Interstellar Visitor 3I/Atlas, Human Longevity Plateau, New No-Sort Plastic & Analog AI

0:00Hello internet. This is your captain speaking Lester Narre joined as always by my co-host and our resident PhD Krishna Chowdery. This is from first principles. We have some great stories for you lined up this week. Starting with an update on the interstellar visitor 3i Atlas as we've seen four new instruments capture images as it makes its way towards Earth. Followed up by a story about longevity. If you were born after 1939, you may not live to see a 100 years of age. We'll talk about that in Brian Johnson, aka the man who wants to live forever. Our third story is about a new no sort plastic that will allow us to recycle more efficiently.

0:41And we will end today with an interesting analog AI story coming out of Microsoft. This is FFP. Let's get after it. [Music] my friend. >> How's it going? >> We're back again. >> Yeah, >> another one. I think we uh we have a little something here. People seem to be excited about talking about science. >> Yes, dude. >> In the I'm really enjoying seeing all the feedback that we're getting. >> 100%. >> Yeah. We're going to start with a very popular science story that has been all

1:22over the headlines. >> That's right. >> Um about 3i Atlas. >> The headline this week, four powerful telescopes agree interstellar comet 3i Atlas really is bizarre. This is from a variety of outlets but from space.com is where we sorted. And the new information here is both NASA and issa, the sort of North American and European space agencies, have used their instruments, Hubble, Sphere X, JWST, and TESS to capture images of the object as it makes its way towards our sun. Yeah. So, what's new here? So, what's new here is um we've we've actually gotten really nice data from the James Web Space Telescope, from Hubble, and from the VLT, and from a lot of these things, and

2:03they're converging on trying to identify really what this object is actually like, okay, and some of the, you know, nitty-gritty characteristics. So, just to recap our viewers about what ThreeI Atlas is, right? It was discovered on the 1st of July of this year by the Atlas program which is the asteroid terrestrial impact last alert system. Okay. And this is a system that is part of University of Hawaii's like institute of astronomy. There's basically a bunch of telescopes all around the globe. There's like two in Hawaii. There's one on Maui, one on the big island. Then there's one in South

2:44Africa, Chile, um, and Spain in Tenneref in the Canary Islands. This particular one was discovered in the one in Chile. Okay. >> Okay. >> The whole point of this guy is just to peruse the sky looking for um, transient objects and objects that move around. Okay. So, it was discovered there and then it turns out that, you know, other telescopes had already seen it but hadn't identified it as this thing. We had talked about this Viki transient facility in Palomar Mountain near San Diego, right? that thing actually had already found it in May but hadn't like pinpointed it as like a thingy that we should be worried about. Yep. >> So, um once we identified it with the Atlas program, >> everybody all over the globe was like,

3:25"Oh, let's just pruise the data cuz we know where it is, right? And where it would have been." And so, from that, we could then um get the trajectory of this thing. >> Yes. >> And once we calculated that trajectory, >> it was like, "Oh, whoa, whoa, whoa, whoa, whoa, whoa, whoa." Okay. It's moving really fast. 61 km/s >> um all the way out in Jupiter. Yep. >> Okay. And if it's moving that fast all the way out there, that means it's not going to stay inside the solar system. You can do the Newton's laws and do the calculation. It'll go right past the sun and then on on its way. >> Mhm. >> Then we got into these things about people had these really blurry images. They're like, "Is it aliens? It's never aliens." Um >> much to my sugar. >> Yeah. Yeah. Yeah. We we always hope,

4:06right? One can always hope, but um time always tells that it's not aliens. And um it turns out to be the third interstellar object, right? Which is um an object that is not part of our solar system that is literally going from star system to star system >> around the galaxy of the Milky Way, right? >> Because the idea is there are these large gaps of void space in between, you know, star systems. And so it's it it's from a probabilities perspective to because of the vastness of space. >> Yeah. >> And our ability to detect Yeah. both of those reasons. We haven't seen a lot of these. >> Yeah, we haven't seen a lot. Yeah, we haven't seen a lot. Now we will be able to because of things like the Vera Rubin coming online and things like that,

4:47right? Um this is our third one. That's why it's called three I. The I is actually for interstellar and threei Atlas. It's because Atlas was the program that discovered it, right? >> Um the first one is one I >> am uam Mua. Um that was the really weird elongated thingy. Um and that one we caught on the way out, so we couldn't really get good data on it. Um the second one was 2i Borisov which turned out to be an interstellar comet. Um and then now this was threei Atlas. And so a key difference between AmuA Mua let's say and Atlas >> is that we are we were able to catch Atlas on its way in >> not when it was on its way out which gives more time to capture more data.

5:28Yeah. >> Uh as well as being able to see how it behaves over a longer time period. Is that fair? >> Yeah. Exactly. It's like that scene from Office when I was like don't panic. I said stay calm like okay now it's like it's happening right so now everyone's like scrambling to figure out how to image this thing because once we've got it like every astronomer who's a planetary astronomer who is uh you know or planetary origins or an exoplanets astronomer trying to understand how solar systems form this is a very big deal okay because the other thing is if you turn back the clock this thing came from above the milky way which is where there's an old population of stars that are much older than our sun right so this thing could be something like 7 billion years old. Our solar system is

6:08tops 4.6 4.7. >> So this thing is older than, you know, anything that we've really had a close encounter with. >> Now the um >> the the problem is when this thing gets close to us, we're going to be on the other side of the sun, >> right? So it's coming in, we're on, you know, like the front side of the sun right now and it's coming in towards us, but we're going to be orbiting and we'll be here. It'll still be coming this direction and we'll be blocked by the sun. >> By the sun, right? It's like during like if you wanted to see it, it would be during daytime behind the sun, >> right? Which because of the atmosphere, >> we're not going to see anything, right? The sky is going to be blue and the air glows more than any stars or >> this three eye atlas.

6:49>> This is why Ver Rubin is doing a survey sky survey at night. >> Yeah. >> And not during the day. >> Yes. Exactly. And and furthermore, we can't really point the James Web Space Telescope towards it either because like then the sun the whole point of the James Web telescope, the reason why it's so nice and like sensitive is because it's it's it's got this giant shade >> that shades it from the sun so that the instruments can be kept at like 4 Kelvin, right? like colder than like actually no even less than like there's like a literal like dilution refrigerator on that thing that is like keeping it colder than outer space because then you get these instruments that are super cold. The instruments the atoms aren't jiggling. So then the only

7:31signal you're getting is not from the instrument itself but from like the starlight. >> There's no there's less or no noise. >> Yeah. Yes. But if if you wanted to like point it towards that thing like that defeats the whole purpose, right? So you can't really do that. Um, so what we what we'd like to do is to observe it right now. >> Yes. >> With our terrestrialbased telescopes like the James Web right now when it's like, you know, sort of away from the sun. Get as much data as possible. >> Yes. >> And then as it goes through the solar system and it reaches Mars. >> Yes. >> It's going to be on the other side of the sun. It's going to be near Mars' orbit. >> Yes. >> Um, we can have all of our orbiters that are orbiting Mars, >> right? >> Sort of. and then point towards this

8:13thing and try to get as much data as possible. >> So, the idea is we have orbiters around Mars that have a certain mission right now, right? A Mars-based mission. >> Um, and >> they can take a break. >> This is one of those like not that many that often in a lifetime opportunity. And it's like we can take a a couple hours a day, reorient, use some of our fuel to do so to capture to provide that data when we on Earth as well as JWST don't have the capability or the angle of of viewing to be able to actually properly capture it as it gets closer uh to our our orbit. >> Yeah. And exactly. And with comets, you really want to look at the thing when it's close to the sun.

8:54>> Okay. >> Cuz that's when it's doing cool, crazy stuff, right? Okay, when the comet is far away, it's just a chunk of ice. >> But as it gets closer to the sun, it's going to start interacting with the solar wind. It's going to start interacting with the photons that the sun is giving out. So then there's going to be chemical reactions and physical reactions happening on top of the comet. And that's what we want to see, right? In the very like way way out there when when we tried to look at three Atlas, right? It was like really bright for how far away it was. And that's what gave rise to a lot of these alternative theories about like it's like artificial and all this. >> The light's not reflecting from the sun. It's generating its own light. >> It's generating its own light cuz it's so bright. Like how's that possible?

