Into the Impossible With Brian Keating - Max Tegmark: Will AI Surpass Human Intelligence? [Ep. 469]

Episode Date: December 9, 2024

Please join my mailing list here 👉 https://briankeating.com/list to win a meteorite 💥 Will AI ever surpass human intelligence, discover new laws of physics, and solve the greatest mysteries of our universe?  This week on Into the Impossible, I explore the potential and dangers of artificial intelligence with none other than Max Tegmark!  Max Tegmark is a renowned physicist and machine learning expert who dedicated his career to uncovering the mathematical fabric of reality, proposing that our universe itself might be a vast mathematical structure and that we could be living in a multiverse of endless possibilities. His work goes beyond physics to tackle the transformative power and ethical challenges of artificial intelligence, an area where he believes humanity must tread carefully. In the second part of our fascinating interview, we discuss the development of AI, the impact it will have on science, and our role in all of this.  Tune in to discover if AI will ever surpass human intelligence! Check out the first part of our interview, where we discuss his mathematical universe hypothesis, the search for extraterrestrial life, and AI’s role in science: https://www.youtube.com/watch?v=EFO5lzA_0Og  Key Takeaways:  00:00 Intro 00:28 Is AI limited by embodiment?  10:52 The challenges in AI research  14:53 The role of regulation in AI development  29:53 How AI is going to impact education and science 39:20 Do we need quantum computation for human intelligence?  44:50 Fantastic final four 50:45 Outro Additional resources:  📚 Life 3.0 by Max Tegmark: https://a.co/d/d08r0FJ  ➡️ Follow me on your fav platforms: ✖️ Twitter: https://twitter.com/DrBrianKeating  🔔 YouTube: https://www.youtube.com/DrBrianKeating?sub_confirmation=1  📝 Join my mailing list: https://briankeating.com/list  ✍️ Check out my blog: https://briankeating.com/cosmic-musings/  🎙️ Follow my podcast: https://briankeating.com/podcast  Into the Impossible with Brian Keating is a podcast dedicated to all those who want to explore the universe within and beyond the known. Make sure to follow/subscribe so you never miss an episode! Learn more about your ad choices. Visit megaphone.fm/adchoices

You said this place was steps from the water. We just haven't found the steps yet. How much did we save? Enough. Enough to get lost! Or you could book a stay with Hilton. Welcome to your oceanfront room. Just steps from the water.

The Hilton sale is on now. Book on Hilton.com or the Hilton app and save up to 20% to get the stay you expected. When you want savings, not surprises. It matters where you stay. Hilton, for the stay. Study and play. Come together on a Windows 11 PC.

And for a limited time, college students get the best of both worlds. Get the Unreal College deal, everything you need to study and play with select Windows 11 PCs. Eligible students get a year of Microsoft 365 premium and a year of Xbox GamePass Ultimate with a custom color Xbox wireless controller. Learn more at Windows.com slash student offer.

While supplies last, ends June 30th, terms at AKA.m.m.S. It's very dangerous to bet against AI progress. I shifted from physics to AI research with my group about eight years ago. Five, six years ago, almost all my AI colleagues were still predicting that something as good as chat GPT4 was decades away. Never say never. Any sufficiently advanced technology is indistinguishable from magic. Open the pod bay doors. I want to pivot to that subject now with a quote from your first. first book, where you say the following. I suspect that if we can build such ultra-intelligent machines, then the first one will be severely limited by the software we've written for it, and we'll have compensated for our lack of understanding about how to optimally program artificial intelligence

by building hardware with significantly more computing power than our brains have. After all, our neurons are no better or more numerous than those of dolphins, just differently connected, suggesting that software can sometimes be more important than hardware. I want to ask you the following question. You know, certainly, what Albert Einstein said was the happiest thought of his life, right? Which was that an observer in free fall would experience no gravitational field. He described it as giving him titillations. I want to ask you, to what extent could we train an artificial intelligent hardware or software,

or software, A, what it means to be happy and be what it would feel like to be in freefall. In other words, are we limited by the lack of ability or our artificial intelligences limited by the lack of embodiment and the lack of these kind of teleologically driven feelings and emotions that people have? In other words, can AI generate AE, Albert Einstein? Can it generate new laws of physics? Right now it cannot. I think within 10 years, very likely, yes, maybe even in two years.

It's very important when we talk about all the crazy stuff that I think is likely to unfold in the AI space that we remember. We're not talking about chat GPT 4O about the AI of today. We're talking about the AI of tomorrow next year, three years from now. just like we're not talking about pocket calculators or the AI that beat Gary Kasparovic chess either. There's been a dramatic evolution. First we started with these very narrow AI systems that could take our butt and chess, but very little else. Now we have things that can arguably pass the Turing test because they've mastered language and knowledge to the point to fooling a lot of people that they're human, but they're still very passive. They act a lot like

oracles, you ask them something, they answer it, maybe make you a funny picture. This year, we're seeing an explosion and people trying to make more agentic AIs, which actually have goals and go out and do things on the internet, operate robots of various sorts, land-based or sea-based or flying ones. We are very soon going to be in the situation where the vast majority of the data that comes in into most of the AIs that they're trained on is not just one modality like text, or whatever, but very multimodal, just like for us, right? You are constantly, Brian, taking in about one megabit or one megabyte per second of visual data from your retinas and acoustic data and sensory data, temperatures and pressures, et cetera,

and smells and flavors, and you're synthesizing all that. And that's what you mean by you being embodied, right? Your brain is trained by all these electrical signals coming in, which, as far as the brain is concerned, all the same kind of electrical. signal. It doesn't matter if it comes from the ear or the eye. That's all interpretation, right? If you're dreaming, it still feels like you're embodied, even though it's all fake in that case, right? So by feeding in this kind of data and having that be something that our AIs are trained on, they can develop all the same insights and potentially emotions and intuitions that a human can,

even if they're not actually in a robot body. And then add to that, that we already have, for example, a whole fleet of Tesla's that very much have robotic bodies and are taking in all their sensory input and using it for training. Very soon we're going to have an explosion in humanoid robots. You've started to see them flooding Twitter, existing ones.

