Elon Enters the Chip Race, the S&P 500 Repricing, and Human Drivers Will Become Illegal | EP #242

[0:00] Without question for me, the number one
[0:02] story this week was Elon's announcement
[0:05] of the Terafab.
[0:06] >> This is the most important endeavor in
[0:08] human history by far.
[0:11] In order to understand the universe, you
[0:12] must explore the universe.
[0:14] >> He's basically building a galactic
[0:16] [music]
[0:17] factory. On the left is 20 gigawatts.
[0:19] It's the current global output. And just
[0:22] the audacity of Elon's vision, [music]
[0:24] it has tremendous geopolitical
[0:26] implications as we discussed on the last
[0:28] pod. This could either accelerate or
[0:31] more hopefully [music] mitigate World
[0:32] War III and a Chinese invasion of
[0:34] Taiwan. We're going to need all the
[0:36] compute we can create. In fact, I'm
[0:37] actually kind of worried a self-driving
[0:39] car uses up basically a full GPU. When
[0:41] is it going to become illegal for
[0:43] [music] humans to drive? I think the
[0:44] thing that would make it later is purely
[0:47] the shortage of chips. Like the
[0:48] technology will be there and the demand
[0:50] will be there long before the chips are
[0:52] there.
[0:52] >> Figure out how to do more compute with
[0:55] less silicon
[0:56] for this exact use case and you'll be an
[0:58] instant [music] billionaire.
[0:59] >> I mean, we're heading towards a hundred
[1:00] trillion-dollar company. Maybe the
[1:02] largest, most important company [music]
[1:04] on and off the planet. Can we get there?
[1:09] Now, that's a moon shot, ladies and
[1:10] gentlemen.
[1:14] Everybody, welcome to Moonshots, another
[1:15] episode of WTF. Here with my incredible
[1:18] moonshot mates. DB2 in Boston. Uh AWG,
[1:23] looks like you're on your home base as
[1:24] well.
[1:25] I am, but without my saucer separated
[1:28] Enterprise 1701D behind me. Oh, yes.
[1:31] I've got I I contracted one of my boys
[1:33] to create finally
[1:36] finally LEGO's come out with a Star Trek
[1:38] uh you know, LEGO set. I'm tired of all
[1:41] the you know, Star Wars LEGO sets. So,
[1:43] yes, 1701D.
[1:46] Uh and it does do a saucer separation,
[1:49] but I'm not going to try it right now
[1:50] because a probably disaster may follow.
[1:54] And of course, we have Dr. EXO Saleem at
[1:57] his normal location at JFK. Saleem, how
[2:00] you doing, pal?
[2:01] >> [laughter]
[2:02] >> Salim is gone.
[2:04] Oh, okay. Well, so much for that.
[2:07] Yeah, it's as you know, listen, we try
[2:10] to be mobile. We're all dedicated to
[2:11] this podcast. But uh
[2:14] let's continue on because the
[2:16] singularity is not going to wait. So,
[2:19] today uh we're working to get you
[2:21] future-ready. A little bit of a format
[2:23] change. Uh we're going to be having some
[2:25] deeper conversations about a more
[2:27] limited number of subjects, still
[2:29] covering the news that's breaking right
[2:30] now and there is a lot. But our mission
[2:34] is to get you excited about the
[2:36] abundance that's coming, show you the
[2:38] opportunities that are coming to you
[2:39] whether you're an entrepreneur, an
[2:41] investor, a student, a parent. And
[2:43] really, you know, this is the time to be
[2:45] paying attention to the supersonic
[2:47] tsunami, the most important tech in the
[2:50] world. And hopefully, this is your
[2:53] number one uh podcast on AI and
[2:55] exponential tech as well.
[2:57] Hopefully, Peter, there's more in here
[2:59] than ever before. It's it's bundled into
[3:02] themes that we can discuss, but uh if
[3:04] you just look at the raw news story
[3:05] count, it's it's as you would expect,
[3:07] exponentially exploding. Uh you know,
[3:10] our goal, all of us, is to make sure
[3:12] that as we're talking about this on on
[3:14] Moonshots, that it's meaningful to the
[3:16] listeners, gets you excited, gives you
[3:18] context, helps you think about this in a
[3:20] different way.
[3:22] Um let's jump in. Without question for
[3:25] me, the number one story this week was
[3:27] Elon's announcement of the Terafab. Uh
[3:30] he's basically building a galactic
[3:33] factory. Uh think of this as putting all
[3:36] the parts of his LEGO puzzle together uh
[3:39] in an extraordinary fashion uh that is
[3:42] going to create massive capabilities.
[3:44] So,
[3:45] uh let me hit these uh these quick
[3:47] points and then we'll jump and discuss
[3:49] it. So, the Terafab is an objective
[3:51] across Tesla, XAI, SpaceX to build one
[3:55] terawatt of AI compute per year.
[3:59] To put this in context, uh the global
[4:01] output today is 20 gigawatts of AI
[4:04] compute. Again, we're measuring AI
[4:06] computation in terms of power, not just
[4:09] chips anymore. Uh so, Elon wants to
[4:11] build 50 times the current production
[4:14] rate of the planet. Uh he's building two
[4:17] kinds of chips, an edge inference chip
[4:19] for robots and cars, but also a
[4:23] high-power rad hard for his space uh
[4:26] Dyson sphere that is coming online. The
[4:28] fab is in Austin uh and it looks
[4:31] eventually like a hundred million square
[4:33] feet of capacity.
[4:35] Uh
[4:36] one terawatt in the near term, near
[4:38] term, you know, uh single-digit years.
[4:42] Uh long-term, a petawatt gets you uh
[4:44] only there from lunar mass drivers.
[4:47] Salim Ismail has joined the story. Hey,
[4:49] Salim, good to see you, pal.
[4:51] Hey, folks. Sorry, I'm bouncing around a
[4:53] bit, but I'm here.
[4:55] All right. In which terminal at JFK are
[4:57] you today? Where are you going?
[4:59] Uh I'm flying to Brazil for 40 hours.
[5:02] Of course, you are. Of course, you are.
[5:04] You're a probability function on planet
[5:06] Earth.
[5:07] So, just to put this in context, check
[5:10] out this chart. On the left is 20
[5:12] gigawatts. It's the current global
[5:14] output. And just the audacity of Elon's
[5:17] vision, uh a thousand gigawatts or a
[5:20] terawatt uh
[5:22] is his objective. I mean,
[5:24] you know, one thing I heard him say is,
[5:26] "Listen, I've been going to all the chip
[5:28] manufacturers out there and saying, 'I
[5:30] will pay you for as much production rate
[5:33] as you will give me. I I don't want to
[5:34] compete with you, but give me more and
[5:36] more and more.' And of course, none of
[5:38] them are moving at Elon speed. And so,
[5:40] he said, 'Screw it. Uh I'm going to go
[5:42] and build my own production facility.'"