9:36Now, as it got closer, actually, so August 6th is when James Web looked at it. And the preprint came out on August 25th. They were fast about this because they know that like people want it, right? Yeah. >> So, um the preprint came out like literally they they had like what like 20 days to do all of the >> people were upset about how long it took. >> Yeah. Yeah. But dude, like 20 days to like massage data is is kind of and like create a story, right? It's not just like you're analyzing the data. You have to like interpret the data, try to create a story about like what you're trying to say, right? Science is all about stories at the end of the day. It's not just numbers. You have to have some kind of narrative. So they they um they analyzed the data from James Webb

10:16Space Telescope. It used um one of the instruments that is a spectrograph. So it can tell us things like the composition of this thing. Yep. >> And what they found was a really high mixing ratio for carbon dioxide and water. >> Okay. So usually the mixing ratio is quite low. Here it's 8:1. >> Okay. >> Okay. There's eight times as much CO2 as water. That's like a lot. Okay. Okay. And that's very unusual for comets in our solar system. Correct. Okay. So that means that this thing there there's two possibilities. Okay. One is that when this thing was born >> Mhm. >> it had a very different birth environment than the comets in our solar system. It could be that this thing was

10:58born in something called the CO2 ice line. Okay. You can imagine when the a star is forming and the solar system is forming around it, the star starts >> making um light and thermal energy that starts going out, right? There's going to be the the disc is going to have a mixture of, you know, silicon, hydrogen, um CO2, water, and all this other all the dust from that primordial nebula that sort of became the star, right? So, all of this stuff is going to be in the disc. um there's going to be a line where inside that line close to the star there's going to be enough radiation to um put CO2 in a gaseous form. Okay.

11:38>> Okay. But afterwards there's not going to be enough radiation, not going to be enough heat >> and that CO2 is going to become dry ice. >> Okay. So there's going to be this ice line where after a certain distance away from the star you just have dry ice and before you have gaseous ice >> because also >> gaseous CO2. Carbon dioxide does not have it does there's no liquid form. Yeah. Go straight from gas to solid. >> Yeah. Yeah. I mean there there is a liquid form but at like very peculiar pressures, right? Like if like in the in the low pressure of the of of the of space and even at like one atmosphere there's like at one atmosphere it'll go directly from dry ice to CO2, right? If you put in high enough pressure you can liquefy CO2, right? But like out here

12:20it's just going to go ice to ice to gases and right back and forth. So, um, you have this like dry ice line. Now, it could be that like the comet was forming in this right past that ice line, right? So, it's like gathering up a lot of CO2. Not a lot is a lot of water, right? Because water, the water line is elsewhere, >> right? >> Right. >> Right. >> Um, so that could be one reason why 3i Atlas has a lot of CO2 versus water. The other thing could be that it formed or it was um subjected to a really harsh radiative environment. Okay. >> Okay. And then what can happen is you can have you can have these reactions

13:01between carbon monoxide which is also fairly common um carbon monoxide and water with this high energy radiation to create CO2. Got it. And that's what happened. Right. Got it. So there's these two possible scenarios, but in any case, it's still very very different from any comment that we've seen, which is I think in what's interesting is there there's this leap, you know, that folks like me who are not as educated always make where it's like, oh, we have no reference point, so that must mean it's aliens. And I think kind of what the through line you're drawing here is we do know that it was birthed in an environment that is different than the environment where we have these data

13:41sets for >> and so those initial conditions matter. >> Yeah. and and the and the data sets to your point the data sets that we have >> on what is normal right >> is like from our solar system as you say right there's millions and millions of solar systems out there with different types of stars different galactic environments like we are we are a tiny little sample size >> right and one of the great things and why everyone is so excited about interstellar objects is because it gives us a sample a sampling of what's out there right and so That's why everyone is like scrambling to try to get as much data as possible on this guy. Right. >> Right. >> Because from that data, like we kind of

14:22talked about at the beginning of the story, >> there are multiple insights that can be gleaned not only at time of capture. >> Mhm. >> But also afterwards when we learn other things and it's like, oh, let's go look back at the Atlas data. Yeah. Now that we have some different context about how these highly rated highly radative environments work. Yeah. Or any number of other Yeah. options. >> Yeah. Exactly. We can like start I mean you know we we're we're kind of scrambling cuz like this thing's going to be out by September, right? So right now we're just in the observe phase. Okay. After this thing gets out and we get all the data, then I'm pretty sure there's going to be a big cohort of people that are going to go into the

15:02modeling phase, right, of trying to understand, okay, theoretically, how is something like this possible? Like, if I were to model it this way, would it match the data? No. Well, then that's not that's probably not what what happened. And and so, yeah, at this point, it's just like everyone's just like, just point it at the thing and get as much light as you can. What what's what's so funny about this is um as I mentioned in a previous episode, I get the pleasure of being able to work with um Dr. Avi Loe on a regular basis who uh as a professor of astrophysics at Harvard. And >> you know, one of the things we had reached out to him to do was to do a white paper >> uh on this new update from 3II Atlas. And to go with what you're just saying, his response was almost a a more uh

15:46expanded version of what you just said is which is like we need to wait for the dust to settle. >> Uh right now it's just an observational stage. We need to focus on capturing as much data and information as possible. Once it's gone >> Yeah. >> and that opportunity has left, then there's going to be time to actually settle in and and do analysis on what it is that we just captured. So I think there's an important distinction to identify here which is like even if it was aliens right we don't have any real baseline to reference for it to identify it as such and if it doesn't land on the White House lawn there's going to be this two-stage data capture and observation which is where we are now >> and then sort of analysis and modeling

16:26which happens after >> and we we will get more clarity generally speaking >> once the observation period is over and we move into this analysis and modeling time frame which is again how science works. Yeah, exactly. >> Um, but but that that's there's there's a there's a big public desire for answers, which is understandable. Um, and like many things, it's just a process. So, it's it's interesting you say that cuz it makes me think of that that letter we got. >> And like, and you know, you know, it's it's it's definitely not aliens at this point, but it's still like one of the coolest objects that's ever visited us, right? And so I mean I don't think people should be disappointed, right?

17:08It's kind of like a copout. I mean, you know, as the Buddhists say, like don't have these expectations because the world itself is a wonderful place, right? Right. Imagine imagine this thing like going through the Milky Way for 7 billion years, visiting like who knows what, right? This isn't we're not the first star system that it's seen. Imagine from its point of view. It's gone from one star system to the next. The fact that it's going so fast means that it's had a pretty close encounter with something pretty massive to like be zooming through like this, right? The sun is not really going to change >> its velocity that much, right? It's just it's sort of like redirecting it in one

17:50direction. But like this thing had a very close encounter like this, you know, like imagine from its perspective. It's been around for like half the age of the universe, >> right? That's a long time. We've only been around for a third. So, I think I think I think it's um an incredible like story, right? And and to get back to the idea of we're trying to observe this thing as much as possible, right? So, that's where the European Space Agency and this new stuff is coming in, right? Where the the European Space Agency has a bunch of orbiters >> around uh Mars that they're like, "Okay, we're going to we're going to repurpose for this." One of them is um the Mars Express and the other one is the Exom Mars Trace Gas Orbiter. Both of these

18:31both of these things have cameras that are pretty good. This thing is still going to be pretty far away from Mars, but you know, getting any sort of getting anything when it's on the other side of the sun and we can't put anything there and it's going to be the most volatile when it's on the other side of the sun. So, getting any data at that point of its orbit of its trajectory I should say, not orbit cuz it's not orbiting, um >> is going to be incredibly valuable. And the guys at the ESA are definitely tempering everyone's expectations. It's like, guys, it's going to be far away. Like, our stuff wasn't designed for this. It's like, okay, dude, like no one's going to blame you if you get data. Like, just get whatever you can, right? And then the other thing is, so we also talked about the Europa Clipper

19:12in one of the previous episodes. So, on its way back to Earth, >> yes. >> Um, it's actually going to cross through where the comet was, >> right? So, so as a comet goes across the solar system, >> it leaves behind a a trail of of particles of debris in its wake. That's actually how we get meteor showers. Meteor showers are literally just the earth moving through cometary >> um paths, right? That have these like particles like and just like just like planets, the particles are going to be going through the the sun like they're they're orbiting the sun as well, right? So like as the earth goes through these these cometary paths any sort of