We're going to get optimists soon. Short answer, yes. Of course, AI, if it races ahead, will be able to do all the science that we do as well. Some people might disagree and think that there are some secret sauce in our human brain that makes it so special that we can never be outthought. But I think the biggest insight, frankly, that has powered the whole AI revolution,

is just the insight that our brain is actually a biological computer and that there are plenty of other ways of building faster and better computers. This closes out, though. You were completely right in the beginning also, I think, when you said that we have really sucky software in the AI systems we built compared to what's in the brain, when you use something like chat GPT for, oh, it uses thousands and thousands of times more energy to do a task than you would. And your brain, all the amazing thing it does, you use about 20 watts.

Obviously, our software, our systems architectures that we use today for our large language models and other AI tools, so-called transformers, et cetera, are incredibly dumb, you know, compared to what's physically possible. So one of the first things that's going to happen when we get AGI that can do all the jobs better than us is it's going to do the job of AI research better than us and realize, oh, we can redesign our hardware to be a thousand times more efficient and we can redefine our AI software, our architectures, to be vastly more efficient. And then, boom, you know, there you suddenly have something which is vastly beyond our capability.

I want to run by a project that I've been working on. And we have a paper that we're about to submit, although that's always a dangerous statement. And it concerns the following thought experiment, popularized by Einstein. And here's a scenario. You're an astronomer in the mid-1800s. The procession of Mercury's orbit is now considered a problem in classical mechanics. And currently, the astronomers of your time have measured there to be a difference of about 40 arc seconds per century versus the predicted and observed procession rate of Mercury's orbit, there's currently no explanation for this unaccounted phenomenon.

And we decided we put this into a large machine learning system, neural network, tensor flow, all sorts of throw everything we had at it with the technology of the 21st century. And C, would AI, whatever that means, would it come up with a whole new concept of gravity? that gravity is not a force, that it is the byproduct of curved space time. And we've come upon an obstacle. And I wonder if you with your mega mind can tell me what you think that obstacle is, is actually making me much less nervous about the future of AI. So I'm a boomer, not a domer when it comes to AI because of this research obstacle that we've hit.

It's not usual that you hit an obstacle and it makes you optimistic. But what do you think is wrong with this formulation of the problem? We're basically trying to derive a whole new paradigm of gravity from the observed data, which we have going back, you know, 10 centuries or more using JPL databases and so forth. If you had a guess, what obstacle do you think we most likely hit in this problem? I think you probably hit the obstacle that I don't know exactly what architecture you put in for the AI model that you're running. But it's probably very much what Conneman would call System 1 still. It's adjusting its way.

way it's training it to fit some data, et cetera, but without using any symbolic reasoning that's common would call system two. If we think about AI now, we tend to think of these neural networks that can do these things as like the new modern thing and the symbolic stuff is like the old thing because we humans, when we did our pocket calculators, which seem very old-fashioned, they could manipulate symbols. But in evolution, it was exactly the other way around. Do you have any dogs right now? We're about to get one. How do you know that? My time machine I looked into the but you know like dogs are very good at catching tennis balls, etc. And eagles probably have way, way better vision than I do and the ability to analyze visual images. This is all

very system one stuff. That's not what makes humans the alpha species on the planet, right? Well we humans are able to do better than any other languages also take this intuitive. Your summer starts now with Memorial Day deals at the Home Depot. It's time to fire up summer cookouts with the next grill, four-burner gas grill, on special buy for only $199. And entertain all season with the Hampton Bay West Grove seven-piece outdoor dining set for only $49. This Memorial Day get low prices guaranteed at the Home Depot.

While supplies last, price invalid May 14th or May 27th, U.S. only exclusions apply. See Home Depot.com slash price match for details. understanding we have which is sort of fits to data and then see patterns in it and abstract it out into a symbolic description. Galileo, if his dad threw a bunch of balls to him when he was four years old, he could also catch them. But then when he got older, he realized, wait a minute, they always go in the same shape. That's a parabola. I can write a formula. You know, Y equals X squared. And he could communicate it to his friends and colleagues in mathematical language or in Italian or in Latin.

What's made us humans so unique, the reason it's humans, not dogs, that invented the Internet, is precisely this. We have the ability not just do System 1, but also reason symbolically. And we can first get intuition from this System 1 stuff,

and then we can abstract things out, distill out a symbolic representation, maybe this curved space time, maybe you have to get rid of Euclidean space here, et cetera, write down Einstein's theory, and then communicate it with others. And that's the remaining thing we're really lacking still.

