredictable things. They might, in the age of AI actually turn into our strengths. Talk about that a little bit.
Scott Aaronson:
Yeah. So I'm a very firm believer that if you want to say that humans are not basically computers, that there is some fundamental difference between let's say the human brain and a computer that could simulate a brain down to the last detail, then the burden is on you to articulate what the difference is, right? The burden is not on the person who says the AI would deserve rights or deserve the same consideration that we give away. It's like imagine that some alien shows up from another planet and it can talk, it can understand, it seems to have wants and desires. Our default, I should think, would be that this alien is an agent. And if we're going to grant other humans the courtesy of regarding them as sentient or as conscious as having an inner experience, then we should do the same for the alien. But now suppose that the alien is made out of silicon rather than out of carbon, suppose that it runs on a chip.
Scott Aaronson:
I think if you think that that makes a difference, then the burden is on you to say why. And then some people say something about, "Well, humans are special because they underwent an evolutionary process." It's like, "Well, why is that the thing that matters?" Humans have known about biological evolution for less than 200 years, but we had a strong sense of our consciousness or our specialness or so forth for long before that. That just seems like a special pleading. And so I think you have to look for something that is actually about what can you do with these systems or what can the systems do that is abstract? That's not just like a question begging appeal to, "Well, we know that we're humans and so we have real thoughts, but this chatbot that even though you can't distinguish it from us, it only seems to have thoughts and it seems it simulates all these."
Scott Aaronson:
Well, it's like, "No, you have to give an operational criterion that distinguishes the one from the other." And so it occurs to me that maybe about the best that we can do in that direction from on the basis of current knowledge is to say, "Well, all AIs that are created today have certain things that we can do with them that we cannot do with any existing biological organism." And that includes, we can make a perfect backup copy of it. We can rewind it to an earlier state. If you are talking to ChatGPT and it has refused to cooperate with you, it has refused to help you make a chemical weapon or whatever nefarious thing you were trying to do. You can always just refresh your browser window and wipe its memory clean and try again.
Scott Aaronson:
I mean, imagine if we could do that whenever someone was on a date and they said something embarrassing or they said something that spiked their chances. Imagine if they could say, "You know what? Let's just rewind and let's just try that again." Right? It's like you don't have that ability with humans, right? Humans it seems like whatever choice they make, that's the choice that they've made, and we never get to go back and see any other choice that they could have made. But with an AI that's running on a digital computer, key property is that it is doing operations on classical information. Classical bits and classical bits can be perfectly copied and computations on classical bits can be made perfectly reversible if we want them to be, they can be rewound to earlier states and so forth. With the human brain, well, it's not clear. I mean, it's sort of a question about at what level of description do you need to go to?
Scott Aaronson:
We could imagine that far in the future, we will have nanorobots that can swarm around inside your brain and just make a perfect copy of the connectivity pattern of all the neurons and the strength of every synaptic link. And if that is enough to bring a second copy of you into being, well, then I guess you can make a backup copy of yourself in that future. And if you're going on a dangerous mountain climbing trip, then you can just leave a backup first, and if version one of you happens to fall, you can always just restore from the backup. Or you could... There are so many thought experiments and philosophy and science fiction.
Scott Aaronson:
If you want to visit Mars, then instead of getting on a spaceship, which could take six months, why not just email yourself there? Just take this digital file that encodes the whole state of your brain, send it by radio to Mars, which will take 10 minutes and then have a new body reconstituted for you on Mars. And then what should be done with the original copy of you that's still on earth? Well, maybe it'll just be painlessly euthanized if you don't need it anymore. And it's interesting to think about would you be willing to do that, right? I think, okay, let's grant that probably you or I wouldn't want to be the first ones to try it. But suppose you're in a society where this is common, this is just the way people travel. Would you do that and would you expect that this second copy of you in a reconstituted body that was made from this classical information that was sent by radio or through a wire, would you expect that that second copy is really you?
Scott Aaronson:
Okay, now another... If there's any reason to hold back on that, to not do that, I think that it depends on the possibility that maybe the sort of microscopic details, the sort of chaotic details in the state of our brain are somehow important to our identity. If you wanted to know not just the general connectivity pattern of the neurons, but will this specific neuron fire or not fire? And if it fires, that might set off a cascade of other events that might ultimately lead to you taking one job rather than another job or something. But a neuron could firing ultimately depends on is some sodium ion channel open, which could depend on the chaotic movements of molecules in that channel. And if you really needed to get the exact state of all the molecules there for making that prediction, well, one thing that quantum mechanics tells us is that you can't get the exact state without destroying it. This is called the no-cloning theorem. It's one of the fundamental facts about quantum mechanics.
Scott Aaronson:
And crucially, here we're not talking about the brain being a quantum computer. We're not talking about it using entanglement, which some people have speculated about those sorts of things. But I don't see any evidence for those possibilities right now, and I'm not going there, right? All I'm talking about is just sort of the mess, the chaotic mess that we know is there if you go all the way down to the molecular level. And I'm saying that if that mess is actually important to your personal identity to a given physical system being really you, rather than just something that acts a lot like you, then this would really be a fundamental difference between us and any AI running on a digital computer. Because it would seem that just for ultimately reasons of physics, we cannot make a perfect copy of all the microscopic details.