19:54particles that get trapped in the earth's gravity fall down to earth and then that's how you get the perciads and all of these meteor showers right but you can now imagine um the Europa clipper probe is going to be going through this thing right so we could be getting observations about >> the particulate matter matter that was a part of >> Yeah yeah yeah and it's not just new Europa clipper I think there's other places like Hira and Lucy They may fly through the commentary tail. They're all NASA. >> Yep. >> So, that's going to be super exciting to get like actual like, >> you know, data. >> It's going to be both optical like imagery capture as well as like physical matter analysis. And and I think this sort of, you know, this this brings up

20:35the when we build these really expensive, highly precise instruments for observation and data capture, >> they're not necessarily single purpose >> um in this exact context, which is like we send a bunch of stuff to look at Mars, >> but we can still use those same instruments for other use cases such as, hey, go just look at this thing that's coming through and let's see what we can get. >> Yeah. And Yeah. It's pretty crazy because like I'm pretty sure when the ESA and when NASA like made these things, we didn't, you know, I uh AU MUA wasn't even a thing. Right. >> Right. They when they were planning like they were made well before >> well before we even and then now it's like now it's it's kind of nice that

21:15humanity has >> a solar systemwide presence. Right. you know, where even though Earth is in a crappy part of its orbit and we're kind of handcuffed in that way, >> we can still get that data >> from other parts because because we have this presence that's so big. >> This is why it's really cool. It's why both near and deep space matter quite a bit. And while it may be the case based on our second article that we may not live past 100, >> yeah, >> uh let's hope that these tools and instruments maybe one day do capture alien life in our local neighborhood. >> Oh, that would be dope. >> That would be fantastic. >> That would be fantastic.

21:56>> But our our second story is an interesting an a similarly interesting one uh about longevity no longer increasing. Headline studies finds generations born after 1939 unlikely to reach age 100. This is covered in the independent. uh but the research was out of it was an international collaboration between the Maxplac Institute of Democratic demographic research the institute national demographic and the University of Wisconsin Madison uh who we've mentioned on this podcast a couple of times ago and they published their paper in PNAS which is the proceedings of the national academy of sciences and so people are probably most familiar

22:36with longevity as a concept or studies either from the Netflix documentary The Man Who Thinks He Can Live Forever, which is about this 100 millionaire Brian Johnson. >> What is he doing? >> So, he's doing all he's getting the best doctors in the world. >> Okay. >> To like give him a regiment. >> So, he's now allegedly like decreased his actual age from his mid-40s, so he's like in his 40s to the the bodily age of a 20some year old, right? That's that's the like And so his whole thing is unless I die in a freak accident like my body is literally >> is is he one of these guys who like gets blood transfusions. >> So I don't want to say specifically but the only two angles are either the Brian

23:17Johnson story or the idea that all these billionaires are vampires getting their young employees to give them blood so that they have young >> because that's like a real thing. I I saw like news stories about that. >> I'm not sure with Brian specifically. >> Maybe Brian Johnson isn't doing that. Okay. But this is this is how in the past everyone's obsessed with it, >> right? Like we people want to live longer, all this stuff. >> It's always been the case over the course of that like as generations go by, we get older. We like are able to live longer. >> Yeah. >> This study is sort of suggesting not anymore. >> Not anymore. We've Yeah. We might have come to the end of the road with the lowhanging fruits. That's what this thing is suggesting. Okay. >> Okay. So longevity, there's been this thing called the longevity revolution over the over the past century, right? life expectancy has been increasing like

23:58crazy over the last century and that has everything to do with advancements in science% >> right um we've like figured out biology at a nitty-gritty level we figured out hey like um sickness actually comes from little tiny pathogens and we figured out ways to advance public health with vaccines with um antibiotics with all kinds of like revolutionary technology um medical innovations we've We've also had a lot of socioeconomic development. So, you know, in sanitation, in like the eradication of infectious diseases, small pox is no longer a thing, which is an incredible thing to think about, dude. Like, smallox doesn't exist.

24:40>> Yeah. This is which is controversial to like bring up the fact that that's an incredible feat for humanity. >> For humanity and science and everything, dude. Like I mean and and you know I I can't imagine it happening today which is so sad. >> Which is sad. >> Which is so sad because like back when in the I think it was in the 1980s when you know we were at the last stages of eradicating smallox and there was just a tiny little bit of small pox in Africa. Okay. And in the middle of a civil war, both sides decided to have a ceasefire so that health workers could go in and because there had been a small smallox outbreak there and health workers could

25:20go in and vaccinate everyone there >> and then and then they started fighting each other again. It's like it's like just stop fighting for like for like a few days, give us the time here and then everyone's like probably like >> planning how to get how to restart the fighting. But then but but even in the civil war they respected the fact that okay nobody wants small pox right >> all right like I want to win the civil war right but >> not at the cost of a nation with smallpox >> I can't imagine that happening today >> there's no there's no I mean a funny story that that someone always brings up to me about this is the former Senate majority leader Mitch McConnell okay >> I think when he was younger had >> small that's crazy

26:01>> and you know leader of the Republican party in the United States for a long period of time. He's been very vocal about the defense of like science in this particular isolated use case. >> Um and to your point imagining any of the current global conflicts going on pausing >> to do a public health >> I can't it's hard it's hard to envision. >> Yeah it really is. >> It's hard to envision. >> Yeah. which is sad. >> It really is sad. But like those are the advancements that got us this longevity revolution, right? Where we had an insanely increasing um lifespan, average lifespan. Okay. And one of the one of the things that happened as a consequence was policy

26:43makers just took it for granted. >> Yep. >> Okay. Where Yeah, of course we're going to just keep living longer, right? you look at the um you look at the the trend of like how average life expectancy is increasing and it looks like a line and so like any good uh high school student they get three data points they're like oh okay it's just going to it's just going to go that way >> right >> and so and and this this paper is really um putting that claim to rest a lot of people have been sounding the alarm though okay cuz any anyone who's reasonable is like okay just because it's linear for now doesn't mean it's going to keep staying linear%. Okay,

27:24there's plenty of systems that go linear and then they plateau off, right? Um but there's been sort of two camps. One camp has been there there is kind of an inherent limitation to how long humans can live. Okay. >> Okay. Based on just evolution, um mutation rates, the ability of our cells and our organs to keep up. Yep. Right. And then there's another camp that's like, well, you know, we could have enough science to circumn >> science, >> like inherent evolution, right? And and just like replace bad cells and do all of this bioengineering to just make us live forever, right? So those are the two camps. This is sort of adjacent to

28:05the transhumanism kind of movement where they're like there's there's going to be the symbiosis between all of our science and stuff and then that way we'll be able to be beyond we'll be this next thing that is >> yeah yeah we'll we'll merge with AI or some some nonsense right and it's like okay yeah maybe maybe in like a hundred years right but not me all right I ain't doing that no ain't no way boy >> but these but these are the two camps >> but those are the two camps right and and it's and it's been hard to sort of um justify one or the other without hard data. And that's what this group is doing from Maxplank, right? The the the paper that's come out in PAS um it's

28:46>> sort of taking a very rigorous approach to this problem. Okay. Because this problem is a hard problem to get good data and good analysis on. Okay. There's let me let me give you a sense of why it's hard. So there's two approaches to how we quantify life expectancy. Okay, there's the period life expectancy and then there's the cohort life expectancy. Okay, the period life expectancy is the easier one to um quantify. It's basically the idea that reflects the average mortality risk in a given year, right? >> Okay. So it's kind of like saying like okay like if a newborn is born now and you say the life expectancy is 80 years but that's a period life expectancy.

29:27What you mean is that like given the climate right now, this newborn, given the mortality climate right now, this newborn would be expected to live 80 years, right? Given the chances of dying on a random person, right? But that is a very local uh local in time measure. Right. Right. Because um if the COVID pandemic happens, the that period life expectancy goes down. If um the Spanish flu happens in 1918, the period life expectancy goes down. But like it's kind of a live metric of life expectancy. Okay. Then there's something called the cohort life expectancy, which is like saying okay >> the people born in 1950 in that birth

30:08cohort, the average lifespan of that individual is 80 years. Okay, >> that's different because that is taking into account this giant time span >> of 80 like 1950 plus whatever 80 years, right? >> And in principle that's really hard to quantify unless everyone has died from 1950. Right? Then you take everyone who died from 1950, you average it out and you say that was the cohort life expectancy. It's >> so it's it can be useful in a retrospective context when you have a generation or cohort that's already >> Yeah. That's already That's already done. >> But you It's not predicted. It's hard. You can't play it forward as like a Well, we have a generation that's like halfway. >> Yeah.