[5:45] And not exactly what he did in the
[5:48] launch industry. Right? Just lapped the
[5:50] entire existing launch industry and the
[5:52] autonomous car and electric car
[5:54] industry. He's playing his playbook over
[5:56] and over again.
[5:58] Um before I get into this data and
[6:00] stats, uh comments, Dave. Well, this is
[6:04] the most important endeavor in human
[6:06] history by far.
[6:08] Cuz it unlocks everything else.
[6:10] And uh you know, no no great surprise
[6:13] that he's announced it at the scale that
[6:16] humanity needs it.
[6:18] But the specifics on how you're going to
[6:20] actually physically do this are unknown
[6:22] because
[6:23] there are
[6:25] fundamental constraints to the number of
[6:26] ASML machines, EUV machines to do this
[6:29] and and many many This is the most
[6:31] complicated product ever made by
[6:32] humanity.
[6:33] Uh and the supply chain is like like,
[6:36] you know, it makes cars look like
[6:37] child's play.
[6:38] So, uh
[6:40] uh
[6:40] so, he announced the mission. It's the
[6:42] right mission. The scale is crazy. It's
[6:44] you know, we estimated this on the last
[6:45] podcast at 50X all current productivity
[6:49] uh of chips and I I guess our estimate
[6:51] was dead on. So, so we got that part
[6:53] right. And he did allude to it last
[6:55] summer when we were meeting with him.
[6:57] And um what what I was most curious
[6:59] about is how he was going to announce
[7:00] this and attract all the talent that he
[7:01] needs
[7:02] without
[7:03] >> [clears throat]
[7:03] >> uh irritating Samsung. You know, cuz he
[7:05] signed a 16 billion-dollar deal for for
[7:07] production with Samsung, which is more
[7:09] like 45 billion if it is going according
[7:11] to plan.
[7:12] And um I guess one of the cover stories
[7:15] here is, "Well, these chips are for cars
[7:17] and they're also for space. They're
[7:19] hardened for space. So, they're not like
[7:21] the other chips."
[7:22] >> he's he said, "I will buy everything
[7:24] Samsung can offer me.
[7:26] But you're not offering me enough. So, I
[7:29] will still build all these chips and I
[7:30] will still buy everything you want to
[7:32] give me." You know, one thing I love and
[7:34] he pointed out in his Austin fab
[7:37] is that it's full vertical integration
[7:40] under one roof. So, that he can run
[7:43] rapid iterations on chip design.
[7:46] That's impressive.
[7:48] >> in our in our Austin podcast when we
[7:50] were talking to Elon,
[7:51] uh I asked him point-blank, uh you know,
[7:53] TSMC is being way too conservative in
[7:56] terms of their production of chips. They
[7:58] should be 10Xing
[8:00] their fab manufacturing. He said, "Well,
[8:02] you know, they're you know, the industry
[8:04] is cyclic. You know, maybe they're being
[8:06] conservative intelligently."
[8:07] >> [laughter]
[8:08] >> Which is hilarious in hindsight if you
[8:10] go back and listen to that audio because
[8:11] in the back of his mind, he's like,
[8:12] "Well, I'm going to build something 50
[8:13] times bigger
[8:14] >> [laughter]
[8:14] >> anyways." So,
[8:16] but uh yeah, it is crazy that Samsung,
[8:18] uh Intel, and TSMC are not racing to
[8:23] build, you know, 10, 20X more
[8:26] production.
[8:27] So, Elon, of course, is well, he's going
[8:29] to do it instead.
[8:31] Okay,
[8:32] can I show you guys some calculations
[8:33] that I found just extraordinary here?
[8:36] Um
[8:38] listening to the presentation he gave uh
[8:41] 48 hours ago.
[8:43] You know, his target is one terawatt of
[8:44] compute per year in orbit.
[8:47] Uh and he said, "Mass to orbit, 10
[8:49] million tons per year." Uh we're talking
[8:52] about an average satellite, uh his
[8:54] next-generation uh Starlink at a ton.
[8:58] Long story short,
[8:59] in order for him to launch that much
[9:01] capacity, it's 274 launches per day
[9:06] on Starship. It's a launch every 5.3
[9:10] minutes. Which, of course, he says,
[9:12] "Listen, in the airline business, that's
[9:13] normal." But just the audacity and the
[9:16] level of uh of you know, thinking that
[9:19] Elon takes on uh is amazing. Uh
[9:23] AWG, you want to you want to jump in? So
[9:26] many thoughts on this. Well, first of
[9:27] all, I think the elephant in the room is
[9:30] if Elon can indeed ramp up capacity for
[9:32] the Terafab in, call it, the next five
[9:35] years, which is the time scale that's
[9:37] being tossed around, it has tremendous
[9:39] geopolitical implications as we
[9:41] discussed on the last pod. This could
[9:43] either accelerate or more hopefully
[9:45] mitigate World War a Chinese invasion of
[9:48] Taiwan.
[9:49] If you look at the the moon aspirations,
[9:52] the the lunar aspirations for not a
[9:54] terawatt, but a petawatt of of from the
[9:57] moon. If you do the back of the envelope
[9:59] arithmetic for what would a a pedawatt
[10:02] of GPU compute that comes from lunar
[10:05] mining take, you run the arithmetic,
[10:07] comes out to be approximately 3/100,000
[10:12] of the lunar mass.
[10:14] So, a pedawatt coming from lunar mining
[10:17] with electromagnetic launches from the
[10:19] moon is starting to have a material
[10:22] impact on the mass of the moon. So, this
[10:24] is
[10:24] >> one big one big crater dug out of the
[10:26] moon.
[10:27] >> [laughter]
[10:27] >> So, that that's that's a pedawatt. As we
[10:30] scale, of course, to an exawatt of
[10:33] compute, and because why not, then at
[10:36] that point we're talking something like
[10:39] 3%
[10:40] of of the moon's mass. And this is, you
[10:43] know, when people think I'm joking when
[10:46] I talk about disassembling the moon or
[10:47] the moon had it coming,
[10:49] it this certainly paints a portrait. The
[10:50] moon did indeed have it coming, and it
[10:54] the moon is is slated for disassembly to
[10:57] to build the Dyson swarm. This is what
[10:58] it looks like. I I think
[11:01] more broadly, there are other sort of
[11:03] secondary implications, really
[11:05] interesting, that this is a joint Tesla
[11:07] xAI SpaceX maneuver. And many folks have
[11:11] speculated over the years, wouldn't it
[11:13] be wonderful if all of Elon's
[11:15] industrial ecosystem came together into
[11:17] one singleton. This is sufficiently
[11:20] cruxy with 20% of its production slated
[11:23] for Tesla and 80% slated for SpaceX,
[11:27] that this starts to look a little bit
[11:28] like maybe a cornerstone for some grand
[11:31] unification of all of Elon's projects.