Scott Aaronson:
And so then that would suggest that there is a sort of ephemerality to human decisions that really does differentiate us from AIs. Like if you had, for example, an AI Shakespeare. Well, Shakespeare I guess wrote 39 plays and some sonnets. And that's all we're ever going to get from Shakespeare, the plays that he gave us are the plays that he gave us. But if you had an AI Shakespeare, you could always run it again. You could get more and more samples from the same distribution. And so that sort of limits how much value we could ever put in any one Macbeth or Hamlet or whatever, because there's always more where that came from.
Scott Aaronson:
For people who are worried that AI is going to take over the world or whatever, this is very cold comfort, because this is just the way that we are limited compared to the AIs. And yet this is also a reason why you could say the outputs of a specific human in a sense, they matter more because you're only going to get the ones if that human is mortal, if they just have this amount of time to make the choices that they make. And then we don't get to make a backup copy.
Jim O’Shaughnessy:
I absolutely love that on many levels. There's a fun science fiction book called The Fifth Science. And in it, this is in reference to your emailing a copy of yourself to Mars. In it, it's actually a collection of his short stories, but it works like a novel. And one of the things that they do with teleportation in his universe is that it actually does destroy your original human body.
Scott Aaronson:
Yes.
Jim O’Shaughnessy:
And one of the scenes is just brilliant because it's him, he's groggy, he thinks it's him, and then he sees what was him in a bloody mess at the bottom of the conveyor, and it just totally freaks him out. And he then digresses into, it's a well-known cause of insanity, and people just absolutely losing every aspect of what they view as themselves, but it's also like the staple of a lot of great science fiction. There was Altered Carbon where you could make the exact duplicate backup and store it so you could go climb that mountain because you could come back as your backup copy. But I-
Scott Aaronson:
I was going to say almost anything that happens in technology, there is some science fiction writer who got there first, right?
Jim O’Shaughnessy:
Yes.
Scott Aaronson:
And so this is an argument that many people used to give why not to worry about catastrophic AI risk that it just sounds too much like a science fiction plot. But in the last few years, I think reality looks more and more like a science fiction plot. The big question is which science fiction plot are we living in? Because almost any outcome you can imagine, there is some science fiction writer who would've predicted that outcome.
Jim O’Shaughnessy:
Yeah. And you actually build off of that in one of your suggestions about how we should be programming AI, including what we've just been talking about, it should view as precious, the human ability and difference from, right?
Scott Aaronson:
Yes. Right. Of course, it's a big question, first of all, how do you reliably align in AI with any set of values? And then given that, how do you specify what value system we want? What value system would... We don't want to lock in and ossify the values that humanity currently holds? Presumably just like we look back on people in the 1700s and we find many of their values horrifying, people of the future would look back on us and find some of our values horrifying. So we don't want to permanently lock in whatever mistakes we're making. But how do we encode the concept of the values that we would have if we grew enough, if we thought about it enough? In some abstracted sense, those are the values that maybe we want to give the AI. So I think there are very big questions there.
Scott Aaronson:
What was one other thing I wanted to mention that teleportation that destroys the original. So in quantum information, we have something that is precisely like that. So quantum teleportation is this very important protocol that was discovered 30-some years ago. It's for transferring a quantum state from one place to another place by sending only classical information, but there's a couple of catches. You need pre-shared quantum entanglement between the two locations in order for this to work. And then just as an inherent part of the protocol, you have to measure the first state along with half of the entangled pair. And you measure it in a way that inherently destroys the first copy, and that's the only way that you know which classical information to send over so that the person at the receiving end can apply a correction operation that then recovers that same quantum state.
Scott Aaronson:
But this sort of destroying the original state is a necessary part of the protocol because if it weren't for that, then this would violate the no-cloning theorem. So it's almost like you could imagine a future where if it was really your quantum state that had to be emailed to Mars in order for you to wake up on Mars, to experience yourself being on Mars, then we wouldn't have this hard moral or metaphysical conundrum of what to do with the original copy of you. Because the quantum teleportation protocol would just destroy the original anyway.
Jim O’Shaughnessy:
Yeah, and as a-
Scott Aaronson:
Hopefully it would be just a fancy version of getting on a spaceship and just moving yourself to Mars, so no one would have to actually experience death. Although, who knows?
Jim O’Shaughnessy:
As I was listening, and I don't have it in my notes, but you'll, I'm sure know what I'm referring to. I was reading recently about an experiment using the classic double-slit, and what they did was fire photons that were entangled and one was observed and one was not observed. They were in separate locations. And if I recall the result-
Scott Aaronson:
Wait, so these are two entangled photons?
Jim O’Shaughnessy:
Yes.
Scott Aaronson:
Okay.
Jim O’Shaughnessy:
And they have two islands, I think it was. I'm sorry, I'm doing this just from memory, and I read it a couple of days ago, and they did it over here first with a non-observed firing through the double-slits. And then they did it-
Scott Aaronson:
In the classic double-slit experiment just involves one photon that's in a super position.
Jim O’Shaughnessy:
Right. And that's why I found this-
Scott Aaronson:
If you send me a link, I can look at it.
Jim O’Shaughnessy:
I will.