That's why we don't have AGI yet because we're not quite there, even though a lot of people are working on it. Now we're in this sort of schizophrenic situation where we've made a big breakthrough also on language with large language models. The two sides still don't really communicate with each other. A large language model cannot introspect and understand how its own brain works and describe things about it. So if a large language model was trained to catch tennis balls, you know, it wouldn't be able

look inside and figure out what the, necessarily what the formulas are. It can discover a trait from data maybe. So I don't think you should blame yourself or the student you're working with. I think at some level, AI isn't quite there yet. I think it's going in that direction. Well, my only argument against it coming from LLM side, which we're not really using the LLM side to do it, but I think the limitation, when I told that to people, I was at a conference, We talked about AI.

David Berlinsky and Peter Thiel were there and other people. And the consensus was this is the worst that AI will be, as right now, and it's growing at this exponentially increasing rate. And I said, okay, hold on a second. One of the reasons it's getting better is because of its training data set and the efficiency of training the data is getting better. Modulo, what I call the mad bot problem where AIs train other AIs, and that leads to like mad cow disease, brain rot, and.

zombie AIs, but ignoring that. And the zombie internet we're heading towards where a large fraction of everything you see on Twitter is written by a bot. That's right. We'll get to that in a moment if you have time. But my point is what's missing here to construct AIAE, you know, artificial Einstein, is not that it doesn't know what happened in Fast and the Furious 12, you know, that it hasn't, we haven't updated the training set.

Now, that's not what it is. There's something of a different character. I don't mean like the ghost in the machine. But I propose that this type of test is a better test and it's a more holistic and it's a more transparent test than the Turing test. The Turing test, people claim that it's already been solved and that passed rather. But I claim constructing a new law of nature, you know, something that was wholly unknown. I would actually settle for constructing a known law of physics. In other words, give it the POSAN bracket and say, now come up with a commutation relate.

Don't tell it anything else. Just give it data. I don't think you can do it. So I think when it does it. Actually, actually, you're right. It can't do it yet. But we're definitely making some progress. I've actually been working at least to know on exactly this problem with my group here at MIT.

For example, we worked on the problems called symbolic regression where you just give it an Excel spreadsheet with tables of numbers. And it tries to figure out what formula do you have to apply to all the columns, other columns to predict the last column. And it took four years for Kepler to stare at his Mars data. and realize it's an ellipse. Our AI discovered that in one hour. We gave it the Schwarzschild solution for a non-rotating black hole metric. It took over a decade for Gull Strand and Panlavae

to realize that there's a different coordinate system where not only do you get rid of the singularity, seeming singularity at the Vent Horizon, but in fact space is entirely flat everywhere outside the center of the black hole. It discovered that also in like an hour. We even discovered a new law and ozone chemistry that nobody actually knew. And then another group of people who actually knew something about ozone chemistry,

some atmospheric scientists wrote another paper and I'm like, oh my God, yeah, this thing that those AI nerds at MIT discovered, we can explain why it is. And they generalized it. And it was quite cool. So nothing at all at the level of what Einstein did, of course. And I really want to downplay the importance of the ozone thing we discovered. But we're making some progress in that direction for sure.

And it's very dangerous to bet against AI progress. I shifted from physics to AI research with my group about eight years ago, right? Five, six years ago, almost all my AI colleagues were still predicting that something as good as chat GPT4 was decades away. Never say never. And finally, can I give you a hard little bit, just give you a hard time a little bit for fun? Yeah, please.

I love it. Because you were jokingly saying you're a boomer not a doomer. And a lot of people like Jan Lacoon and Andrew Eng like to call me a doomer on social media. Upcoming guest, Jan Lecun. Yeah, yeah, yeah. So you can ask him about this also. I don't think of myself. as a Duma at all.

But you mentioned things like pauses, like six, six month pauses and, you know. Yeah, well, yeah, but why, you know, and what, you know, so Dumer was clearly a pejorative term invented by some people from that crowd to make ad hominem attacks about people who disagreed without ever having, so they could have something to say without actually having to rebut any arguments, right? It's just pure ad hominem. Fear, right, yeah. If you tell me that you have some.

some technical questions about equation five in my last paper, you think I'm missing a factor of two. And my response is just that, you know, you're a racist. It feels a little bit like that. Right. I have been called a racist on the podcast by none other than Neil deGrasse Tyson, but I'm not going to get into that now. Okay. But like Dumer, first of all, I'm, I'm very, I love technology, super excited about, about the potential of humanity to flourish with technology. which is why I've dedicated my life to building technology. I don't think that should count as doomerism. In fact, I've talked more, even in my books,

about how amazing life can be with advanced AI than Andrew Eng or John Lagoon ever have. You've worked out how long it would take to go to other galaxies and how you could get there efficiently and stuff like that. So, no, I love to think very big about things I find very hopeful. There are two kinds of doom, can have you can you can you can be very pessimistic and say oh it's impossible AI is never going to work techno pessimism tech you might call that or techno skepticism there are people like rodney brooks who told me some years ago he thinks we won't get

aGI for at least 300 years you know andrew eng talked about how it's worrying about superintelligence is like over worrying about overpopulation on marks Mars that to me is actually pessimistic you're sort of assuming that people are too stupid to solve to solve technological problems. The other big split among people is if we get AGI or super intelligence, whether it's automatically going to be great or whether it might bring problems. And we know that every time we've built some powerful technology, it could be used for good or it could be used for bad.