[11:33] >> you, Alex, we talked about in a previous
[11:35] podcast, the idea that, and Elon said
[11:38] this, we'll see the first hundred
[11:39] trillion dollar company. And when we
[11:41] look at the numbers here, I want to show
[11:43] another set of calculations I did on
[11:45] what might the terrafab be worth in the
[11:48] ecosystem. I mean, we're heading towards
[11:50] a hundred trillion dollar company, and
[11:52] can we get there? And And at the day, my
[11:56] I don't know how you guys feel, but the
[11:58] Musk world ecosystem here looks like it
[12:02] will lap by an order or two of magnitude
[12:06] what Nvidia's done.
[12:08] Um
[12:09] It may be the largest most important
[12:11] company on and off the planet. Yeah, and
[12:13] I don't think Elon wants to unify all of
[12:15] his projects just for the sake of having
[12:18] one unified company. I think he wants to
[12:20] unify the the capital raising and the
[12:21] capital leverage with his massive
[12:23] multi-trillion dollar IPO
[12:26] and the massive joint mission unlocking
[12:29] an unprecedented amount of capital,
[12:31] which is what it's going to take to do
[12:32] these fabs in parallel at this scale.
[12:35] Cuz that's what That's the thing that's
[12:36] holding back Samsung and Intel and TSMC.
[12:39] They actually could do it. He needs 25
[12:41] billion initially
[12:43] to turn on the terrafab and and get it,
[12:45] you know, get it the buildings started,
[12:47] so to speak. Right. I saw that in the
[12:49] analysis, but 25 billion is is just one
[12:51] fab.
[12:52] And here we're going to 50x the US or
[12:55] the world production. 50x the world
[12:57] production. So, he needs 50 of those 25
[12:59] billion dollar investments to to achieve
[13:00] this mission. Yeah, there's a there's a
[13:02] lot to do. I thought I thought I had two
[13:04] three thoughts. I mean, one is like talk
[13:06] about patron saint of exponentials,
[13:08] right? Like this guy thinks at scales
[13:11] that very few people do. And it sounds
[13:12] incredible. It's classically the future
[13:14] of anything he's looking at looks
[13:16] vertical, and the past looks like flat
[13:18] and boring. Um what I what I thought was
[13:21] great was this is like amazing XLR logic
[13:24] cuz he's going exponential at the
[13:26] bottlenecks.
[13:27] You you stop competing. You're competing
[13:29] You're just redefining the game. And
[13:31] you're challenging anybody to dare to to
[13:33] come with you. I think that's like the
[13:36] amazing
[13:37] uh part of this. The launch cadence is
[13:40] is unreal, every 5.0 minutes. And I
[13:42] think that's exactly right. It forces
[13:44] the operating model to completely
[13:45] change. And it's forcing everybody to
[13:47] rethink that, including all the
[13:49] engineers and all the infrastructure,
[13:51] etc. If this is done in any kind of
[13:53] normal industry, um one thing to point
[13:56] out is that his predictions on timing
[13:58] tend to be about 15 to 20% accurate. So,
[14:02] you know, but but it doesn't matter if
[14:04] he's if it takes him three times as
[14:06] long, who the hell cares? Like the fact
[14:07] that
[14:08] >> thinking at the scale and he'll get
[14:10] there is the fact that he's putting his
[14:12] planting up Yeah. Yeah, he's
[14:14] directionally correct. idea of like you
[14:15] shoot for the stars, and if you get to
[14:16] the moon, who the hell cares? We're
[14:18] landing somewhere amazing.
[14:20] >> [laughter]
[14:21] >> That's A W S G's plan. So, this is
[14:23] Here's the next question. a a
[14:25] a rapid schedule to disassembly of the
[14:27] moon, I think is what's on the table.
[14:30] >> [laughter]
[14:32] >> Okay. Good thing I don't [clears throat]
[14:33] have a glass of wine. Everybody, you may
[14:35] not know this, but I've got an
[14:36] incredible research team. And every
[14:38] week, myself, my research team study the
[14:40] meta trends that are impacting the
[14:41] world. Topics like computation, sensors,
[14:44] networks, AI, robotics, 3D printing,
[14:46] synthetic biology. And these meta trend
[14:48] reports I put out once a week enable you
[14:51] to see the future 10 years ahead of
[14:53] anybody else. If you'd like to get
[14:55] access to the meta trends newsletter
[14:56] every week, go to
[14:58] diamandis.com/metatrends.
[15:00] That's diamandis.com/metatrends.
[15:03] So, a terawatt of compute per year, if
[15:06] physically achievable,
[15:08] is this aspirational? What time frame?
[15:11] You know, he says five years. Uh
[15:14] that's still 50x is is crazy, but if
[15:17] he's able to achieve that,
[15:19] you know, what happens to the you know,
[15:21] to the terrestrial data centers?
[15:23] And to the investments made in
[15:24] terrestrial data centers.
[15:26] Uh you know,
[15:27] gentlemen, questions on that? Uh every
[15:30] every chip, every investment in power
[15:32] and data centers is going to pay off
[15:33] tremendously. They won't cannibalize
[15:34] each other. We're going to need all the
[15:35] compute we can create and so much more.
[15:38] In fact, I'm actually kind of worried
[15:40] that the self-driving car is going to
[15:42] get cannibalized. You know, driving a
[15:44] self-driving car uses up basically a
[15:45] full GPU. And by the end of this year, a
[15:48] full GPU can also do brain surgery, or
[15:50] it can discover new math or new physics.
[15:53] And it's not clear that driving somebody
[15:55] around is going to make the price cut
[15:58] uh as the demand for compute goes to
[16:00] near infinity. So, I think the
[16:03] terrestrial data centers are going to be
[16:04] critical for national security for every
[16:06] country in the world because if
[16:08] something goes wrong in space, you got
[16:10] to fall back Your whole society will be
[16:11] running on these GPUs. Yeah, you can't
[16:13] you can't have an outage. So, anyone
[16:15] investing in this does not have to worry
[16:16] about one thing cannibalizing the other.
[16:19] And you The other thing you're going to
[16:20] see, I think it's later in the deck, but
[16:22] all the different um
[16:24] process nodes All the fab process nodes
[16:26] are going to get used, even the older
[16:27] ones, the 3 nanometer and the 5
[16:29] nanometer, are going to be running full
[16:31] throttle now, even if they're not as
[16:32] good as the new two and 1.6 nanometer.
[16:35] It doesn't matter. We need all the
[16:37] compute we can crank out.