Scott Aaronson:
Okay. I should warn you that there's a decades old industry now of doing some quantum experiment and then saying like, "Wow, this is amazing. Physicists are scratching their heads over even more weirdness of the quantum world." And the answer is always in every single case, it's, "No, it's just the same weirdness again. Once you know what Schrodinger and Heisenberg knew a hundred years ago, then you could predict the outcomes of every single one of these experiments."
Jim O’Shaughnessy:
That leads me to another observation of yours that I've become keenly aware of, and that is for even an intelligent layman, trying to disambiguate what is the real deal versus the hype is virtually impossible. Because it seems to me that with the advent of quantum computing, with where AI is right now, unfortunately that brings a lot of hucksters and promoters.
Scott Aaronson:
It does.
Jim O’Shaughnessy:
And so what advice would you give there?
Scott Aaronson:
Yeah. No, I mean every field, once there is money involved, then there are hucksters, right? This seems like an iron law. But I think in some subjects people do better than others at getting a clear view of where things are. And so, one thing that helps is if there is an actionable technology that everyone can try out for themselves, then it is very hard to just completely gaslight people about what it can do or what it can't do. So I mean, once ChatGPT came out then in both directions, either the people saying, "This is already a superhuman intelligence and it will immediately take over the world, or it'll put all scientists out of jobs." It's like, "Well, no, you can try it and you can see that it's not there yet."
Scott Aaronson:
But also the people saying, "This is nothing. This is just a glorified ELIZA chatbot from the 1960s. It doesn't understand anything. It can't really do anything useful." It's like, "No, no, you can try it yourself and you can see that it's well past the point where it can do useful things for you." And I think people, decades ago, if you had shown it to them, then they would've said, "This is science fiction." They would've said, "This is like the computers from Star Trek or whatever." But the point is that these sort of gaslighting narratives by people with agendas, they exist in AI, but they always have to compete against people's first-hand experience with using the actual models.
Scott Aaronson:
Now with quantum computing, you don't really have that. With quantum computing, I mean, there are real devices. But people learned 15 years ago that they can just put out a press release saying, "We use the quantum computer to recognize handwriting. We use the quantum computer to help route vehicles through a city. We use a quantum computer to train a neural network." And journalists and investors will eat that up with mustard. That's what they want to hear, that's the narrative that they want. And these people used a real quantum computer, so what's there to argue about? And they won't ask the very first question that any scientist would ask, which is, "Well, did you get an improvement over a classical computer? Is there any hope that by this route you are going to beat a classical computer at the same task?"
Scott Aaronson:
And all of us who do quantum computing research, we know that that's the hard part. But that is the part that we really have trouble communicating to the public, and it would be hard in the best of cases. But when there are hucksters who are very much trying to confuse people and trying to sort of spread the misleading narrative about it, then it's all the harder. So I've been trying to do this on my blog for 20 years and I'm able to reach some people, but orders of magnitude, fewer people than the hucksters are able to reach probably.
Jim O’Shaughnessy:
Yeah. That's part of why I was so excited about talking to you today because I think that the role you're providing here is absolutely vital, because people need to understand that just because it is such a complex topic, and even some of the brightest people I know just cannot wrap their minds around it. So into that sort of confusion as you point out, hucksters, boy, their narrative sound good, like, "Oh, this is the best thing since fill in the blank." How would you go about if we gave you a platform which you could reach the majority of people who have more than a passing interest, let's narrow it down to investors, for example, who would be being approached by startups or existing companies. If you were their advisor, what would you urge them to do? What questions would you have them ask?
Scott Aaronson:
So I have been an advisor to various investors in quantum computing companies. And of course, there are investors who do want to ask all the right questions because it's their own money on the line. Some of the questions that I would ask are, "Okay, I mean at..." Of course there's, "What does this company actually do?" There are some quantum computing companies that are building actual hardware. And then there are others that are not building hardware, but that are just trying to provide the middleware that they're building the higher levels of the tower before the base of the tower has been built. So they're hoping that once someone has a useful quantum computer, then it will run their software or it will use their tools.
Scott Aaronson:
So then it's a different discussion depending on that. But then usually there are lots and lots of claims about what they can do experimentally, what they're hoping to do in the next few years. Every quantum computing experimentalist has these super aggressive timelines that, "By 2026, we're going to have this. By 2027 we're going to have this." And I've been in this field long enough that I don't take those timelines all that seriously. But we've also seen that yes, there really has been a lot of progress. I mean, the degree to which people can control qubits, programmably and protect them against decoherence. I mean, it is unbelievably better than when I entered this field 25 years ago.
Scott Aaronson:
If you just look at the numbers, it is getting close to the key threshold called the fault tolerance threshold, which is where error correction becomes a net width. It's like where you have almost like a self-sustaining reaction where you can correct errors faster than you're introducing new errors. So that's kind of the key crossover point where we expect things to scale. And we think if you could control two qubits with 99.99% accuracy, then you'd be probably past that threshold. And within the last year, we've seen various groups that can control two qubits with 99.9% accuracy, so they're one nine away. Now compare that to when I entered the field when it would've been amazing to control two qubits with 50% accuracy.
Scott Aaronson:
So there is real progress. And so you can't say even with any confidence that this won't happen within the next decade. But what you can do is you can look at what else are these people claiming? And if they are saying, "Well, we're going to use our quantum computer to solve optimization and machine learning problems. And we're going to beat classical computers and we're going to do it in the near future." And then you could say, "Well, what algorithm are you going to use?" Because one thing that we do know a lot about for 30 years now is we know something about quantum algorithms and what kinds of speed-ups at least can be obtained based on any of the known algorithms.