30:49>> And like let's like maybe predict >> Yeah. Exactly. >> these unknowns about the future environment. >> Exly. Yeah. Yeah. So, so you see now the struggle, right, that life scientists have in in defining a cohort life expectancy of a generation that's still alive. Yes. >> And that's what these guys are trying to forecast. >> Okay. >> Okay. And so the way they did it was was really cool. They they got six different types of models. Mh. So you you know in social science you you if you just do like a single model lots of compounds you want to do a bunch that's why I like this paper right they they were pretty rigorous about it they you do six different different approaches and you see if they all give the same answer if they do then it's a pretty robust finding got it >> right on the other hand if it's a giant

31:30spread that means that the nitty-g gritties of the model and the initial data that you put in that's what's messing it up right so they did six different models um like the Lee Carter model the cohort segmented transformation, agitate death model, all of these like different kinds of models. They're basically like fancy data analysis techniques where what you do is you take a time series of data of different age groups and how they're doing in terms of mortality and then from that you project forward and try to get like a sense of okay when is the time series going to end. Yep. Okay. You try to do this for six different things and what they came up with the same they came up with the same >> end point >> end point which is that this life expectancy is going up but it's not

32:11going to get to 100. It's going to plateau. >> Okay. To validate it you might be thinking okay well these are just models right? It's nice to validate it they they forecasted life expectancy for people from 1919 to 1938 who are all dead. >> Okay. Everyone born in that time period are all dead. So you can take that data, right, as your test data in some sense, right, and be like, okay, I'm going to do the same analysis for this data from 199 people born in 1919 to 1938 and then try to project >> Yep. >> when they're going to die. And it worked. >> And we have the answer for that. >> And we have the answer for that cuz we have the data about when they die. >> So we built a sort of predictive model to be able to work for current cohorts that are still alive. And we validated

32:52it and it worked across six models and then we used a historical cohort who everyone's dead to put it in to see if it mapped onto the actual results we see saw from that cohort that's now all >> Yeah. And it was a little bit off but not enough to justify what they were seeing with the present cohort. And it might be a little bit off because like you know when when you're trying to do like obviously the the the environment of the data back then is different from the environment now, right? But then you can make arguments about how off the result should be. And the result should not be that off compared to what we're finding now, which is that this thing is definitely plateauing. >> Mhm. Okay. >> Um, >> yes. >> The cooler part of this is they actually

33:34did a kind of causation, >> okay, >> of what is causing this plateau, >> right? >> And what's causing this plateau is the fact that we have exhausted all the lowhanging fruits when it comes to longevity of human lifespan. We we've we've taken all the low branches on the tree for extending life and we've actualized >> we've actualized them all because young people are now living longer right no one's I mean people obviously people are dying before the age of five but much less infant mortality is way down okay people dying before the age of 20 is way down because infectious diseases are way down right like the problems that you have as a young person are no longer the

34:16kinds of problems that you used to have before. Okay? Like >> because of vaccinations, because of fundamental science, because of all of the advancement that we've been doing over the past century, >> right? >> The hard stuff is now the actual aging. >> Okay? That's the cancer, >> right? Right. >> That's the Alzheimer's. That's the That's the That's these >> higher order things that are notoriously hard because they're no longer >> they're no longer agents that are trying to kill you, okay? They're yourself. >> That's just like sort of >> at a cellular cellular level giving up in some sense, right? Like the neurons

34:56are sort of dying and the nervous system and the brain is in some sense deteriorating just because of >> like time right and in cancer cells it's like the mechanisms that control the growth of cells is sort of giving up and then now you have these tumors growths right so so it's a much more complicated problem now >> right >> that we haven't really had the kind of progress that we have had in cases like polio and smallox and all of these other sort of you know quote easier problems. They weren't easy at the time, but now that we've looking in hindsight, it's like, okay, they were all like agents. And now we have these like modular

35:37mechanisms to like solve that, right? Oh, just make a vaccine. Corona virus, like within 6 months, we had a vaccine. Easy. >> Now we have to solve for our own body trying to kill us over time. And and and the open-ended issue with this is it's not clear that in the short or midterm that we're going to actually make this. sufficient or or significant progress that would lead to outcomes that were at the rate of progress we've seen historically. >> Exly. Yeah. Yeah. It doesn't it doesn't seem that like it's not going to we're not making progress at the speed that would like temper this plateau. >> You know what this reminds me of funny

36:17enough um is in technology there's this you know markets and and cycles are defined by this scurve right. Yeah. where you start, you know, with a low rate of change and then you have this explosive growth and with every technology cycle eventually it tapers off. Mobile phones is the perfect example, right? We went through years of Apple innovating and doing new things and combining all this stuff and then all of a sudden everyone catches up and then we kind of get to the place where we are now where is a foldable versus that really >> yeah, >> you know, or are we now just kind of making things for marketing? Yeah. >> And so this similar idea of sort of the S-curves that you see in technology cycles, >> it's happening in fundamental science when it comes to biology when it comes

36:58to this longevity issue. It's like we kind of have reached the end of the first S-curve. Yes. >> And so in technology the idea is S-curves move, right? So it's like you first have like vacuum tubes, all this stuff. Then you get to microprocessors, then you move to mobile, PC, mobile, etc. >> The open question is arguably we just went through the first, let's say, longevity cycle, curve cycle. >> Yes. And it's not clear what the second S-curve cycle may or may not look like and whether we can traverse it or not. >> Yes. Exactly. And like if we if we get to that S-curve, what is the real >> upper limit? Right. Because right now the upper limit I think the record is 122 years for someone to live. Okay. And um >> is that the real limit? Right. Is it at

37:40that point the body is just like nah. >> Right. Right. evolution has literally programmed us to be like >> no no no we're still evolving right the human genome has a really high mutation rate compared to like >> like for example sharks right sharks have like a insanely low mut mutation rate which is why they're fossils 300 300 million years ago look exactly the same there's no humans like even 10 million years ago right that look like us so the the mutation rate is really high on top of that like >> you know we're we're incredibly like agile creatures You know, we're mammals at a high body temperature. So, high metabolism. So, all of these things sort of like start contributing to tempering,

38:22>> right, >> an infinite lifespan, >> right? The limits. We we have these limits that exist that are fundamental to our the the existence of our being. >> Yeah. >> And it is not clear that they are necessarily surmountable by a human manufactured solution. >> Yeah. In in human time spans. >> Exactly. Yeah. And what I liked, I mean I guess not what I like about this story, but what concerns me about this story is that I think we need to again start taking this very seriously as a policym substrate. Right. >> Right. As like as like now a scientific fact that we have learned and >> um urgently need to act on. Okay.

39:04Because from a policy perspective, I mean there's to extrapolate this there there's huge arguments right now over the the debt in the US uh becoming being at un untenable levels and one of the biggest line items in that is social security. >> Yeah. >> Right. And it makes assumptions about retirement age and who gets Medicare. Right. >> And what happens if you're the line of average life expectancy shifts or moves? >> Yeah. It's no longer I mean everybody has made this policy based on an increasing life expectancy forever and now this paper is saying that actually no life expectancy has plateaued. we're at the end of the road when it comes to like how old we can get, right? like all

39:45these like you know with with pension funds with social security systems >> there there was a riot in France right the other year when when like um the government tried to raise the retirement age the reason why they're raising the retirement age is a pretty like simple logic right which is like if people are living longer then they can work longer and if they can work longer we should make them work longer before they get the retirement well this isn't the case anymore people aren't actually living longer they're living the same as they were like 10 years ago. Right. >> Right. We're not actually we're not actually making it >> better. Right. >> And the worst part about these kinds of policies that sort of raise the

40:25retirement age is um >> it actually impacts those of worse socioeconomic background, right? Because the rich people, rich people can just retire. You could retire at 50. You probably already did. It's the poor people that are still going to have to work until like 70 now and and then what? die at like 72, right? Like, you know, >> 100. No, it's a really it's a really important point about how a lot of this very fundamental science research >> historically, especially in the United States, has both informed like military and intelligence posture as well as policym posture. That's right. Yeah. And there's just been this disconnect where for whatever reason um there's no longer