[16:39] So, yeah, that it's it's going to be
[16:41] just an all hands-on-deck race, and Elon
[16:43] is is just documenting the up upper
[16:46] bound of what we can achieve as a as
[16:48] humanity. My expectation
[16:51] My expectation would be that in the
[16:52] process of I think Elon likes to call it
[16:55] production hell. In in in the process of
[16:57] production hell for realizing the
[16:59] terrafab over the next five years, I I
[17:02] strongly suspect we're going to discover
[17:03] some new not semiconductor physics, but
[17:06] more material physics and process
[17:08] engineering. It it seems improbable to
[17:11] me that Elon will just build the
[17:13] terrafab based on the existing stack as
[17:16] say TSMC did of ASML plus the existing
[17:20] optics plus all of the the conventional
[17:23] semiconductor processing techniques. If
[17:25] he really is looking to disrupt the
[17:27] space, he's going to want much more
[17:29] disruptive unit economics. So, maybe
[17:31] some of these technologies for
[17:33] semiconductor production and and
[17:35] fabrication that have been waiting in
[17:37] the wings for their right time in in the
[17:39] light, maybe uh this is purely
[17:41] speculative, maybe he'll look, for
[17:43] example, at alternatives to
[17:45] photolithography. The we It's not like
[17:47] as civilization, we don't have lots of
[17:49] alternatives. Exactly. And he's got
[17:52] and he's got superintelligence to help
[17:55] him get there, to design new systems.
[17:57] >> going to need the humanoid robots to be
[17:58] building the fabs. We don't have the
[18:00] workers.
[18:01] Salim?
[18:03] I think this is a really important point
[18:04] being made right now, Alex. Thank you
[18:06] for this, because, you know, you think
[18:07] about the secondary technologies and the
[18:09] benefits, like like all the carbon fiber
[18:12] that came from the space industries and
[18:14] cascaded down to everyday life. The
[18:16] secondary inventions that will be needed
[18:18] here will be massively beneficial to
[18:20] humanity. Well, this this is one of the
[18:21] most exciting things in tech history, in
[18:24] fact, the most exciting thing in tech
[18:26] history is what Elon was talking about
[18:28] in Austin, is laying down single atoms
[18:30] using a some kind of a self-organizing
[18:31] process.
[18:32] And I feel like Alex is exactly right.
[18:35] Something will get discovered in the
[18:37] next year or two using current LLM AI
[18:40] running on GPUs that will then dictate a
[18:42] very different non-lithography future.
[18:46] But it'll probably be five or six or
[18:47] seven years before those start getting
[18:49] manufactured at the same scale as
[18:51] lithography.
[18:52] Uh but it's super exciting to to watch.
[18:55] Dave, I asked Claude to give me an
[18:56] estimate based upon all the data that we
[18:59] got from Elon during his his talk on the
[19:02] value of future terrafabs, right? And
[19:04] so, it makes the point here that initial
[19:06] capex, as you said, Alex, you know, it's
[19:09] 20 25 billion, but the real capex for
[19:12] buildout is going to be on the order of
[19:13] 150 billion dollars
[19:15] um at a minimum. It may be a half a
[19:18] trillion
[19:19] there. And then he talk And then the the
[19:21] model here looks at what's the annual
[19:23] cost savings for his captive
[19:25] opportunities, right? So, if he's
[19:27] building the chips himself and he's
[19:28] putting them in Optimus and then
[19:30] cybercabs, and then there's external
[19:33] revenue, and then there's an applied
[19:35] enterprise value. And, you know, it's on
[19:38] the order of a trillion to multiple
[19:40] trillions. TSMC is valued at 1.7
[19:43] trillion, and what we talked about last
[19:45] week was that terrafab is expected to
[19:48] produce on the order of 70% of TSMC's
[19:52] output.
[19:53] So, yet we're layering on another
[19:55] multi-trillion dollar opportunity. These
[19:58] numbers seem low to me. If you're really
[20:00] generating Yeah, if you're generating
[20:02] 50x Yeah, 50x the total output of AI
[20:05] chips on the planet.
[20:07] Yeah. Uh
[20:08] this this is operating
[20:10] This is operating a different still a
[20:11] car company and it's not.
[20:13] Yeah. Well, not only that. I mean just
[20:15] right out of the gate, this is not just
[20:16] doing what TSMC does. This is TNC plus
[20:19] Nvidia.
[20:20] You know, Nvidia's worth four and a half
[20:22] trillion. Yeah. But but I mean even
[20:23] that's ridiculous as an analysis. We're
[20:25] talking about 50 times the production of
[20:26] the world's current compute.
[20:28] Uh so, you know, out of the gate you
[20:30] would take TSMC plus Nvidia and multiply
[20:33] by 50 to get uh you know, a
[20:35] a starting point for an estimate. So,
[20:37] this is off by
[20:38] over an order of magnitude, well over an
[20:41] >> Any of you Any of you concerned about a
[20:43] monopoly here?
[20:45] No. If you're following the the
[20:47] prediction markets for the SpaceX IPO,
[20:50] this is already starting to get priced
[20:52] into the SpaceX IPO. So, SpaceX IPO was
[20:54] originally going to be 1 and 1/2
[20:56] trillion. Now prediction markets favor 2
[20:59] plus trillion dollar SpaceX IPO pricing
[21:01] in the SpaceX portion of the terrafab.
[21:03] So, in in some sense again, I'm not
[21:06] quite clear on what the government
[21:07] structure is here is going to look like
[21:09] and and how clean it's going to be, but
[21:11] to the extent it falls mostly in the
[21:13] SpaceX bucket,
[21:14] SpaceX IPO, not an investment advice,
[21:17] obviously, could end up being as you
[21:19] say, SpaceX plus Starlink plus Nvidia
[21:23] plus ASML plus TSMC all rolled into one.
[21:27] I I just as a as a piece of advice for
[21:30] entrepreneurs out there, um just
[21:33] understand the level of audacity that
[21:35] Elon is looking at. He's building in in
[21:39] a um
[21:40] >> [clears throat]
[21:40] >> sort of multiple orders of magnitude
[21:42] beyond anybody else. And from his first
[21:44] principle thinking, he's looking at
[21:46] where are the blockages for my growth.
[21:50] And we had this conversation. People are
[21:52] not generating enough chips. I need to
[21:54] build a chip fab. And then he doesn't
[21:56] just go out to say I'm going to buy
[21:58] Intel or build a chip fab equivalent to
[22:01] what TSMC is building in Arizona. No, if
[22:04] I'm going to build a chip fab, I'm going
[22:05] to build something that is 50 times
[22:07] bigger than the world supply.
[22:09] Um amazing.