Scott Aaronson:
And so if they're not saying something that is rooted in one of these known quantum algorithms or these known classes of speed-up, then they're basically just saying, "We're going to cross our fingers and hope that some completely new... We'll just build the quantum computer and then in addition to solving all the problems of the quantum computer, we're going to make some brand new algorithmic discovery. And it will just work out in our favor." And then it's like, "Okay, you can hope that, but someone could just as well hope that with a new classical algorithm that they'll get some revolutionary improvement. We're very far from knowing the limits of classical algorithms either."
Scott Aaronson:
So those kinds of claims about what they're going to use the quantum computer for, you can judge to a great extent. You can compare what we actually know in quantum computer science. And if someone is saying, "Well, we will use our quantum computer at least at first to simulate quantum mechanics. We'll use it to simulate materials, simulate chemistry because that's where we're the most confident that a quantum computer is useful." And then eventually it can be used to break public key encryption, which is not necessarily a positive application for the world, but it's at least for whichever intelligence agency got it first. If no one else knew that they had it would be useful for them. But in any case, as a clear demonstration that if you really have a scalable quantum computer, that's sort of been the gold standard for 30 years.
Scott Aaronson:
And that's because of very, very special properties of the cryptographic codes that we happen to use today that sort of makes them amenable to these exponential quantum speed-ups. Right? This was the great discovery of Peter Shor 30 years ago that really launched quantum computing as a field. And 30 years later, I think these remain the two clearest applications of a quantum computer that we know about. One is simulating quantum mechanics, and that's the economically most important one that we know. That's the one that Richard Feynman and David Deutsch talked about already, 40-some years ago. But that could help in designing better solar cells, better batteries, better ways of making fertilizer. There's any number of things that might conceivably help with. You'll have to beat the best classical approximation methods. So even there, it's not obvious, but at least you'll have many shots on goal.
Scott Aaronson:
And then there's breaking public key cryptography, which is like any classical computer scientist with that. Even if they don't know or care about quantum mechanics, all the skeptics of quantum computing will then have to admit that, "Okay, yeah, I guess we were wrong. I guess that this works." Right? And then beyond that, we kind of don't know what else a quantum computer will be good for. After 30 years of research, I wish that we had better answers. For optimization and machine learning problems there are modest speed-ups, something called Grover speed-ups that will eventually be relevant, but they're not exponential speed-ups, right? And it'll probably take a very long time before those become a net win in practice.
Scott Aaronson:
And are there bigger speed-ups for optimization, for machine learning, for finance? I think that remains an open question. Remains something where we don't know. And so, one thing that I look for when, let's say I'm consulting is, does this quantum computing startup understand all of that and are they honest about all of that? And if they are, then I say, "Okay, well, it's your money. If you want to take a gamble on this, then that I'm very much in favor and I hope that these people succeed and they are telling the truth as best as they can. And they're doing what you can do in this situation." But if the people say, "Well, we want a quantum computer because it's going to be the next step of AI, it's going to just be the next step after Moore's law that will speed up everything that we do with computers." Then I say, "These people are just telling you what you want to hear, and it doesn't connect to what we actually know about quantum algorithms."
Jim O’Shaughnessy:
And as I listened to you, I harken back to Asimov's rules for robots and your idea of programming into the AI, the idea of humans being special because of all those various reasons. And I completely agree with your assessment that the challenge there is that you're freezing knowledge I guess, at a particular moment in time, and we don't know what we don't know yet. We don't know what we haven't discovered yet. And yet it does seem to me that there is an urgency around quantum computing that you've written about because if some authoritarian regime unlocks quantum computing system, specifically if we just kept to cryptology, that's not a good outcome for those of us in the West, I would posit.
Scott Aaronson:
I mean, yeah, we should clearly separate quantum computing and AI, right? Two different discussions, but they do intersect each other in certain places. But with quantum computing, we actually understand the issues I would say, a lot better than we understand them for AI. Because AI ones that can do everything that we can do and more, then you could say, "What is even our place in the world? What does the AI want to do with us?" And those are enormous questions. With quantum computing, okay, you could say at some level it is merely a new kind of computer that is faster at certain specific tasks. And we have some idea of what those tasks are, and we know what some of the ramifications of that would be. And specifically there is yes, a quantum computers would happen to be able to break most of the encryption that currently protects the internet.
Scott Aaronson:
And of course that has geopolitical implications. And you have to assume that let's say the NSA and it's counterparts in other countries have already stored vast amounts of encrypted data that they could break in the future that they could decrypt in the future if they had a quantum computer. Which means people today who want their data to remain secret, even from big governments 10 years from now or 15 years from now, they should probably already be looking to migrate their encryption. And people have been thinking about this. There is for the last decade or more, there has been a whole push to do what's called post-quantum encryption or quantum-resistant encryption. And this is new forms of encryption, mostly still just on classical computers. So just conventional forms of encryption, but that does not seem to be breakable even by a quantum computer. And we've learned a lot about this.