If you say, hey, you know, we live in this big wooden house. why we put a smoke detector in and get a fire extinguisher, you know, are you a doomer? No, I would say that you're the one who's, who has great positive vision for how your house is not going to burn down and you're going to actually make sure that it doesn't burn down and the future goes well, right? When cars came along and some people were like, hey, maybe we should put seatbelts in them. Car industry was very against it, lobbied really hard against it, saying this is going to destroy the car industry. I say they were the doomers. They were talking negative, making very negative prognostications, which turned out to be bullshit because you know what actually happened

after we passed the law in the U.S. requiring seatbelts and cars? Do you know what happened to auto sales? It exploded because it turned out the main reason holding people back from buying cars where they were scared. And once death started to really plummet, people got the confidence to buy way more cars, saying that we should treat AI and powerful AI like every other technology and have some safety standards to make sure that they get used for good things and not for bad things. I think that's exactly the kind of safety engineering and common sense that we've successfully used for all other powerful tech in the past. I didn't mean to refer to you as a doomer. I just got a little triggered because I very often do get. No, I know. And it's now impact. You know it's serious when it's

impacts a California legislation and someone in Boston, like yourself, opines upon it and very, very supportingly so. And I appreciate that. The Chinese government and the European governments both were concerned about AI risk and have put in place various regulations to make sure that their companies don't do too much crazy stuff, just like with seatbelts, but for AI. U.S. so far has no meaningful AI regulation at all, except for an executive order that Biden put in place that just says that the companies that do the most extreme stuff just have to let the government know at least. So now there is this first ever law that would actually do something.

It doesn't do much, in my opinion. It mostly just says that the stuff that Open AI and Anthropic and Google DeepMind have basically promised to do anyway is voluntary commitments now is actually required, not just by them, but by their competitors too. And then you get this hilarious. uproar, just like with the seatbelt law, where like Open AI came out and said, oh, this is going to cause companies to leave California. It's going to doom. People are saying it's going to doom the California economy, et cetera. And then I remember that actually, that's exactly what they were

saying also when the EU was debating the EU AI Act. Yeah, we're probably going to have to leave Europe. Then they act passed. It's peak pollination season, and my business is scaling fast. To keep the nectar flow, I need a phone plan with top priority data speeds. That's why I chose GoogleFi Wireless. My connections stay strong even when the hive is buzzing. Plus, unlimited plans started $35 a month. Now that's a deal that doesn't stay.

Explore Google Fi Wireless plans today. Plus taxes and government fees. Google Fi Wireless is not subject to data traffic deprioritization during times of high network usage. And surprise, they're still there. All the companies and life goes. on, you know, just like the immediate death that was supposed to happen to the car companies when seatbelts were enforced, mysteriously didn't happen either. So we're seeing all this drama all over again, except people have such amnesia. They always act like these

challenges of good governance of AI is like first time in history of humanity that we ever had to govern a technology. Like, hello? What about cars? What about like codes for electricity, or your house doesn't burn down? What about food safety? What about the FDA? What about us? What about us as professors? You know, so you and I are, I joke, we have the second oldest profession after, you know what. I mean, back in the year 1080, there are professors in the University of Bologna, Northern Italy. And they were, you know, guys scratching on a piece of rock with another rock, as your friend David Kaiser calls it. our profession really hasn't changed that much. Are we the last of a breed, you and I? Will we have training by, I don't even, I never learned to really code with Python. And I'm glad I did.

I see where you're going with this. I comment on the oldest profession first. Okay, go for it. The first thing I think we scientists have to do with, if we are the second oldest profession, is to not become like the oldest profession and become intellectual prostitutes. I hate to say this, but like if you were to go to a, a, public health conference and there's a speaker up there talking about what public policy one should have tobacco regulation and stuff. And then you realize halfway through the talk that this guy is funded by Philip Morris and nobody told you and he didn't disclose it, right? And then you walk out after this talk into the expo area and you see all these booths from all the tobacco companies.

Wouldn't that feel that would feel like really weird. You would really feel like you are the oldest profession. That's completely taboo. You could never get away with doing that in a public health conference. Yet, you know, I was just at ICML in Vienna, the second biggest AI conference. I was at Nureps last year. It's exactly like that. You have all these speakers, sometimes even speaking about social impacts of AI, and they don't mention all the funding they get from big tech. And then in the coffee break, you go out to the expo area and all the tech companies have their booths there. There was an amazing study recently. You can find it on the archive, Abdallah and Abdallah are the authors, where they found that MIT is one of the worst.

So many people even who say they work on AI ethics are taking all this funding from Facebook and companies like this. And, you know, that is kind of intellectual prostitution. Upton Sinclair, I think, was spot on when he said it's hard to make someone understand something when their paycheck depends on them, not understanding it. And one more thing. Also, just ask yourself, Whenever you bring a new powerful technology into the world, how is it actually going to be used? Do you know Tom Lier? Yeah, the songwriter. He's like the master of dark humor.

He has a song about Werner von Braun. Yes. He goes up, once the rockets go up, who cares where they come down? That's not my department, says Werner von Braun. That to me just epitomizes. what some scientists are doing now in AI. You're just going to build these, we're going to try to build super intelligence

and get super rich or whatever. And figuring out what to do with it, that's not my department. You have one of your kind of renowned articles in Time magazine from a couple of years ago. You talk about the things we shouldn't do. And I don't think at this time, everything was done,

but you say don't teach it to code, don't connect it to the internet, don't give it public API, Don't start in arms. All those things have been done. What do you say to somebody like Jan, respectfully, obviously? He's coming on. People talk about the maximalist.