[22:11] Well, also he's thinking two moves ahead
[22:12] in the big chess game and you know, two
[22:14] moves used to be 20 years. Now it's six
[22:16] years, but but two moves ahead of
[22:18] everybody else and this came up. Alex
[22:20] and I were talking earlier this week
[22:21] about fusion energy
[22:23] and you know, Elon doesn't talk much
[22:25] about fusion energy. Why? Well, because
[22:27] he's visualizing solar in space
[22:29] and and you know, solar on Earth is a
[22:31] great stepping stone to solar in space
[22:33] and it requires panels and batteries and
[22:35] cooling,
[22:36] but it doesn't require turbines and and
[22:39] fusion reactors. So, he can skip a
[22:41] couple of hard steps and go straight
[22:43] after the next move in the big chess
[22:44] game. So, it's really interesting to
[22:46] watch how those those timelines have
[22:48] inverted. You know, the Dyson sphere is
[22:50] now right on our radar and even Google
[22:52] is talking about it, which means
[22:54] everyone's talking about it. So, it that
[22:56] came to the forefront really just in the
[22:57] last month or two and so the whole
[22:59] timeline of humanity got shifted. So,
[23:02] the Dyson sphere will come before anyone
[23:03] even figures out how to get licensed for
[23:05] a fusion reactor.
[23:07] Can I be cynical Can I be cynical just
[23:08] for a second? Um Yes. The plans are
[23:11] ridiculously grandiose and if anyone is
[23:13] going to achieve them, it's Elon, but uh
[23:16] curious that the timing of this is just
[23:18] leading into the IPO to get everybody
[23:20] excited about things. So, that would be
[23:22] the cynical view, but I still I still
[23:23] just love the audacity.
[23:26] Yeah. Well, I think Saleem, your your
[23:27] first observation was dead right. If you
[23:29] shoot for the Mars and you end up at
[23:30] moon, you're still a way up.
[23:32] >> [laughter]
[23:33] >> So, I I want to ask the I want to ask
[23:34] the question again that AWG said,
[23:37] "Nope." Which is
[23:39] monopoly concerns. Do you believe I mean
[23:41] if he's really generating, you know, I'm
[23:44] not worried about it because when you
[23:45] have an MTP, which he does, you're
[23:47] basically operating on this massive
[23:50] mission. You may have ethical issues
[23:52] here and there, but generally the trend
[23:53] is so positive and so beneficial for
[23:55] humanity, who the hell cares?
[23:58] Well, I mean humanity becomes There's
[24:00] always another person.
[24:02] So, somebody who's like Brendan foodie
[24:04] maybe. You know, somebody like that
[24:05] right now is like, "Wow,
[24:07] I'm going to do this, too."
[24:09] And you know, it's just the way it is
[24:11] and that person, we don't know who they
[24:13] are yet, but they'll emerge and you
[24:15] know,
[24:16] you can't exist in the US without
[24:18] antitrust action if you don't have a
[24:19] competitor. So, Elon will invite that
[24:21] competitor, whoever it is and and it'll
[24:24] be great. I'll say one of the things
[24:26] here. If
[24:27] If we label this monopolistic behavior,
[24:29] then don't we have to label everyone
[24:30] with an MTP a monopolist? Mhm. Mhm.
[24:34] Just asking the question, my dear mates.
[24:36] Market dominance if you if you would get
[24:38] there, then that's true.
[24:40] But you haven't even picked the real
[24:42] estate site yet [clears throat] for the
[24:43] the terrafab. It doesn't He hasn't even
[24:45] picked a final location. I I think it's
[24:47] way too premature to declare this
[24:49] monopolistic behavior being so ambitious
[24:51] as to build the Dyson swarm. We're going
[24:52] to have multiple Dyson swarms.
[24:54] >> I'm with I'm with you. Love it. I'm with
[24:55] you. All right.
[24:56] Well, well, look, Google Google isn't
[24:58] going away. You know, Google has 300
[25:00] billion dollars of revenue, 100 billion
[25:02] plus of cash flow, their own chips and
[25:05] design. Uh they have all everything
[25:08] except the rockets.
[25:09] So, uh Google's not going to go away as
[25:11] a Well, I don't forget Eric Schmidt Eric
[25:14] Schmidt is trying to bring Relativity
[25:15] Space online. So, that the rockets are
[25:17] at least part of Google family. Oh, you
[25:19] know, they're going But let me let me
[25:20] just close Let me close this
[25:22] There there's plenty of them to go
[25:24] around. Let's not forget.
[25:26] One thing that's important here is every
[25:28] time there is a constraint, uh the judo
[25:31] move is to realize it's a massive
[25:33] opportunity. And so, this is an abundant
[25:35] story once again. This is a a massive
[25:38] increase in abundance of AI compute uh
[25:42] beyond what anyone was speaking about
[25:44] just a week ago.
[25:46] So, I'm glad he's focused on this and
[25:49] less on the politics.
[25:51] Mhm. Yes. I'm in.
[25:54] All right. Here is our second
[25:56] conversation story, one that I'm excited
[25:58] to have with my moonshot mates. It's
[25:59] about the future of human
[26:01] transportation.
[26:02] You know, robots are getting their
[26:03] driver's licenses, flying cars are
[26:05] taking flight. Uh here's some of the
[26:08] data and uh I want to go deep on this
[26:10] because I want everyone listening to
[26:12] understand
[26:14] how this is going to impact how and
[26:16] where you live, how you commute, every
[26:18] aspect of our lives. So, Waymo uh hit
[26:21] 170 million fully autonomous miles,
[26:25] uh equivalent to 200 human lifetimes of
[26:27] driving
[26:28] with 92%
[26:30] fewer serious crashes. So, a significant
[26:33] reduction uh in crashes. At current,
[26:36] they've got 3,000 vehicles in 10 cities.
[26:38] Still early, right? Uber has now
[26:42] invested 1 and 1/4 billion dollars in
[26:44] Rivian with plans to deploy 50,000 fully
[26:47] autonomous robo-taxis.
[26:50] Here's a look at Waymo versus human
[26:51] drivers. Um Waymo is doing an
[26:55] extraordinary job of of saving uh
[26:58] not necessarily lives, but saving
[27:00] crashes, uh you know, minus 92%.
[27:05] So,
[27:06] I think cyber cab, we've seen incredible
[27:08] data like this also on full self-driving
[27:11] from from Tesla.
[27:13] Check out this image. This is Joby
[27:15] Aviation. Joby had started this company.
[27:17] He started Velocity 11, took that money
[27:20] after he sold it. Uh he was partnered
[27:22] with Rob Nail, uh Saleem. And uh Joby
[27:25] started Joby Aviation God knows over a
[27:28] decade ago and here it is flying over
[27:31] the Golden Gate Bridge in San Francisco.
[27:33] It's a a beautiful image. This is a
[27:35] eVTOL,
[27:37] electric vertical takeoff or landing.