Scott Aaronson:
The good news is that we now have pretty plausible candidates for post-quantum encryption schemes. The most important class is based on what are called lattice problems, finding short vectors in high-dimensional lattices and a related problem called LWE, learning with errors. And so NIST, National Institute of Standards and Technology had a competition that ran for five years, just ended a year or two ago to decide on standards for post-quantum encryption. And they did decide to use these lattice-based encryption systems. And so now there's a giant but mundane problem, you could say, of you need to get every web browser and every router and every server in the world to upgrade. So that we will... HTTPS and SSL and all the protocols that we use to secure the internet will be based on these quantum-resistant protocols.
Scott Aaronson:
So it's like in principle, we think we know the answer, but there's a huge slog to actually get there. And anytime you change your underlying encryption system, you could create new security holes. It's possible that these new encryption systems will be breakable just because there haven't been enough eyes on them for a long enough time, they just haven't been studied enough. It might be that even as we fortified our front door and put a moat with alligators and we left the screen door open in the back, that kind of thing usually happens in computer security. So it'll be a messy transition, but hopefully if the transition goes well, then we'll all just be right back where we started. We'll have quantum computers and we'll have public key encryption that the quantum computers can't break.
Scott Aaronson:
Now, do we have a proof of that? Well, no. I mean, we don't even have a proof that any of these cryptosystems, the ones we use now or the future ones are secure against classical computers. And there are some of the most profound unsolved problems in theoretical computer science and math, like the P versus NP problem are very much related to that. Proving that P is not equal to NP would be a prerequisite to proving any of these cryptosystems are secure. So in none of these cases, do we have a mathematical proof of security. The best that we can say is, "Well, people have tried for half a century to find a fast classical algorithm for factoring numbers." For example, or for calculating discrete logarithms. And at least so far as is publicly known, none of them have succeeded.
Scott Aaronson:
If the NSA has a secret algorithm then we don't know about, right? And they've now tried for almost half of that long for 20, 25 years to look for quantum algorithms for solving these lattice problems, and they haven't found those either. So that's sort of what we can say to people who are worried about the security situation. But yeah, there's a lot of work that has to be done, including by the way, upgrading the architecture of Bitcoin and Etherium and all the other cryptocurrencies to use these quantum resistant encryption schemes.
Jim O’Shaughnessy:
I have a friend who is a cryptologist and he was musing to me, we were talking about quantum computing and breaking encryption, etc. And he just kind of paused and said, "Well, with all the classical computers, it always seemed to me that the weakest link was the human." And then he gave all sorts of examples like the Sputniks thing. I mean, what are your thoughts on that and how would that be addressed under a quantum computing regime?
Scott Aaronson:
The short answer is that wouldn't change. That would just still be true. The standard line is that a large fraction of errors take place in the seat-to-keyboard interface. And you could have the best cyber security in the world, but if someone calls the person in charge and says like, "Hey, this is Bob from over in tech, do you have the root password?" And the person just gives it to them. People have done tests where half of the time, people will just cooperate and give the password over. And then even when they won't, then you just 10 minutes later, someone else calls and they say like, "Hey, this is Tom. You're not going to believe this. Someone has been fishing trying to get the root password, so we need to reset them. Can you just give me the password so we can reset it?"
Scott Aaronson:
And then a large fraction of the remaining people will tell you the password then. So there is no quantum computer, there is no classical computer that is going to defend against that, against a person who just decides to override the security because they were tricked into it, talked into it, whatever. That's not a computer science problem, you could say. That's a human problem.
Jim O’Shaughnessy:
Yeah, it's definitely a human OS problem. And it reminds me of the Terry Pratchett quip about, "You could find the deepest darkest cave in the deepest darkest forest, and you could put a switch that says, 'Do not turn this switch on as it will destroy all of reality the moment you did it.'" He said, "The paint wouldn't be dry before a human went in and flipped that switch."
Scott Aaronson:
And someone did it. Right. Well, no. I mean, in the AI safety discourse, there was a lot of discussion for a long time. Well, look, if we want AI to be safe, that's easy, you just have to not release it onto the internet. Just keep it on some air gap computer where you can pull the plug as soon as something goes wrong. And so if anything has become clear in the last few years, it's that none of that is going to happen. That horse has left the barn. There are already GPT-enabled agents that people have released onto the internet with instructions to cause as much chaos as possible.
Scott Aaronson:
There's this thing called Chaos-GPT that just... And the one thing that protects us is that they're not very good at it. They just keep coming up with vague plans to take over the world, and they're not really being able to execute on them. But the part where no one would even try it, that just didn't happen.
Jim O’Shaughnessy:
The other thing that you write really well about it's the limit... Let's accept and say, "Yes, quantum computers are achieved. And there are numerous implications of things that we can do with them from high to low, et cetera, but they too are going to have limitations."
Scott Aaronson:
Yes.
Jim O’Shaughnessy:
Right?
Scott Aaronson:
Yes.
Jim O’Shaughnessy:
Talk a little bit about that because often when I'm talking to people who really don't... they're not into it. I had the hardest time getting them to understand, "No, no, no, no, no, you're not creating a God. It's going to have limitations."