You mentioned the quote from Upton Sinclair. He's the chief AI scientist at Meta. He's obviously got a huge vested interest, although he is pivoting more, according to him to research at NYU and going back to his roots in computer science. But if you were Mark Zuckerberg at this point, and you really did, you weren't mendacious. And you cared about humanity's future.

What would you tell him to tell Jan or what should Jan be telling Mark? I should say this as a proud owner of a pair of meta, Rayban spec, which I actually think is the best AI product that I've used, including chat GPT, because it has some virtual aspects of it. It's actually doing something useful all the time. But what would you tell Jan or tell Mark Zuckerberg? The reason I try to never make an hominem attacks is not only, because stooping very low, but it's also very important because I actually believe that all these people think they're doing the right thing

and what they do comes from a good place. It might surprise people who only see me having Twitter spats with Jan. But whenever we meet in person, we get along great. When we did the monk debate, I'm actually the one who invited him. And he said, yeah, I like to debate with you. But I do think it's a little bit hard for him to probably subconsciously to separate this massive conflict of interest that he has, Especially when I speak with the CEOs of these companies, I really empathize because I think they all in their heart want to do the right thing.

I don't think any of them, except maybe one, just wants to build a successor species that replaces humanity. They would like to take it as carefully as they can without being outcompeted by their competitors. But they're all trapped in this race to the bottom. and they all feel it, where if one of them were to be like, hey, I'm just going to pause this thing for two years until we can figure out and make it safe, they will ruin their own company.

And before the two years are up, they will be replaced the CEO by the shareholders, most likely. It's like if you're a tobacco executive and you decide one morning that, you know, my best friend just died of lung cancer, this is not right what I'm doing. Mike, now I'm going to Philip Morris from now on is going to do something else.

They would just get replaced. And they, so the only way to fix this is, for example, the U.S. government to step in and say, hey, you know, here are the safety standards. They apply to all companies. Now, things are much better for the CEOs because they, they don't have to be the bad guy. They can redirect corporate efforts to figuring out how to meet the safety standards so they can make money. We talked earlier about how we want incentives in the world to bring out the best in us, right? that's what I'm talking about here. You know, before we had something,

before we had the FDA in the U.S., you know, there was this drug called thalidomide. Are you familiar with it? Yeah, it was given. Chiral asymmetry that led to birth defects. Yeah, they marketed to pregnant women and said, this is great if you have headaches and whatever during pregnancy.

Born in sickness, yeah. Then kids were often born without arms. Things like this prompted the creation of the food and drug administration exactly for the same reason, that you want the incentives to be such that here are the safety standards for all companies, regardless of how ethical or not the CEO is. And whoever meets the safety standards first, they get to make a lot of money first.

So biotech companies have exactly the right incentives there to make their things safe as fast as possible. They don't have to write a pause letter written by scientists in biotexting. We should pause development of thalidomide or new drugs. of questionable safety, right, until someone could figure out that it's safe. No, it's not the politicians' this problem. It's the company's this problem. If someone starts selling a new wonder drug that's supposed to cure all cancers, the FDA is going to be like, wait a second, where is your clinical trial?

Oh, you haven't done one yet? We'll come back when you have, you know. That's the solution to all of this. We need the U.S. to tell that the American companies, here are the safety standards. and the Chinese company government to say, here are the Chinese safety standards, the European. This is how it happened with drugs, right?

The Chinese also have an FDA, not to make the Americans happy, but just to prevent Chinese consumers from getting harmed by Chinese companies. And then once you have these safety standards that have independently been formulated in different countries around the world, you almost automatically start to get the international corporation, You know, the American FDA and the European EMA in Europe, they go for drinks, they have meetings to harmonize their standards so that an American company approved for a drug here can more easily get approved in Europe, et cetera. The rest sort of takes care of itself.

I can hear my audience screaming out, you know, the conspiracy theorist in the audience that, yeah, sometimes they're a little too cozy when we fund gain of function research at the Wuhan Institute of Irology. and recently came out, they actually had a polio patient potentially, that also there. But I'm not going to defend the regulatory capture problems, which we obviously have here, which are bad. But I have never met anyone in the U.S. who works in biotech who has called for just completely getting rid of the FDA and going back to letting anyone sell telenamide. What we should obviously do is get rid of the revolving doors and prevent the regulatory capture. Getting back to the most important subject of all, me and you, how will AI impact? I mean, I love it. I have to say, part of the reason I'm an optimist is because I really enjoy it. It's like, it's like I have you with me 24-7. I don't have to sleep on your couch anymore. But, you know, I can ask it anything. It'll even make slides for me. I've written scripts. I never learned Python to a level that my graduate students, you know, could do in their first year of our undergraduate. But I didn't have to. It's like when we were kids, they said learn Mandarin. And it'll be really important. Or learn Japanese.