[27:39] It's a name that rolls off the tongue
[27:41] onto the floor. I'm calling them flying
[27:43] cars cuz that's what they are. And
[27:47] here's what's going on in the eVTOL
[27:49] world. I think really important. So,
[27:51] Joby uh just is now in testing for its
[27:55] first FAA conforming aircraft, meaning
[27:58] it's
[27:59] it's demonstrating to the FAA that it
[28:01] can build a reliable design over and
[28:04] over again. It just had demonstration
[28:06] flights on the Golden Gate Bridge.
[28:08] Joby and Uber announced Uber Air powered
[28:12] by Joby.
[28:13] In fact, when uh when Travis left
[28:17] was before he left, he had created
[28:19] something called Uber Elevate and they
[28:21] were doing their earliest work on flying
[28:24] cars.
[28:25] Uh I keynoted their their talk there,
[28:27] but Uber Elevate got sold to Joby
[28:30] and now Joby and Uber have a
[28:31] partnership. The other company in the
[28:33] United States that is a competitor to
[28:36] Joby is called Archer Aviation. Uh they
[28:38] have a beautiful aircraft called uh
[28:40] Midnight and they're the first company
[28:43] to achieve 100% FAA acceptance of its
[28:45] eVTOL uh aircraft uh means of
[28:49] compliance. Long story short, we've been
[28:51] waiting a long time and flying cars are
[28:54] almost here. We're going to start to see
[28:56] them operating in the US in the next uh
[28:59] 18 months. Should be here in LA in 2028
[29:03] in a big way.
[29:04] Uh and so, here are some of the
[29:07] conversations, gents.
[29:09] Um first off, when is it going to become
[29:11] illegal for humans to drive? Saleem?
[29:14] Yeah.
[29:17] I think pretty you know, you'll start
[29:18] with city centers, right? And you'll be
[29:20] illegal to drive in the city centers.
[29:22] Then it'll slowly broaden out from
[29:23] there. I think what's I'm I'm the flying
[29:26] car is the most exciting technology for
[29:28] me personally given that I'm commuting
[29:30] to airports a lot that I could ever ask
[29:32] for. So, this is this is 10 years later
[29:35] than I wanted it, but finally it's
[29:36] finally happening. What I like about
[29:38] this These are not transportational
[29:39] stories. This is full urban redesign is
[29:42] what the narrative is. Essentially, you
[29:44] make Our becomes abundant. Now, land has
[29:47] always been scarce. Real estate has been
[29:49] scarce. Real estate becomes abundant. If
[29:51] you fly across the US, it's empty.
[29:54] Right? And we've talked about the
[29:56] statistic just around
[29:58] between Toronto and Chicago airports,
[30:00] there are 10,000 islands and lakes,
[30:02] right? So, we do not have a scarcity
[30:04] problem. We have a mobility and
[30:05] accessibility problem. So, I think I'm
[30:07] super excited by this particular model.
[30:10] I've got my 2 years for
[30:13] us to get to full autonomy before my son
[30:15] gets his driver's license. He's 14 right
[30:17] now. So, I'm I'm pushing hard on this
[30:19] race. Mostly, he wants to get it to get
[30:21] away from parents, but that's fair. Um
[30:23] so, we'll see what happens. But, you
[30:25] know, you compress travel time, you
[30:26] re-price real estate. This is such a
[30:28] huge thing. And I got to just shout out
[30:30] to Joe Ben just because it's hard to
[30:33] build a hardware platform like this and
[30:35] to do it over a decade with all the
[30:37] inevitable regulatory and market
[30:39] structure against you and infrastructure
[30:41] against you. This is a huge This is like
[30:43] a Nobel Prize in patience here.
[30:46] Yeah.
[30:47] Incredible.
[30:48] Dave, you were going to say?
[30:50] Oh, the EVTOLs are going to move very
[30:52] quickly
[30:54] cuz they don't they don't run the risk
[30:56] of you know, crashing into houses like
[30:58] cars on a self-driving cars on a road
[31:00] do. They're all going to be autonomous
[31:02] on birth. That that's the new thing.
[31:04] EVTOLs have been in the works for years,
[31:05] but the AI that makes them self-flying,
[31:08] self-driving, and super safe is here all
[31:10] of a sudden now.
[31:11] >> True, but the first airplanes are going
[31:13] to be piloted, right? They'll be single
[31:14] pilot, four passengers in the back. The
[31:18] the
[31:19] goal is rapid recharge at the vertiports
[31:21] when they land recharge. Probably an
[31:24] average length of flight of under 10 km,
[31:27] I think. Uh you know, going from Santa
[31:30] Monica where I am to the Dodger Stadium
[31:32] and avoiding
[31:33] uh the 10.
[31:34] But, autonomy will come with enough with
[31:37] enough data and enough demonstrations.
[31:39] Wait, why why won't they be Why won't
[31:41] they fully autonomous from the get-go? I
[31:43] mean, it's there's Cuz the It's called
[31:45] the FAA.
[31:47] FAA. The FAA is not happy until you're
[31:50] not happy.
[31:51] >> [laughter]
[31:53] >> Yeah, that's exactly it. The
[31:54] manufacturing of these wants to happen
[31:56] right away.
[31:57] And the the AI command and control is
[32:00] being worked on for the car, not for the
[32:02] EVTOL yet. So, there'll be a couple year
[32:04] very short period of time, in my
[32:05] opinion, 2 years or so. Cuz you know, in
[32:07] in the Middle East, they're already
[32:08] doing the self-driving, self-flying
[32:10] version of this. So, should be very
[32:12] short window where people get to fly
[32:14] these. Uh you know, the bullet here
[32:16] though is when does it become illegal
[32:17] for humans to drive? I think that's
[32:18] going to happen very quickly as well. Uh
[32:21] very similar to indoor smoking
[32:24] um or drunk driving.
[32:26] There's a tipping point where a lot of
[32:28] voters
[32:30] say, "Wait, you're putting my children
[32:32] at risk with your crappy driving."
[32:34] Mhm. That that's ridiculous. We got data
[32:37] and proof here that the the self-driving
[32:39] is 90% safer, [clears throat] soon to be
[32:41] 95, 97% safer.
[32:44] And and you know, the human tragedy that
[32:45] comes from car crashes is is
[32:47] unbelievable, shocking.
[32:49] And so,
[32:50] >> for under 5-year-old under 5-year-old
[32:52] kids, it's the number one cause of
[32:53] death.
[32:55] Yeah, accidents. Yeah. Oh, and and you
[32:57] know, the the it's devastating to
[32:58] families, too. It's just absolutely
[32:59] tragic.
[33:00] >> in the first world. Yeah.