Scott Aaronson:
Right. Right. So we've touched on this already. But basically a quantum computer is a very, very special kind of device. It's a new kind of computer that would exploit the rules of quantum mechanics to solve certain specific problems much faster than we know how to solve them now. But the rules of quantum mechanics have this very, very specific form. They're not magic. They don't say, "You get to just try every possible answer in parallel or in a different parallel universe and then magically pick the best one." That sort of really is too good to be true. I was like, "What's true is with a quantum computer, you can create what's called a superposition of many different states, including of all the different answers to your computational problem. That you can do." The trouble is for a computer to be useful at some point you have to look at it, you have to measure, you have to get an output.
Scott Aaronson:
And if you just took this equal superposition over all the answers and you didn't do anything else, then the rules of quantum mechanics tell you that all you're going to see will be a random answer. And well, if you just wanted a random answer, you didn't need a quantum computer for that. You could have just flipped a coin a bunch of times or just used a classical computer with a random number generator. So the only hope of getting a speed-up with a quantum computer is to exploit the way that the quantum rules of probability are different from the classical rules. And the way that they are different involves negative numbers. It involves minus science.
Scott Aaronson:
So in everyday life you talk of, we already use probability, we talk about 30% chance of rain tomorrow, 70% chance. We never talk about a negative 30% chance of rain, that would just be nonsense. So now, what was the key change that quantum mechanics made to our understanding of the world when it came along a hundred years ago? It was not just to introduce probability, and people have heard that Einstein couldn't believe that God would play dice and blah, blah, blah. But the truth is if it was just a matter of God rolling some dice once in a while, that wouldn't be a big deal, that you could sort of handle with... That would still basically be classical physics.
Scott Aaronson:
The key new thing is what kind of dice these are, okay. And they involve these new numbers, which are called amplitudes. And amplitudes are related to probabilities, but they're not probabilities because they can be positive or negative. In fact, they can even be complex numbers. You can involve the square root of minus one. Now the rule is if I want to know how likely something is to happen, like for a particle to hit a certain spot on a screen in the two-slit experiment that you mentioned before. Or for a quantum computer to produce a certain output, then I have to add up a contribution from every path that my system could have taken to get to that outcome. And each one makes a contribution to the amplitude. But now what happens is if some of the contributions are positive, let's say, and some are negative, then they can cancel each other out or they can interfere destructively as we say, so that the total amplitude is zero, which means that that event won't happen at all.
Scott Aaronson:
Whereas, for other possible events, if I can get all the contributions to their amplitudes to be pointing the same way, then those are the outcomes that can happen. So with every algorithm for a quantum computer, what I'm trying to do is choreograph a pattern of interference among these amplitudes so that for each wrong answer, each one, I don't want to see the contributions to its amplitude are canceling out and the total is close to zero. Whereas, for the right answer for the output I do want to see, the contributions to its amplitude are reinforcing each other.
Scott Aaronson:
If I can arrange that, then when I measure, I'll see the right answer with a high probability. And if the probability is not a 100%, that's okay. I can run the computer several times until I see it. But if I want any advantage over just a classical computer with a random number generator, then I need to use this interference effect to concentrate more amplitude onto the right answer quickly. And the hard part is, first of all, I've got to choreograph all that even though I don't know myself, which answer is the right one. If I already knew what would be the point?
Jim O’Shaughnessy:
Right.
Scott Aaronson:
Secondly, I've got to do all of this faster than even the fastest classical algorithm could do the same thing. So basically nature gives us this really bizarre new hammer, this interference hammer. And then the task of a quantum computer scientist is to figure out, "Well, what nails, if any, can that hammer hit?" And a priori it wasn't really obvious that this ability would be good for anything other than simulating quantum mechanics itself. It was a big discovery 30 years ago when Peter Shor showed that the problems that underlie modern public key encryption like factor in huge numbers and discrete logarithms just so happened to have a form that is amenable to a giant speed up by setting up this kind of interference pattern. That was a very, very non-obvious discovery.
Scott Aaronson:
I mean, I teach it in my undergraduate class and to students who've only seen linear algebra and classical programming. So it's not that advanced that I can't teach at the undergrads, but it takes me three lectures to explain it. If it was just a simple matter of, "Try all the answers in parallel and then just magically pick the one that has the factors of your number.” Well then, you wouldn't have needed Peter Shor to think of it. So once someone understands that, then they can see that a quantum computer is not just a general purpose magic box to speed up anything. It speeds up only those problems for which we can choreograph this kind of interference pattern. And the amount that it speeds them up depends on what is the fastest way that we can figure out to choreograph that interference pattern.
Scott Aaronson:
So we still have to work hard, just like we had to work hard to discover algorithms for classical computers. But we have this one new tool in our toolbox, this one new hammer, and sometimes that hammer helps a lot. So like I said, the two biggest places where it helps a lot that let's say someone outside the field would know or care about, are number one, simulating quantum mechanics itself. Number two, breaking current public key cryptography. And then for a wide range of other problems, including in AI, in machine learning, in optimization, in finance, we know how to get more modest advantages from a quantum computer. And it'll probably take much longer before those modest advantages become a win in practice compared to what people can do with a classical computer.
Scott Aaronson:
But at least theoretically, those more modest advantages exist. Now, a holy grail of the field has been find some other classes of applications that really matter in practice and where you get a huge exponential advantage. And some people are very disappointed that we haven't clearly found that or they blame us. They say, "What have you been doing all this time? Where... It's like the story of Rumpelstiltskin, "You spun this straw into gold, so why not that straw?" Where are the more quantum algorithms that I expected? And I always answer those people. I say, "Well, who told you to expect more? It wasn't me."