When I was a kid, it was learned Japanese. I never learned it. I'm glad I didn't. I'm glad I dedicated that to learning, you know, cosmology and astrophysics. But how is it going to impact our profession, the oldest, second oldest profession, being a professor, teaching this job that we call an avocation, a love? Do you see it? Do you see it existing in this form as you and I practice it? That is a question I really struggle with, particularly this one.

week, you know, last week was the first week of two classes again, because I feel MIT is totally in denial about where this is all going. I bet you mean how education is going or something else like protests or things like that? What are they in denial about? Ultimately, all aspects of life. But even what is the, what's the value even of having a university if AI can do all this stuff better? What is it that's valuable for me to teach the students now?

So many people are so excited now by using LLM-powered AI tools, the teachers. teach students better than we can do it. But once the AI can teach the students, obviously the students are not going to get paid to do a job that the AI can do good enough to teach it to them either. And I think I'm not going to give you a glib answer to this because this is really tough. But what I do want to say is, I think,

asking the question this way, what will happen? Ambition comes in all shapes and sizes. At First Citizens Bank, we roll with your goals because we're built for what you're building. Fit for your ambition for Citizens Bank. It's really the wrong way to ask it. It sort of presupposes that you and I are just passive bystanders here sitting and the couch eating popcorn waiting for the future to happen to us. Now, you and I and all of the people watching this podcast, we're building.

this future. So the right question to ask is, what do we want to happen? What kind of future are we excited for us and our descendants to live in? That's where it really needs to start, a shared positive vision. And then we can ask, okay, what does that mean about how we should and shouldn't deploy our technology? People often say things like, oh, it's so inevitable that all the jobs are going to go away because AI will be able to do them more cheaply or whatever. None of these things are inevitable. There are many things that we can do with technology that we've chosen not to do. Human cloning, for example, it could be very profitable.

If I could make a lot of clones of you, Brian, I could make a lot of money. Yeah, you know, you have. We don't want to mess too much with the foundations of our species. It's not worth the risk. and we can have a lot of fun anyway. If there are activities that we humans find very meaningful and to give us a lot of joy and purpose, we don't have to stop doing them

just because there are machines that can do them. If you like playing tennis, you wouldn't replace yourself by a tennis-playing robot. I might need a surf robot because I hurt my knee on Friday. If we ask ourselves what kind of future we want to live in and we want our kids to live in, obviously the answers we can come up with It can be much more exciting if we also allow ourselves to have a lot of high tech in there.

We were promised AI and robots that we're going to do all the boring chores, clean up the kitchen, etc. While we could do the art and write beautiful essays and stuff, instead, now we have AI that can do art and essays, and I'm still stuck cleaning up the kitchen myself. Hey there, fellow Voyagers into the Impossible Tizai, your fearful host. Professor Brian Keating here with a tiny little homework assignment before we get back to the episode. And that's to make sure that you're subscribed to the podcast, either following it or subscribing to it depending on your podcast, catcher of choice. I did some research of my own and found out that only about half of you are actually following

or subscribing to the podcast. So please do that. And for some extra credit, if you're looking to boost your position on the grading curve, please leave a rating or review. It really helps us out tremendously. Do it. Do it now. Before you forget, let's go back to the episode. Your youngest child's too young to do the dishes, but you know from a past experience, you can get the older ones to do autonomy, autonomous tasks, except driving.

I'm not ready to make that leap of faith yet. I would say, no offense, but you were one of the greatest cosmologists of your generation. You made original contributions, predictions, You aided and abetted scoundrels like me and experiment predictions ranging from first application of a very abstract logic and statistical proofs in computer science, bringing that to cosmology in the early 1990s, to deep treatises on math and foundations of quantum mechanics in the late 1990s. You were one of the first with Wayne Hugh and Dan Eisenstein to predict the effects of barrier on acoustic oscillations, which has become a whole new tool akin to some of the things that I do using the CMB. Do you miss it? Do you miss being a day-to-day cosmologist? Or, you know, because you gave up, as I see it. Again, I'm offensive.

I don't mean it, Max. I'm just saying you were at that level. That's very rarefied. And now you're doing something extremely important. Is that a sign that I should be more worried? Because when someone like you pivotes a whole career that's, you know, very, very high level to a completely different direction starting from scratch, should I be worried? Wengretzky famously said that don't, you shouldn't skate to where the puck is, but to where it's going. And I did, I did feel when I was doing my PhD that it was quite obvious that the puck then was going towards these huge C.

experiments coming online, micro background experiments, including the ones you worked on, large galaxy surveys, et cetera. So I decided to skate in that direction, and it was absolutely fascinating to work with you and so many others on harvesting all the gold that was in this data. But then, as you know, there were these three big mysteries. Did inflation happen? Still haven't found the gravitational waves. What is dark energy? We still have not found any evidence that W is not just equal to minus 1 and it's not just boring constant. And what is dark matter?

Do we keep pushing down the experimental limits? We still haven't found anything. So I started this feel that the golden age of cosmological data was waning at least a little bit. There's plenty more cool stuff to do, but still coming off the peak a little bit. Well, at the same time, I felt the puck was very much going towards AI. It was so obvious.

this field was exploding. On top of that, I'm very much just driven by curiosity, and it's just so much fun to work on understanding how intelligence works and the mind and so on. And my group at MIT now, we've mainly focused on what's called mechanistic interpretability, which, in less nerdy speak, you could think it was just artificial neuroscience. You take an AI system that's doing something really smart and clever and interesting, and you try to figure out how it actually works under the hood.