[33:02] Well, there'll be a there'll be a TV ad
[33:03] campaign probably 3, 4, 5 years from now
[33:06] with lots of ugly images in it. And then
[33:08] there'll be massive amounts of voting,
[33:10] and then people will say it's it's
[33:12] inconceivable that you would drive on a
[33:13] public road. That's inhumane. Go drive
[33:16] on a test track, that's fine. Maybe you
[33:18] know, some country roads, that's fine.
[33:21] But, uh but no way, don't put my
[33:23] children at risk. So, I I think that's
[33:24] going to come as soon as we have the
[33:26] manufacturing for the cars themselves.
[33:29] But, I I think the thing that would make
[33:30] it later is purely the shortage of
[33:33] chips. Like the technology will be there
[33:35] and the demand will be there long before
[33:37] the chips are there.
[33:38] So, if you want to unlock this as an
[33:41] engineer, figure out how to do more
[33:44] compute with less silicon
[33:46] for this exact use case and you'll be an
[33:48] instant billionaire.
[33:50] Alex, your thoughts?
[33:51] >> this format, Peter. This is like an
[33:53] internal AMA. So, I'm going to try for a
[33:55] lightning round on on all of these.
[33:57] Oh,
[33:58] >> Leave some Leave some room for Leave
[33:59] some room for the rest of us. Let's take
[34:01] it one at a time.
[34:02] One at a time. At what point does it
[34:04] become illegal for humans to drive? I
[34:06] think never. I think we'll simply
[34:08] redefine driving to represent higher and
[34:10] higher levels of abstraction. So, right
[34:12] now with like FSD14, you tell it where
[34:15] you want and if if you're running the
[34:17] most recent subversion, you can have a
[34:19] conversation with Grok and you can do
[34:21] minute steers along the way. I think
[34:23] that notion will get refined such that
[34:24] driving gets redefined to be
[34:26] sufficiently abstract that it's always
[34:28] safe for pedestrians. It's always the
[34:30] human on the loop of the AI driver. So,
[34:34] it's effectively a human-machine hybrid,
[34:36] if you will, that has the safety of the
[34:38] machine, but makes the human feel like
[34:40] they're in the driver's seat still.
[34:41] >> I I said I said this with when Dara was
[34:44] on stage with Salim and myself. I said
[34:45] there's a version in the future of
[34:47] self-driving where you're driving
[34:49] and you can push the car as fast and as
[34:51] hard as you want and the car knows its
[34:53] own limits. It knows the traction of its
[34:55] tires. It knows the road surface and it
[34:57] prevents you from doing something
[34:59] stupid, but you're in control of it 99%
[35:02] of the time, but the 1% where you're
[35:03] about to do something that will destroy
[35:06] you, a person, or the car, it stops you.
[35:09] Exactly. I think like the the future the
[35:11] future of the accelerator pedal isn't
[35:13] the accelerator pedal. It's if you use
[35:15] FSD, it's turning the driving mode up to
[35:18] Mad Max. That's sort of like a an
[35:20] abstraction of acceleration.
[35:22] >> that's that's all I use is Mad Max and
[35:24] it still doesn't go fast enough. So, I
[35:25] have to step on the pedal.
[35:28] >> There used to be an ad saying, "Friends
[35:29] don't let friends drive drunk." And so,
[35:32] you can just keep that ad and drop off
[35:33] the drunk part and go, "Friends don't
[35:35] let friends drive." Period. So, all the
[35:37] all the messaging is there.
[35:39] All right. So, so AWG, why don't you Why
[35:41] don't you kick us off on question two
[35:43] here?
[35:44] Okay. Question two, with Uber partnering
[35:46] with Waymo and a bunch of other names,
[35:49] will the cybercab be able to compete? Uh
[35:52] I think we mean compete here.
[35:53] Yes, of course. Uh it's going to be very
[35:56] competitive market. Period.
[35:58] Yeah. And I love the fact that this is
[36:00] driving us towards abundance, right?
[36:01] This is driving us towards UHI. If
[36:04] you've got a dozen companies delivering
[36:07] autonomous uh
[36:09] uh vehicle services in your city,
[36:11] they're going to be competing against
[36:13] quality of service and price and just
[36:16] bringing the price down to a minimum
[36:17] amount. Now, one of the things that's
[36:19] interesting about cybercab is that
[36:22] that's going to be priced at probably uh
[36:24] 30K is what roughly what Elon's
[36:26] announced. And he's going to allow
[36:27] people to buy it. So, uh you know, one
[36:30] of my goals is can I buy, you know, 25
[36:33] or 50 of them here in Santa Monica and
[36:36] own them, but have them going out and
[36:39] and basically generating revenue for me
[36:42] and for my my cybercabs. I'm sure
[36:44] they'll have some level of
[36:46] of personhood by then, Alex. Uh I think
[36:49] so. I I I
[36:49] I never would have guessed, Peter, that
[36:51] that your next gig would be as a cabbie,
[36:53] but the singularity makes for strange
[36:55] bedfellows. Fleet owner. Fleet owner.
[36:58] Um Um I'm I'm a comedian and I I think
[37:00] >> Yeah, please.
[37:01] >> the big impact for me when I see this is
[37:03] the complete collapse in the market
[37:06] structure of cars. Today, we make close
[37:08] to 100 million new cars a year.
[37:10] And they sit empty 94% of the time. So,
[37:14] even if you drop that by 50% in utility,
[37:17] um
[37:18] um you you basically collapse the need
[37:21] for past the car industry instantly. And
[37:24] if these cars maintain for a long time,
[37:26] the lifetime should be near infinite. My
[37:29] Tesla Model 7 2017 should it'll go a
[37:31] million miles. There's nothing wrong
[37:33] with that car. So, this is going to
[37:35] completely change the nature of car
[37:37] services, car
[37:39] maintenance people, like the complete
[37:40] industry gets reshuffled from the bottom
[37:42] up.
[37:44] Yeah. Yeah. We'll think about the
[37:45] implications of that, too, Salim. Right
[37:46] now, if you take an Uber from SFO to San
[37:48] Fran for like 200 bucks or whatever the
[37:51] hell it is,
[37:52] it's almost all driver costs. So, even
[37:54] before you shrink the number of required
[37:56] cars by a factor I think the estimate
[37:58] was 5x.
[37:59] Is it 10x that savings? Is it 10x?
[38:02] So, then then but the driver's already
[38:04] the majority. So, you take the driver
[38:05] out of the loop. So, the cost of that
[38:07] ride should go down, you know, at least
[38:09] 10x cuz the car's coming down 10x and
[38:11] the
[38:12] and the driver is more than the car
[38:14] anyway.
[38:15] I think the number I've seen is 20
[38:18] between 10 and 30 cents a mile.
[38:21] Yeah, the the I've seen it as a four- to
[38:24] fivefold cheaper than owning a car.