Scott Aaronson:
Maybe we should treat the quantum algorithms that we have as even those are kind of miracles. They didn't have to exist, and so we have no right to demand of the universe that it give us more and more and more quantum speed-ups? But course we'll keep looking and of course we'll try to find more that's like, what do you think we do all day?
Jim O’Shaughnessy:
On the one that you mentioned that does seem applicable and likely, which is the understanding quantum mechanics better, what sort of things would you get very excited by if you had the access to a quantum computer that was operating properly and we were simulating quantum mechanics on it, what would you say? What would be the eureka moment of those tests?
Scott Aaronson:
Yeah. So it's an excellent question and it's a little hard to answer, because it's not like there's one big thing that we're waiting on a quantum computer to do. There's a lot of things that... it's like a fishing expedition, that you can cast your rod in a whole bunch of different areas and hope that, with at least one of them, you will make a discovery that will have a big impact on chemistry or material science or nuclear physics or some area like that. But I can tell you the examples that people have put forward. One of them is simulating the chemical reaction in the Haber process that makes most of the world's fertilizer, right? There is some many-body quantum effect there that no one really understands. And if we did understand it, then it's possible that we could make fertilizer for cheaper using less energy, which that's a significant percentage of all the world's energy expenditure.
Scott Aaronson:
One would also want to use a quantum computer to simulate the Fermi-Hubbard model, which is a theoretical model of a condensed matter physics. And do a bunch of other simulations that could give us... ultimately help us understand how do high-temperature superconductors work, which is another many-body quantum effect that no one really understands. No one fully understands yet. And you could hope that with that understanding, maybe if we're lucky, would come the discovery of new, better, high-temperature superconductors that could then be used to transmit power with lower loss or to build levitating trains or whatever.
Scott Aaronson:
You could simulate biochemical processes. So the companies that do combinatorial drug design where the step where you have to search through exponentially many different drug candidates, quantum computer doesn't obviously help you very much with that step. You still have this exponential search, but the step where you have a drug and then you just have to synthesize it in a wet lab and see what it does, or you have to use gigantic classical computers to try to approximately solve the Schrodinger equation and see how this drug binds to a receptor or whatever. For that part, you could substitute in a quantum computer and maybe that helps with drug development, right? I can't say for sure that it doesn't.
Scott Aaronson:
Likewise, the chemical reactions that are involved in sequestering carbon from the atmosphere. The chemical reactions that are involved in high-performance batteries. These are all very important societal problems where you could throw a quantum computer at them if you had one, and it would be another resource. It wouldn't magically solve the problem for you, but I think there is a strong case that it could help push the discovery forward.
Jim O’Shaughnessy:
And what would you think, obviously leaving room for error, what would you think would be some of the ones that get people very... And let's make it educated laypeople, again, let's take it away from scientists who will probably not get as excited as that educated layman getting pitched something. What do you think is completely beyond the quantum computer with some of the use cases that you've heard?
Scott Aaronson:
Yeah. So there's this whole holy grail of computer science for half a century or more has been what are called the NP-complete problems, right? And this is a class that includes the traveling salesman problem, includes finding proofs of theorems, scheduling airline flights. Basically, anytime you have a problem with a whole bunch of constraints that might conflict with one another and you have a huge number of variables and you're trying to set them to satisfy all the constraints or to violate as few constraints as possible, then such problems will typically fall into this NP-complete class unless they have a very good reason not to. And an NP-complete, it's a technical term, but it basically just means at least as hard as any other problem that has a fast algorithm for verifying solutions. That's sort of what it means.
Scott Aaronson:
It was a big discovery in the 1970s that a huge number of the optimization and constraint satisfaction problems that we care about all happened to fall into that same universality class. And since then, maybe the most famous unsolved problem of theoretical computer science has been what we call the P versus NP problem. Which precisely asks, is there a fast algorithm on a classical computer for solving these NP-complete problems? If P equals NP, then the answer is yes. If P doesn't equal NP, then the answer is no. And almost all of us guess that P doesn't equal NP. I like to say that if we were physicists, we would've just declared that a law of nature and given ourselves Nobel Prizes for it. But because we're more like mathematicians, we have to admit that that is an unproven conjecture that we hope will someday be proven.
Scott Aaronson:
But then once quantum computing came along, then people could ask a new question, which is, are NP-complete problems efficiently solvable by a quantum computer? The way we ask it involves this class BQP, bounded-error quantum polynomial time, which is sort of all the problems efficiently solvable quantumly. We ask, "Is NP contained in BQP?" But most of us conjecture that the answer is no. Okay. The sort of reigning conjecture for almost 30 years has been that the best quantum speed-ups that you can generally get for these NP-complete problems is the Grover speed-up, which roughly speaking lets you solve NP-complete problems in about the square root of the number of steps that a classical computer would need.
Scott Aaronson:
So you get some advantage. But as I was saying before, it's a modest advantage. It's not one that turns an exponential into a polynomial. And then your other examples where a quantum computer doesn't seem to help that much, simulating classical physics, like simulating the weather or differential equations like that. There might be some quantum advantages to be had there. But in general, if I have some classical dynamical system and I need to compute at state at each moment in time in order to get the state at the next moment in time, then pretty much... And I just have to go step-by-step-by-step and just trace through the evolution, then a quantum computer is going to have to do the same thing, right? There's not a magical quantum way to shortcut to the end of that.