41:07the through lines from fundamental research and the insights that are gleaned from that into bleeding into these areas of that are in the control of nation states. Yeah. >> Uh in terms of how do you sort of organize society around it? And this is this is like a meaningful uh and important point because people do still generally do still think like oh yeah we're just going to live longer than our parents much like every other aspect at least in the US of life right we are less like the millennials and younger yeah >> we're less well off financially than our parents everything is less affordable than it has been for our parents and we're the first generation >> um to be less welloff than our parents and so the idea that we're going to also

41:49now live just as long >> makes like it You can kind of see like all these things are like aligning. >> Yeah. And then and then when they when when you know our parents generation retires, the the the population pyramid is now basically inverted for all of these developing countries where this is true, right? >> The Japan problem, >> right? Like it's it's now there's fewer and fewer young people to support this older population for social security. Turns out given that population pyramid and the mathematics of just 2 plus 2= 4 and it's never five, like >> this might just be a scam, compound on top of that this idea that like >> actually the fundamental premise that like the younger generation is going to

42:29live longer and so can contribute more to the social security pile is now no longer correct. And we need to start asking questions like is this like it's there's a fundamental shift now I think in my mind from is this a tenable strategy >> to is this even a correct way to do business in a developed society. >> Right. >> Right. I don't want to sound communist because I swear I am not a communist. Okay. But like at some point you got to start asking questions about fundamental things like the social construct, right? Like is social contract sorry yes like you know is this true like is the social

43:11contract of the post-war era still applicable today right and the fundamental variables in that equation have changed so drastically and we're still saying the outcome >> are going to be the same even though all the variables on the other side of the equation are fundamentally very very different. >> Yeah. Yeah. Like this is a reality now and it's like going to catch up with us as we grow older. Right. >> This is this is a really really important story. I'm sure we're going to come back to the subject of longevity because I think your the nexus you brought up from longevity studies to policy >> is a really important note especially when we're seeing not only in you know western let's say quoteunquote western developed nations but any developed

43:52nation th this similar issue of of generational friction yeah that arises from 50year-old to 80year-old policy conceptions that uh are based on fundamentals that are now statistically, data wise, and scientifically are proving to not necessarily be the same. >> And yet we've not changed any of our >> posture. >> Yeah. Yeah. Yeah. >> And then we're just going to run into a brick wall at some point. And this is it's it's it's a problem. >> Yeah. >> Um >> Yang Gang. Sorry. >> This it's it's it's interesting because it it does dovetail somewhat to our our third story. But one aspect we didn't

44:32talk about the factors of longevity that change is you know climate is you know a big part of that >> it's getting bad >> and our our next story is about a new no sort plastic recycling being near. The headline new nickel catalyst enables no sort mixed plastic recycling. This is a uh new brand new breaking research coming out of Northwestern University. They put their paper out in nature chemistry. Yeah. Um I just went to Japan earlier this year and one of the interesting things about in Japan unlike in the US is they have different recycling bins for different types of recycling glass cardboard etc. In the US we just put it all in the same thing. So my assumption is this no sort plastic is

45:15me saying like there's no need to xyz like this this seems like it's on the on the sort of like pollution and our need to be more effective with our resource usage. This might have impact, but tell me tell me what we need to know about this story. >> Yeah. Um, plastic, we love it. >> Yes, we do love it. >> We do love it. As humans, we are cheap >> and we want things. >> We want things to be easy. We want them to be cheap and so we love plastic. Okay, plastic is um really bad for the environment. It can it it sticks around for like millions of years or something like that, right? Um and we are producing a lot of it. something like

45:56400 million tons a year. Um something like like 11 million tons go into the ocean, which you know I am I am I'm very I really don't like that. Okay. But I'm also um a lazy selfish American >> and um >> I like my things, right? I like I like my mustard and ketchup bottles in plastic and my milk jug that comes in a gallon in plastic, you know? I don't want to go with a glass bottle and refill the same milk jug. It's just so convenient to just to just go to the store and grab a new thing and then throw away the old one and just not

46:36worry about what the old one is doing. Right. So as as Americans, we are very selfish and we're very lazy and um you know it's not just something that Americans are, it's something that the whole world is. This is a world problem. And you know, everyone who says that like we need to start using less. >> Mhm. >> It's just simply not a practical way to save the planet. Okay. >> Like we love our planet, but we love ourselves and our comfort way more. And that is just not going to change. Okay? Even the developing especially the developing countries, right? Like China and India and the continent of Africa

47:17that are also contributing immensely to this chaos of of plastic and consumables and all this stuff. They're looking at the developed countries being like, "No, you guys did that. Well, you guys just had 15 years of doing that and now we're >> and oh and now it's like, "Oh, I'm sorry. I can't use coal." like you know when that's the one resource that's super cheap to use to make electricity which is something I need so I can like educate my kids. >> Yeah. >> So so this idea that like you know using humans doing less to save the planet is not going to work. >> Okay. The solution that is going to work is scientific innovation to make our current lifestyle better for the planet >> and more sustainable.

47:58>> More sustainable for the planet. I know this is a hot take about like how we need to deal with human directed climate change because it is real, right? Humans are changing the planet. But my my fundamental take is like we're not going to stop. I I mean at the end of the day the the idea that you know people using paper straws instead of plastic >> but then all every corporation and manufacturing entity and these large people who are actually producing the largest volume of the problem >> uh is really like where the problem is in terms of having the most impact for our efforts. >> Yes. Yes. It's we need we need we need a bulk strategy because even if you made

48:38every human like individual human like feel bad corporations who are apparently also people they're not going to feel bad because the mob is crazy right >> no exactly >> and so and so no they're going to be like oh yeah just dump it right there it's fine you know so so we need a strategy from fundamental science perspective that helps us recycle efficiently and in a better way okay plastic recycling is a key problem that we as humans I think we need to solve. Okay. And this is a really big step in that direction. Okay. So, let's talk about plastic. Plastic is really bad. I just told you like something like 11 million tons go into the ocean every year. It's going to just stay there for

49:18thousands of years. It's not going to decompose. Plastic is essentially chemically it's just um polymer chains of hydrocarbons. So, um carbon hydrogen hydrogen hydrogen carbon hydrogen hydrogen hydrogen just chains of carbon and hydrogen. Okay? Like that's all plastic is. It comes from from like petroleum. Yes. >> Which is also hydrocarbons. So another like you know it just it's another use of petroleum in that case. Um and in order to recycle it there's basically like a few ways that we do it right now and they're all pretty bad. Okay. One of them is the thermomechanical method which is what you got to do is you got to basically materially degrade this thing. Um you've got to like heat it up.

50:02you've got to make it into like lower worse plastic and then use that plastic somehow for some other purpose. Okay, it's a pretty bad way of doing things. Um, but we do it 9% 9% of the plastic globally is done this way. Um, it's really expensive. Nobody really wants to do it because you got to sort all these plastics >> from for all their different like types of plastic, right? >> Because they break down to different fundamental component parts and you don't necessarily want those mixed. >> Yeah. And then if they mix then they make even worse >> like stuff and then and then they just go to the landfill and it's like okay what did you even do? You just Yeah. Um so so that's that's one way. The other way is pyrolyis which is just just heat

50:45it up. Heat it up. Degrade this thing and then you you get some kind of low value type of um GHG which is like some kind of like low value fuel. Right. The problem with that is okay fine. How much energy did you use to like heat this thing up like to get a a low value fuel? You're getting it's like it's like nuclear fusion, right? It's like why why are you even doing that? You're putting more energy in to actually get like some kind of fuel out, >> right? >> On top of that, if oil prices are super cheap, then like why would you even do this, right? Um so >> the and whenever you're doing anything like this, there's also all of these

51:25contaminations that happen, right? Because plastic has like food products sometimes they have PVC which is polyvinyl chloride and if that gets in here and as it's degrading so this pol polyvinyl chloride it's it's found in like pipes flooring and if that gets in there even in a batch a small thing in a huge batch that whole batch is done because you get this hydrochloric acid um all the equipment is done and at the end of the day you're dumping into the link >> and you fail anyway. >> Okay. So, so it's like it's really like kind of kind of bad. >> We have a lot of bad options, right? >> We have a really a lot of bad options. Like the only good option is to take the plastic. Usually it's in this thing called polyolifin

52:05um which is a CH2 carbon two hydrogens attached to the to a C and a H and then some other thing an R group a residual group. Um those are the things that are basically 2/3 of the global plastic chain. This is like, you know, your milk cartons and um plastic spoons and all like plastic wrap and all this other kind of stuff. Um single use, really short lifespan, >> just gets used, gets thrown. Okay. That's what we're trying to at the end ultimately we're trying to we're trying to recycle. >> We're trying to repeal and replace it. >> Yeah. Yes. Exactly. And the what I just

52:46told you, right, the the options are pretty bad, >> right? >> Okay. There's one option that has a little bit of hope. >> Okay, and that is this thing called um hydrogenolysis. >> Okay, >> hydrogenolysis. Okay, it's basically breaking it down in the presence of hydrogen and what it does is the hydrogen sort of comes in along with some kind of catalyst like that helps this reaction and that hydrogen comes in and with the carbons and the hydrogen's there it creates hydrocarbons that can then be used for either fuel or like waxes or like these higher sort of um lubricants things like that that you can recycle them to. Right.