So we had a big conference about on that they're organized here at MIT last year. And then I was part of organizing one in Vienna this summer. And it's crazy how fast this is progressing. It reminds me a lot of the heyday of the microwave background from one year to the next. Oh, WMAP came out, you know, everything revolutionized. The reason for that is quite obvious. If you're doing neuroscience on an actual human brain, it's painstakingly difficult to measure the output from even 1,000 neurons and you have

a hundred billion of them. And you also have to get ethics board permission to even do the experiments. Whereas when you're studying an artificial intelligence, you can measure every single neuron, every single synapse, like all the time. And you don't even have to ask permission. So this is also just very much a fun field. So the emotional, personal answer here is, yes, of course, I think back fondly about the good old days when you and I were doing micro background papers together. But it's also just an enormous amount of fun working on this question.

How can we understand the inner universe of intelligence how it works? It's every bit as much fun as a lot of fun. as it was back then with the cosmology. So I've got one question from the audience that I want to ask you, and then I've got four quick response answer that I hope will be rapid fire. So the first and maybe the only question that I want to ask from the audience has to do with the past guest, Nobel laureate, Roger Penrose, and Stuart Hammeroff, both of them multi-guests on the podcast.

And it relates to, you know, something that is obviously very interesting to you, which is artificial intelligence and also the human mind and mathematics. And, of course, Sir Rogers' book, The Emperor's New Mind was one of the first, if I think it was the first science book I ever read as a high schooler, didn't understand it. But he made a very powerful case that the brain really isn't a computer. And yet, you know, to this very day, and there were certain problems that weren't computable and it has to do with Turing tests. It's very interesting and still quite relevant book. But you're very convinced that that approach is wrong that the microtubule orchestrated collapse or without attacking him personally, you've been very skeptical about that. Can you comment on why you don't believe that's the correct approach or, in your opinion,

it's inferior to other ways of getting at this question of the ghost in the machine, so to speak? I'm a scientist. So my job is not to believe stuff or not believe. leave stuff. It's to look at the data and make educated guesses, happy to place bets. I would have loved for the brain to be a quantum computer because it's much cooler than a classical computer. But back when I was a postdoc in Princeton, I decided to just calculate how long quantum superpositions would survive in a warm, wet brain. And disappointingly, they would last way less long than they would need. to. If a neuron is firing and not firing in a quantum superposition that gets destroyed in 10 to the

minus 20 seconds, you know, I don't know if Roger Pendros can think 10 to the 20 thoughts per second, right? I certainly can't. And so I just, I wrote that nerd paper. Then some journalists put a little bit of a mean spin on it and those guys, I think, got a little bit annoyed. But I was really not going into this with an agenda at all. I think what's happened since, unfortunately vindicated my conclusion. that you don't need quantum computation for human level intelligence, because chat GPT, 4-0, and other large language models are able to do a lot of the things already that we humans do, demonstrably without any quantum computing. They're purely classical, right?

So I think what this is telling us is that even though quantum computing is very cool, it is possible to build very intelligent machines also classically. Yes, and I guess that kind of dovetails when it makes me think of the computational aspects of it. When I talked to Sean Carroll, who's a big proponent of the mathematical universe number three multiverse, which is an ever-ready in many worlds theory. And I asked him, I want to ask you the same question. How could we experimentally test branching rates and so forth? I mean, are these things just completely out of the realm? I mean, our colleagues at NIST and other places are measuring, you know, kind of at a second level, making clocks that operate at a second level.

What level of technology would be needed to get falsifiable or even evidence for, you know, the Everettian branching view of reality in that level three multiverse? That's an easy one. Basically, what we should do is just try to build quantum computers that can do things, do computations, that no classical computer could do even if it were the size of our universe, right? That would totally demolish the idea that all the resources we have access to somehow are sort of limited to our classical universe. You know, people argue now about whether we already have demonstrated quantum supremacy or not. I'm going to stay out of that mudslinging contest.

But I applaud experimentalists working hard to really try to do this. And if you could, if you can build a computer, which can actually run chat GPT 4-0, for example, in superposition, in quantum superposition. And suppose you can build a version that actually will talk to you and describe how it consciously experiences things, right? If you have someone as smart as you, who is demonstrably having two different experiences

in the level three multiverse at the same time, and how much more evidence do you want than that, than being able to talk to it and so on, We've already put atoms in two places and ones, molecules, the Carbon 60 Bucky Ball, Nurgis Mavala, and her colleagues at Lago even took a mirror that weighs one kilo, and we're able to demonstrate that it behaves quantum mechanics. Yamava Resort and Casino at San Manuel is California's number one entertainment destination for today's superstars. Catch the Jonas Brothers return to the Yamava Theater stage on April 30th,

the powerful vocals of Demi Lovato on May 17th, and the signature Southern Country Rock of Eric Church on July 19th. Tickets on sale now at Yamava Theater.com. Only at Yamava Resort and Casino, celebrating its 40th anniversary. U.N. must be 21 to enter. Once we can get to the point that something that is intelligent as a human brain is able to do two different things in quantum superposition, I think that kind of clinches it because that's what the whole point is of U. Everett and his work. So we won't need to look at microtubules anymore.

That'll make things a little bit easy in the wet. I think people should look at microtubules. And like I said in the beginning, if there's ever any cool experiment that can be done at a reasonable cost that will teach you something, we should totally go ahead and do it. Yeah. This is a part of the show I call the fantastic final four. We're basically existential questions that I haven't asked you in our past conversations.