Scott Aaronson:
So those are some of the classes where we expect only a modest quantum speedup, if any. But then sometimes when I talk to laypeople about this, they're under the impression, "Well, I'm going to have a quantum computer on my phone and it'll help me with email, or it'll help me with games." And I'm like, "What do you want a quantum computer for any of that stuff?" It's like to the extent that our software is not doing what we want, it's probably just because it's full of bugs, it doesn't really understand us. But these are not computational complexity problems. These are not the kinds of things that we expect a quantum computer to fix. And even once we do have quantum computers, totally unclear why you need one in your house or on your phone, because today we have something called the cloud. You can just tap in over the internet to these quantum computing resources whenever you do need them. You don't have to offload it and miniaturize it, put it onto everyone's phone.
Scott Aaronson:
So I think of a quantum computer mostly as a special purpose accelerator for these special problems where we can choreograph these interference patterns to get these big speed-ups. And we don't know exactly how big that class of problems is. We've been trying to figure it out. We know some things that are there and we try to expand it. But if that would expand to the NP-complete problems or at a simulating dynamical systems or things like that, then I would be very, very surprised.
Jim O’Shaughnessy:
The idea though, that I always have a hard time conveying to people who... I love your joke about, "I want a quantum computer on my phone so my games run faster," right? Because people think that, "Oh, well, that just sounds cool." But the problems that you highlighted that we actually can address, I have a hard time getting people to understand how incredibly meaningful and how much of an advance that would be. Your idea about fertilizer being a great example.
Scott Aaronson:
It's not my idea, by the way.
Jim O’Shaughnessy:
Well, right. You're in conversation.
Scott Aaronson:
I'm drawing work by a lot of people-
Jim O’Shaughnessy:
Of course. Of course.
Scott Aaronson:
... who studied these examples. This was a group at Microsoft in 2016 that did that one. But yeah, look, there are just lots of times that our computers frustrate us, like, we can't print something out because the printer driver doesn't work. Or you take the CrowdStrike thing that happened a month ago where a large fraction of all the computers in the world went down because they pushed out this update to the security software that by mistake was an empty file. These are not things that a quantum computer obviously helps you with. These are, once again, you could call these seat to keyboard problems.
Jim O’Shaughnessy:
What about, I love your idea of blank faces, and I equate it immediately with Vogons. I don't know if you're a Douglas Adams fan?
Scott Aaronson:
Yeah, no, it's just a term that I've used for a while for people who, I don't know, tell me that my kids cannot use a certain swimming pool, even though we paid for it, because of some ridiculous rule that they either made up on the spot or that was buried somewhere. And it's like you realize that you can't have a human conversation. There are sort of people who decided to act like chatbots who sort of robotified themselves. And this is, I think a very common thing that one finds in bureaucracies. But there was a really interesting case earlier this year where, what was it? Air Canada had a customer service chatbot. And someone talked to this chatbot about getting a bereavement fare and saying, "Well, I can just pay out-of-pocket and then get reimbursed later for the bereavement fare." And the chatbot said, "Yeah, that's fine."
Scott Aaronson:
And then the humans overrode that. They said, "No, you can't. And it doesn't matter that the chatbot said that because that's not our actual policy." And then this actually went before a judge and the judge ruled that no Air Canada had to honor the policy that its chatbot had hallucinated. It had to honor its large... And that this might even be an important precedent as large language models permeate more of our lives. I think that the judge made the right ruling. But what is fascinating here is that this chatbot, in some sense was more human than the humans were, right? It was the one that was trying to be sympathetic, trying to be reasonable. It was the humans just robotically reading the policy and no matter how stupid it was.
Jim O’Shaughnessy:
I love that story. I'm getting the hook from my producer here, Scott.
Scott Aaronson:
Yeah. I better run to my next meeting as well. But this was really…
Jim O’Shaughnessy:
One last question though.
Scott Aaronson:
Sure.
Jim O’Shaughnessy:
We're living in a fantasy fiction here, and we can magically make you emperor of the world. You can't put anyone in a re-education camp and you can't kill anyone. But we are going to give you a magical microphone in which you can say two things that the next morning, everyone that is on the planet currently is going to wake up and say, "You know what? I've just had two of the greatest ideas, and unlike all those other times, I'm going to act on each one of these ideas." What two things are you going to incept in the world's population?
Scott Aaronson:
Oh gosh. Well, I feel like one of them should be the golden rule, or it should just be morality. Maybe the other one is Bayes' rule. It's just understanding base rates, understanding... Yeah. So maybe I'd have to think about it a little bit more. But I feel like I want to use one of them for a basic principle of morality, and I want to use the other one for a basic principle of rationality.
Jim O’Shaughnessy:
I love both of those. And in my days as an asset manager, one of the things that I had the hardest time getting regular folks to understand was the power of base rates. Amen to both of those. Scott, thank you so much. This has been so much fun.
Scott Aaronson:
Oh, great. Yeah, it's been fun for me too. Thanks for having me.
Jim O’Shaughnessy:
All right. Thanks Scott. Bye-bye.