53:27So, so that's one way of doing things. That's like kind of a promising way of like recycling this plastic and make putting it back into simple hydrocarbons, right? And and that's something that's like much less worse for the environment, right? Like animals aren't choking on it and like that. So, so that's what we want to do. Now usually this hydro hydrogenolysis takes catalysts and those catalysts have um are require like higher order elements rare earth elements like platinum palladium >> we don't got we don't got like the amount of plastic we got >> compared to the amount of platinum and palladium like this is the stuff that Iron Man used to make his like >> new element okay so that's how rare it

54:09was that Marvel was using it in a movie so we we can't be dependent on platinum and palladium to to like recycle the tons and tons of plastic, right? How many mines we going to get? Right. >> So So the idea is you want you want to have a catalyst that is not that rare. >> Mhm. >> That's what this team is doing. >> This this makes >> fundamentally. Okay. Fundamentally, they're they they found a way to catal catalyze hydrogenolysis using a catalyst made out of nickel which is very abundant. Furthermore, this catalysis only targets a specific type of this polyolifhin. Okay? Because

54:50as I said, there's the CH2 CH2, right? And then there's a R group. That R group can be sometimes H which makes it polyethylene >> or it can be sometimes polyropylene which is the the the the other group is another comp complex thing like another CH2. So you get these like you get um with polyethylene you just get this chain of carbons with hydrogens's attached. Yes. >> Like a caterpillar. Yes. >> Or you get for um polyropanine, you get um the chain of carbons with carbon sticking out. Yep. >> Okay. >> And so what you want to do is you don't want to recycle these together. >> Okay. >> And usually that requires separation. And that's really really just a pain.

55:31>> Yeah. Yeah. Yeah. Yeah. >> Nobody wants to do this. >> Both both time and cost prohibitive. >> Both time and cost prohibitive. And more importantly, it's cost prohibitive. Right. And the consumer is not going to do it cuz again, we're lazy. We don't want to do that. Yes, I love the earth, but I don't want to check under the bottle what kind of plastic it is before I throw it in the garbage. >> And this goes back to the sorting issue, which is like ideally this is why you want to have the humans who have the products put it into different buckets so that when it gets to the facility to recycle, it is pre-sorted. >> Yeah. >> But Yeah. And it's like even even if you got the humans to do it though, do you really trust that human? >> That's the point. like I don't know if this human can read. Right. Right. So

56:13it's like it's like I don't know. Right. So you would rather have a way of doing it within the chemical reaction, right? That targets one type of plastic and not the other, >> right? And that's what this thing is doing. So they developed um a catalyst out of nickel. Um it's got aluminum, nickel, the usual carbon, and all this other kind of stuff. And um one of the cool things they did was this catalyst, you can imagine it as again a Lego block. Okay. But this Lego block has only a single piece of like active sight. Okay. Whenever you have a catalyst, you have something called an active site, which is the part that does the locking and the moving

56:53around of the atoms. Okay? This thing has a single active site. Okay? And what that gives it is >> the benefit of being extremely specific to what shape it wants to actually >> attach to. attached to, right? A lot of catalysts have different shapes, which means that when you try to put it into something like polyropylene or polyethylene, they're going to attack every single carbonarbon bond. Okay? And then you're going to get a bunch of methane, which is fine, I guess, but like it's not like great. Methane is like super cheap. It's like this low what what you what you'd like to do is with poly polyropylene, I told you, right? There's these carbon and then there's these carbons that are sticking out. You want to just take out the carbon that's sticking out. Okay? And then and then and then this thing just

57:34becomes like polyethylene and then you can like then do those two together in some other sort of >> industrial process, right? Um so you want to take those out and then and then only target specific bonds and that's what this new catalyst does very sorting within the chemistry >> within within the at the facility where they're going to be doing the recycling. Yeah. Regardless of the input, the chemical process itself is doing the job of separation. >> Exactly. >> That we otherwise are in some cases in some places depending on humans to put things in different buckets. >> Yeah. Exactly. And and the the results are they speak for themselves. So they had um they they tested this catalyst

58:16with isotactic prop polyropylene which is this thing that I was telling you about with the branches coming out and 99.5% conversion into lighter lighter compounds in 20 minutes. And then when they did the same thing they did the same thing with um the uh polyethylene. >> Yes. >> Nothing happened. >> Mhm. >> Right. >> Yes. >> It's like and it took 4 hours. >> Yeah. Yeah. Yeah. Yeah. >> And then they finally got a little bit of signal. So what you can do is you can have a mixture of this stuff. You can apply this catalyst and then you can just let it run and then within like within a short amount of time all the polyropylene will be decomposed but the polyethylene will be in this native state and then you can like go forward >> and now you have these two buckets of of

58:58distilled compounds that that you then have functional use for in any number of other downstream or upstream depending on what way you want to look at it industrial process. >> That's right. Yeah. And then the best part was that you know that PVC that I was talking about the polyvinyl chloride which is a contaminant for all of these plastics that basically makes plastic unrecyclable in these previous strategies >> somehow. And I don't think they fully understand how this works. This made it better. >> Really? >> Yeah. The the the introduction of PVC made the catalysis reaction better. >> No freaking >> Yeah. So then they were like, "Oh, that's that's free. That's great. >> Just like free upside. It's like, oh,

59:40that's great. Yeah. Which then removes the crosscontamination issue. Yeah. >> Which goes back to this no sort like again. >> So then so then we don't have to sort it. We don't even have to worry about the PVC. The PVC could be in there >> in there and it's just going to make it better. >> Yeah. It's just like fine. So I think this is a really cool um use of like fundamental chemistry to try to like tackle a very very real problem. >> This is really good. That that's actually and that's a great explanation. Um because I actually had no context for >> how the recycling process really worked. >> Yeah, I didn't either until I started researching this >> and this is like a clear like it's a clear problem set. We don't have good solutions currently implemented at scale. >> No. >> Um and we we >> but it's something that I really do care

1:00:20about, right? Just you know, >> right? But do it at the factory. >> Yeah. Yeah. Like I'm I'm lazy. I'm sorry. as as are the major like again it's also should not be incumbent on we all there's the commons is a concept we all should invest in in the commons and trying to upkeep but at the end of the day >> the biggest kind of issue is the industrial grade industrial level >> exactly it's hard for me to yeah >> like the impact is most there and we have to address it at that level and one of the ways you can do so is like when they get bring the stuff to the facility it can actually repurpose it in a way that creates functional valuable uh output that is not the current state of affairs where it's either the output

1:01:02sucks and we wasted a bunch of energy, time and money. >> Yeah. Or >> just it just went to the landfill. >> It just went to landfill anyway or the output is not really is not as like doesn't have real value in terms of the concept of recycling it back into society. >> Exactly. Yeah. And and your point about the commons I think is important, right? There is a commons that I am heavily invested in. I pay taxes and I want those commons, my tax dollars to go into this kind of research because that is the future of saving this planet. Saving the planet is not making humans somehow be less comfortable. >> Okay? That's not going to work. I know humans. I know several humans. They're all like me. Okay?