And the first one relates to your countryman, Alfred Nobel. who created a Nobel Prize, which had a twofold component. It gave away money and encouraged people to do stuff in science, peace, literature, et cetera, and medicine. And I've had on 21 Nobel Prize winners, Max, in this short time. Maybe you'll be the 22nd someday, we'll see. But Alfred also said that a winner had to do his or her work for the betterment of mankind. And that's an example of what's called an ethical will, not just a monetary or material will. In Hebrew, we call it a Zava'a.

There's sort of things of great wisdom that you want to leave to your descendants, not just biological, but maybe ideological like me and all those that come after. Don't ask you, Max, if you were to leave as a document or some sort of will for the future of humanity, what would you put in it? Or what would it be about? I would say that we should go forth and not only build great technology, but also create incentive structures. They bring out the best in us humans so that we use it wisely. Otherwise, we'll be like Werner von Braun again. Once the rockets go up, who cares where they come down? That's not my department, says Werner von Braun.

Well, speaking of technology and things that go up and go down, as you know, I'm the associate director of the Arthur C. Clark Center for Human Imagination at UC San Diego. You probably know the movie 2001 of Space Odyssey, which has not a small amount of artificial intelligence. In fact, the sign in the back over there, if you can see it, do you know that the word podcast comes from that movie, Max? Did you know that? I'm sorry, Brian.

I can't do that. I have a computer in this room. Actually, I just said a trigger word. I've named it how, but I could say, how computer turn off the plug. Now, if you saw that, I actually turned off my magical neon sign. But my ultimate Keating test, number two maybe, is can you program, in AI that will cause itself pain, that will turn itself off or blow its own capacitors or

things like that. But I don't want to get into all that. What I want to say now is you remember in the movie there were these monoliths, right? There were these structures and the apes are hitting it with a bone and then they're found on moons of the planets of the solar system. I want to ask you a different question. I want to say, if you had a time capsule that could last for a billion years like these monoliths placed by a sentinel species, What would you put on it or in it? What would you put on a time capsule that you knew was guaranteed to last for billions of years? I would just put all the books and movies we've ever made in there and then let our descendants figure out what they were interested in, I think.

Now here's another question. This one comes from Richard Feynman, Nobel Prize winner Richard Feynman. He said, if there was some cataclysm, maybe caused by AI, and only one sentence could be passed on to the next generation of creatures, what sentence would you put? on there. You know he put on the atomic hypothesis. What would you put on? What would be Max's sort of a time capsule of a scientific variety? You're going to be disappointed that I'm going to pick exactly Feynman sentence here. The idea that everything is made of these tiny particles bouncing around called atoms, et cetera, et cetera, because it was actually when I read that sentence in volume one of the Feynman lectures, that it finally clicked to me. And that's what made me

really fall in love with physics, actually. What I previously thought was the most boring subject in high school. So it's really because of this that I'm here. Yeah, for me, it would be the CMV antisotropy power spectrum. Because that actually has within it the atomic physics hypothesis, you know, because it comes from hydrogen. So it's a little more economical, Max, but that's my choice. Okay, another saying by Arthur C. Clark, when a distinguished but elderly scientist,

I'm not calling you elderly, but when a distinguished, but elderly scientist says that something is possible, they're most certainly right. When they say something is impossible, they are very probably wrong. And Arthur called these things failures of imagination. Nowadays, we call these limiting beliefs. I want to ask you, what have you been wrong about? What have you changed your mind about in science or outside of science, if anything? I was thinking the other day about what I was wrong about in my second book, Life 3.0. And there were actually two things. First, I was wrong thinking that. it would take much longer than it actually did to get so close to artificial general intelligence.

So it actually turned out to be easier than I had thought. The second thing I was wrong about was I never in my wildest dreams when I wrote that book thought that world leaders would let companies get so close to taking over the world without doing anything at all. That's why I opened that book with a story about how some people take over the world entirely in secrecy. And then, Max, my final question. You've been so gracious with your time. It's so-called third law of Arthur C. Clark, who said the only way of discovering the

possible is to venture a little way past them into the impossible. I want to ask you, Max, if you could go back, meet 20-year-old Max when he used to be called Mad Max? What would you tell him? You've got 30 seconds with 20-year-old Max. What would you tell him to give him the courage to do as you've done to go, to the impossible. Never underestimate what he might be able to actually do. And if he has an idea that he really believes is correct and everyone around tells him that it's bullshit or impossible,

keep pursuing it anyway. Let the people talk. Max Tagmark, my good old friend, it's great to see you. I wish we could be in person, but maybe we will be in the near future. Some follow the noise. Bloomberg follows the money. Because behind every headline, is a bottom line. Whether it's the funds fueling AI or crypto's trillion dollar swings, there's a money side to every story. And when you see the money side, you understand what others miss. Get the money side of the story. Subscribe now at Bloomberg.com. You can't reason with the sun. Trust us. We've tried. This summer, it's time to put that angry ball of fire on mute. Columbia's Omnyshade technology is engineered to protect you from the sun's

harsh rays that can burn and damage your skin. The sun is relentless, but so is our gear. Level up your summer at Columbia.com to spend more time outside and less time slathering on allotion. You're welcome. Columbia, engineered for whatever.