1:01:43>> You know, and what we want is solutions that actually make sense, right? Right. That don't compromise my state of life. And like I yeah we we need we need fundamental science research to do this. >> I think we've had enough innovation on our mobile phones. Let's apply that brain power. Totally fine. >> Yeah. With the current state of my iPhone is sufficient. Uh maybe let's like not try to iterate on an incremental version for next year and spend that brain power, time and energy on things that will create a better common environment which then drives the consumerism that capitalism loves. Great great story. >> Yeah, I like the story. Um, speaking of the private sector and uh, multi-t trillion dollar companies, we're going

1:02:25to end today with a story about analog AI coming out of Microsoft. We almost cannot avoid AI every week. >> Nope. >> Headline on the story, Microsoft's analog optical computer cracks two practical problems and shows AI promise. This was published on Microsoft News's blog, but they did put out a research paper associated with this in nature. Yeah. So, it's not just a big deal. It's not just marketing. Uh they did sort of go through the peer review and so I don't really actually even understand what they mean when they say analog optical computer. Yeah. >> And there's several problems to crack, but they cracked two of them. So let me know what what are we talking about here? >> That's right. Um it's an analog computer

1:03:05instead of a digital one, which is very cool to me. Um digital, you got your zeros and ones, right? >> Um and we talked a lot about digital computers in the past and some of their problems, right? um the fact that Mo's law is slowing down because we're sort of reaching this limit of how small we can go >> before the capacitors and the transistors that we're using start leaking off so much power that we just have to constantly keep powering it in order for the data to actually stay alive. We also talked about the um vonoman gap, right? the Vonoan bottleneck which is the separation between >> your processor

1:03:47>> and the memory there which is your RAM and your hard disk or your actual memory. Yes. Right. And the where you're storing stuff is different from where you are doing the computation. Yes. Right. This is the thing that plagues CPUs but it also plagues GPUs. Okay. And um as AI becomes increasingly big, there's going to come a point where we're going to really have to start having the conversation of again, you know, what we were talking about last in the last story. Is this worth the environmental impact? Right. >> Right. And of course, the corporations are going to be like, yes, and they're going to make bigger and bigger data

1:04:28centers. For example, the energy usage has tripled from 2014 to 2023 and it's projected to be 20% of global electricity by 2030. >> Okay, the data centers and the AI, all of that AI computations for you making a cat that wears a hat or whatever. Like that's 20% of global electricity. Now, I'm not saying that we shouldn't be using AI like because I use AI. AI is great. Um but I think all of these companies including Microsoft are starting to realize that it's no longer going to become tenable right >> 100%. >> Um there needs to be an alternative way to do the computation that they are

1:05:10seeking. I'll just make a brief note here which is that every single one of the uh either uh frontier model builders or uh sort of hyperscaler cloud computing platforms Google cloud product AWS um they're all spending a ton of their capital on directly investing into power infrastructure. >> Yeah. because they know that that's going to be a bottleneck. >> Every single hyperscaler or frontier model company is investing if not hundreds of millions, billions of dollars >> into partnerships with these mini

1:05:51nuclear reactor companies. They're literally returning on the big power facility that's just north of us in California, the nuclear one. Facebook has a little al cove there. So, they all know the story. Yeah. Like it's not they see it. They see the writing. It's not controversial for any of them. Yeah. And and on top of that, there's a bunch of these companies that are investing in alternative computing technology, right? Like we were talking about Spintronics the other day. Um this is an analog computer. It's doing sort of the same type of deal. Not the same physics, but the same sense of let's go away from transistors and um >> physical stuff and let's try to find a

1:06:32different completely different way to store data and manipulate data. Okay. Um, if we go back to the vonoman bottleneck, right, from John von Noman, he um championed this idea of that early computer where there was going to be a separation between the processor and memory, right? That's become sort of the bedrock for modern computation. But it's really become a huge handicap for AI. Okay? Because when we want to do AI, there's two core operations that we have to do. Okay? One is matrix multiplication, right? Where we take the weights of your neural network and you multiply it by some vector and then you multiply by another

1:07:13matrix and another matrix and all of these are weights that you've stored. Um, and the other thing that you have to do is you have to do a nonlinearity which is this idea that if if my input is above a certain threshold only then do I let it through. If it's below a certain threshold then it's a zero. Okay, that's one of the types of nonlinearity that you have to do. Those are the two fundamental things that you have to do as someone who wants to implement a neural network. Okay, >> then there's other things like momentum and annealing and all this other stuff that you have to do. But fundamentally, if you really want to do neural network stuff, it's matrix multiplication and implementing a nonlinearity. Okay. So

1:07:54now, in order to do these things in a traditional computer, what you have to do is you have to store your weights in a certain spot. Okay. And then you got to load them up in your processor. And then you got to you got to multiply the weights to this thing. And then you got to load up new weights. And then you got to multiply to this thing. And you know, now we've gotten clever where maybe we can load up all of the weights to this thing and you know do a forward pass and things like that. But fundamentally a lot of the source of latency that comes from like you waiting while you you've typed something in a chat GPT and it's like thinking is doing this, right? It's like loading up and then there's like copper wires in whatever server that is like transferring data and the latency

1:08:37is a huge part of it too, right? So there's a lot of power that's being dissipated for no reason and then there's also a lot of latency. This thing is solving both of those. >> That's very interesting. >> Okay. This thing is solving both of those by creating a computer out of light, >> lenses, and filters. >> Very interesting. >> Okay. Okay, >> it's very cool to me because it's like fundamental physics in used in a whole new way. >> Okay, >> so to to understand this, let's let's look at the two things that we need to do in order to do a neural network implementation. Okay, the first one I talked about was matrix multiplication. That's this idea of you take a vector of numbers, which is let's say a row of

1:09:18numbers, and then you apply a matrix to it. So this this number gets applied gets multiplied to all of these. this number gets multiplied to all of these and then you sum it up and you get a new vector kind of like things like that, right? Um what these guys figured is we could do that with like light and filters. >> Okay. >> Okay. So the vector is an array of numbers, right? We could represent that with the brightness of LEDs. >> Do you see where I'm going now with the whole thing? >> I already see where you're going. >> Okay. I could represent that with a brightness of LEDs. >> Okay. >> Then I've got a matrix that I need to multiply it with. Right. I could

1:09:59represent that with a bunch of filters. >> Oh no. I >> that each light goes through. >> I'm so mad. >> It's so nice. >> It's so nice. Right. And then I've got an output vector, right? That output vector could be a CCD that senses how much light went through. Okay. So completely analog. >> Right. Right. Right. There's no zeros and ones. It's like something it's there's a continuous value here of the brightness. There's a continuous value of the filters and then there's a CCD that captures whatever comes out. Right? And then now okay the second thing I have to do is implement a nonlinearity. Right? In order to implement the nonlinearity they've got like a bunch of

1:10:41filters that they can use like the bipolar difference filter. So they've got so now they've got these CCDs right that converts it into current. The current now goes through analog. It's still analog. A lot of these implementations, what they've done is they've done a hybrid approach of digital and analog. This is all analog. >> Pure analog. >> Pure analog. It's wonderful. You got these you got the you got the light that's coming out like this. >> It is now going to go through this sort of filter block. Okay. It's going to go through a bipolar differential pair which is going to do the nonlinearity. So if the light if the current is below a certain amount, it's going to give a zero. But if it's above a certain amount, then it's going to let through a bunch of current. Um, you also want to do this thing called annealing, which is

1:11:22this idea of like taking the the the current representation and sort of changing it a tiny bit so that you don't get stuck in like a state of the neural network where you're not like moving and you're getting the wrong answer, so to speak. So, so you can do that using these things called variable gain amplifiers, >> which are VGAs used a lot all over tech. Um, you can also do this thing called momentum, which is something that you use in neural networks. If you're like converging on an answer, you want to keep going in that direction. That's something that you can implement with VGAs, which is again something that everybody uses. And then you can feed it back >> into the thing and have this loop. Okay. Now, this is a fundamentally different

1:12:04architecture not just for the physics of it, right? because we're not using GPUs, we're using light and matrices in the form of filters and things like that. But also you can imagine it as like there's a single sort of neural network that we're like passing over and over again right? >> So what these things are implementing is something called deep equilibrium models. There are different type of neural network where you've got a neural network and instead of a forward pass that gives you a right answer. Yeah, >> what these models do is they call the model over and over again until the model doesn't change its output. >> Okay, so what it's doing is it's in the

1:12:44space and it's going down kind of like a gradient descent algorithm where it's grow going down this sort of energy landscape until it gets to this minima where as it moves it's not really moving anywhere. Right? Before it was moving here, then it's moving here. It's trying to find the answer. But when it once it gets to the answer, the more you apply it all the way to infinity, it's not really going to move. Right. Right. That's what this thing is doing. The problem with deep equilibrium models on GPUs was it was like extremely I like how many times am I going to like but this the archite the the physics itself is geared towards these deep equilibrium models. >> Yes. Right. >> And so you've got this you've got this entire like system that is uh digital

1:13:26analog. I mean, sorry. It's an analog computer, >> right? Right. Right. >> Which is which is I think which and and because part of the this goes back to like the energy and cost issue with running, you know, GPUs are both expensive on their face to produce at the scale we need to to do these like chain of thought reasoning level models, these frontier models. Um, and then the power usage to not only to acquire the physical hardware is cost prohibitive. But then to run them Yeah. uh to be able to get to that level of is also cost prohibitive. And this sort of solves both issues both on again accounting for the fact that this is obviously not an industrial manufacturing scale and all these other things.

1:14:08>> This is not industrial manufacturing, but they're not using like crazy tech. >> That that's that's using LEDs. They're using the the matrix that they use for the multiplication. That's just like the filter in your projector. >> That's very interesting. >> Okay. And then the CCD is like the the the sensor in your phone. >> So this >> and then you got VGAs which everybody knows how to like everybody has. We know how to make it super cheap at this point. You don't need like an insane fab to do any of this stuff >> which is this becomes very interesting because of the I mean there's so many implications here not only uh from a just from the fundamental perspective but if you extrapolate both like geopolitically and economically you know one of the issues with the sort of whole

1:14:49AI debate is Nvidia has now become a trillion dollar company in record speed because they're the only ones that can build these chips these GPU sorry these GPUs uh that can really run the frontier models effectively um and to be able to get to Nvidia scale because of the complexity of the manufacturing process and the amount of proprietary knowledge necessary there's like no IBM has tried they can't like and there's no there's no one else right >> yeah yeah because they've done they've done 20 years of work to get to where they are right and and this what's interesting is this sort of upsets the apple cart in the sense of if there is a lowcost high efficacy you know hardware substrate that you can utilize to do the same type of work that

1:15:31you're seeing out of a G5 drive faster with less power with less power >> like one caveat is this thing is only going to work for inference >> okay it's not for training >> which is which is >> because in training you'd have to change the weights and things like that right so in order to train what they did was they created a digital twin >> which they then trained y >> and the digital twin matched the hardware's y >> performance so then when the digital twin said okay you want to set your filters this way and you want to set your nonlinearities this way they could go in implement that with the hardware and then they could like do the inference time stuff on this. But to be fair, >> like all of the like the reason why all

1:16:12of these companies are building all of these data centers is not to train. It's for inference. >> It's for inference, right? The training happens in like San Francisco. Yeah. >> In their facility or you know whatever in like some like tiny like the hundreds of millions of users of chatbt and Gemini are just using it for info. All it's all inference. I was going to say like where the usage is going as more >> like the power is going for inference. >> Correct. As more users start using you know LLMs and other versions of AI in everyday life and we get to the order of billions billions billions of users doing so they what they're using is the inference. And so this can be a whole you can still have your proprietary training stacks with all these fancy

1:16:52GPUs and all that stuff, but now you don't need to have that high cost infrastructure in order to serve the end user of the models. If you can miniaturaturize this thing, then >> you're good to go. You can I mean again we're extrapolating here, but you can imagine I have my own little analog. >> Yeah. Yeah. And it's just And the great thing is this uses LEDs. It's not using lasers, right? So I don't need coherent light. This is just like normal LEDs. So incred That's so clever. >> Yeah, >> that's so I love >> I love that. Yeah, >> I think >> they I think they did a really good job. I quite like it. >> Creativity is alive and well and fundamental science research. One of the ways you get breakthroughs is >> by being creative, not dogmatic. Um that

1:17:33is really really I coming from the tech space this obviously my light bulbs are going off. >> Yeah, dude. And it's like it's like you know 20 nanconds per search. >> Mhm. per um iteration of this thing the the the idea is called a fixed point search because you're searching for a fixed point to get to the answer right what I was telling you earlier about um so yeah 20 nanconds per fixed point search and this thing is projected to be 100 times more efficient than leading GPUs going at 4.5 terra operations per second per watt right that's 100 times more efficient than GPUs >> that's actually crazy yeah no that's actually crazy >> it's a

1:18:14um >> like in terms of per watt like the yeah >> having having a 100x decrease in watt watt like w like usage >> that's that's that's >> not much that's so sick actually yeah and and they proved it um you know with their validation obviously you make something like this okay what how are you validating that something like this works so they got their dig digital twin to train on the emnest data set which is the classic um handwritten d hand handwritten digits data set to identify like two four um did really well on that. Train it on the emnest fashion data set did really well on that. The other thing they did was um they can really do so this this was AI inference.

1:18:55The other thing they can really do is um cominatorial optimization which is this like this class of NP hard problems like the traveling salesman problem and things like that that they can efficiently they're not solving P equals NP but they can efficiently solve NP hard problems using this technique and one of the things that they were saying was um as I was reading was like you know in MRI scans the way the MRI scans work is like the patient has to go in >> and be there for like half hour to an hour Because what you got to do is you got to take a bunch of different little puzzle pieces of effectively what you're doing is you're you're taking like scans

1:19:35in um frequency space because of the way that the physics of the MRI works. You're you're trying to look at the procession of hydrogen nuclei in the patient as you like disturb it and then there's a magnetic field. So then the the the proton is going to relax and as it relaxes it releases light and you were trying to gather this light and that'll tell you some it'll give you like a frequency map >> of the patient and then from that frequency map you go into like a a positional map right um so in order to do that it takes a lot of scans in this forier space to then reconstruct the actual thing right but there's a way to do it with um something called um

1:20:16sensing right like Uh and what you can do is you can very selectively probe this frequency space. >> Okay. And do it in a very smart way with this compressed sensing. It's like a zip file for sensing. And um >> you can then you can then reconstruct the image. Yes. >> Right. So they could do that. That's a that's an optimization problem. >> Yeah. I get what you're saying >> at the end of the day. Right. And the compressed sensing is an optimization problem that they could then figure out and they projected that you know instead of a 30 minute to an hourong MRI it'll take a patient 5 minutes and then this thing can just like reconstruct it

1:20:57pretty quickly. What what's an important point here is that the the analog infrastructure the analog computer infrastructure that they've built has this application as sort of a replacement for GPUs at inference time in the AI use case uh because it's it's good at one aspect but it's also good at this combinatorial piece which gives it uh implications in the medical field for example in decreasing MRI times yeah because it's fundamentally so much faster and better at this very specific function that is a part of the MRI process >> that currently is using digital computers and if you just swap out the hardware infrastructure and a variety of

1:21:37other details to get it integrated into the MRI workflow >> you're now just able to actually just do more with less. >> Yeah. Yeah. And that's like super huge, right? Like like now I can do like 12 patients instead of one. Right. Right. >> That's so great. I love >> Yeah. So innovation is innovation is >> it's for everyone who said innovation is dead uh watch from first principles because we will uh dispel that rumor quickly. We touched on four very very fascinating stories this week. We started with three Atlas. Yep. >> We had some new imagery. Uh it's not aliens. It's still very interesting. Uh we may not live to 100. Uh so Brian

1:22:18Johnson, watch out. But longevity studies are showing that there is a real plateau being reached in this sort of increase in life expectancy. There's huge implications for policy um and economics related to how we shift to the zeitgeist opinion about that to be >> connected to the fundamental frontier research in that area. We don't have to sort plastic anymore. >> We talked about this new nickel catalyst that enables this no that was a fascinating one. And last but not least, analog AI from Microsoft uh which has potential to solve for the energy uh issue with AI but also has other applications. >> Yeah, we want to make AI better,

1:22:59>> right? >> And cheaper, cheaper, faster, better for the environment. >> Yeah. Who knows? Maybe it'll solve like our the problems, >> right? Right. >> If it also solves the MP problem, it might solve other problem types that we didn't we didn't have the tools for before. This is what we love to do here at From First Principles. Talk about this breaking science research. Break it down so you don't need a PhD to understand it. The implications here are crazy. My name is Lester Nar, joined as always by my co-host and our resident PhD, Christian Chowdery. We will see you all next week. [Music]

1:23:42Hey. [Music]