[38:27] The the next question I want to ask and
[38:29] and offer my points of view is I think
[38:32] one of the most important one for our
[38:33] listeners. This is going to have a
[38:35] profound impact on your real estate
[38:38] holdings, where you live, what you do
[38:40] with your real estate. So, if we have
[38:42] autonomous vehicles uh and we've reduced
[38:45] the number of vehicles on the road by
[38:47] 10x, let's call it that. Um and these
[38:51] vehicles don't need to park. Again, my
[38:54] my current version [clears throat] of
[38:56] this is I get up from the breakfast
[38:58] table with my family, I walk towards the
[39:00] front door. My AI knows that I'm moving
[39:03] to open the front door. It knows where
[39:04] I'm going. It's ordered an autonomous
[39:07] vehicle, what I call automatically, for
[39:09] me. I haven't had to ask. And so, all of
[39:12] a sudden, you know, we had In our home
[39:14] here, we had a three-car garage. We
[39:17] already converted one of those garages
[39:19] into an extra bedroom. Um the other two
[39:23] garages have become effectively storage
[39:25] and I'll build out, you know, probably a
[39:27] workout gym and so forth. I think the
[39:29] idea of a garage, a personal garage in
[39:31] your home, goes away. So, start thinking
[39:33] about what you're going to do with your
[39:34] garage space. What are you going to make
[39:36] it into because you're not going to own
[39:38] a car.
[39:39] You might want access to a car, but most
[39:42] of the time do you really like driving?
[39:45] I mean, when you get into an Uber, do
[39:46] you ask the Uber driver to get out and
[39:49] let you drive? So right. And you know,
[39:51] this part of the conversation is
[39:52] incredibly actionable for all of our
[39:54] listeners. It doesn't [clears throat]
[39:55] rise to Alex's level of, you know,
[39:58] like change the world tomorrow, but it
[40:00] really matters to almost everyone who
[40:02] listens. Uh
[40:04] Everything Salim said and Peter said is
[40:06] dead right. If you're young and you're
[40:09] you've got a job and you're living in a
[40:11] city, which is 60% of you,
[40:13] um
[40:14] you might not want to buy in the city.
[40:18] Uh keep renting
[40:19] and look for something that becomes your
[40:22] second home later in life that's in a
[40:24] beautiful spot. Yes.
[40:26] >> That's a little harder to get to
[40:28] that is going to be incredibly coveted.
[40:31] Imagine a world where there's 10x more
[40:33] wealth about 2034, 2036.
[40:37] And this is a spot that anyone in their
[40:38] right mind would want. And actually, uh
[40:41] if my wife is listening, close that
[40:42] transaction that you kicked off this
[40:44] [laughter] weekend,
[40:45] even if you have to pay a little more.
[40:47] Um
[40:48] But yeah, that's what you that's the
[40:49] life plan you want because
[40:50] accessibility, not not just getting to
[40:52] it, but also delivering things to it.
[40:54] Like, you know, your Starbucks or
[40:55] Dunkin' Donuts is going to come by
[40:56] drone. Absolutely.
[40:58] >> changes what you want. Think about it.
[41:00] And there aren't, you know, we have a
[41:02] huge country like Salim said, but the
[41:05] really great spots
[41:07] are limited.
[41:08] So really do your soul searching and
[41:10] look look for that thing and don't buy
[41:13] near an airport in a city. Island real
[41:15] estate is going to become, you know,
[41:17] 10x, 100x more accessible. That will
[41:20] drive the value up. And in a downtown
[41:23] LA, you know, it's like I don't remember
[41:25] the figure. It's like like 30% of the
[41:28] blacktop is parking. All of that gets
[41:30] released to become new.
[41:33] >> 60%
[41:34] >> 60 60% of the land area is parking spots
[41:37] in Los Angeles. That's crazy number.
[41:39] >> That's insane. Well, that becomes that
[41:40] becomes gardens. It becomes green land.
[41:42] It becomes parks.
[41:44] That's incredible.
[41:45] >> and think of the unbelievable space we
[41:47] use in in stadium parking lots, right?
[41:50] Acres and acres and acres of rows and
[41:52] stuff. So we'll have to rethink, uh you
[41:54] know, tailgating and everything.
[41:56] There's so much available, uh you know,
[41:59] business opportunity here. If you can
[42:01] think ahead of what you will do with
[42:04] that. And if you're in the parking
[42:05] garage in business,
[42:07] you got to think ahead as well.
[42:09] Yeah.
[42:10] A couple of other second and third order
[42:12] implications, if I may. I mean, so we've
[42:14] already touched on I I think the more
[42:15] obvious ones, parking garages, et
[42:18] cetera, need to be reprogrammed for
[42:19] other purposes. Another, I I think,
[42:22] borderline cliche implication of full
[42:25] autonomy everywhere is the re-spread of
[42:28] suburbia. Why invest so much in urban
[42:31] center real estate if you can be
[42:33] effectively connected to an urban center
[42:35] or even not even need an urban center if
[42:38] AVs take you everywhere. Basically a
[42:40] virtual subway from from anywhere to
[42:42] anywhere. So re-suburbanization, if you
[42:45] will, at least in relatively low
[42:47] population density countries like the
[42:49] US. I think these are pretty cliche
[42:51] implications.
[42:52] A a less cliche implementa- implication,
[42:55] in my mind, is what if we just take this
[42:58] trend and extrapolate it fully to
[43:00] completion? What happens? I think
[43:02] there's a future I I put out a request
[43:04] for startups around this idea of why not
[43:07] just
[43:08] create autonomous Winnebagos, the the
[43:10] equivalent of having entire office
[43:12] buildings that are themselves autonomous
[43:15] vehicles. One could imagine living in an
[43:17] autonomous vehicle. It's all part of a
[43:19] social network. When you need to take an
[43:21] in-person meeting with someone, your two
[43:23] AVs are part of the social network and
[43:25] they connect and synchronize all of your
[43:27] locations. So maybe you're in Boston in
[43:29] the morning,
[43:30] but you're in Washington, D.C. in the
[43:32] evening. This is all handled
[43:33] automatically to synchronize your
[43:35] calendar with your AV location. And then
[43:38] it's a sleeper car and then you're in
[43:40] Chicago or wherever the next day. So
[43:42] Your bedroom car You become Your bed You
[43:45] become humans become to join you.
[43:48] Humans become internet packets that are
[43:50] being routed by the autonomous vehicle
[43:53] system. Yes, love it. Love it. Love it.
[43:55] Love it. Um you know, the EVTOLs, uh I
[43:59] it's taken a while. There is still a lot
[44:01] of doubt people have about EVTOLs. Um
[44:04] you know, the opportunity we have uh is
[44:08] going to be limited by the size of
[44:10] these, being able to land locally. So
[44:12] there needs to be sort of local hub and
[44:14] spoke uh vertiports. Um you know,
[44:17] somewhere within 5-minute driving and
[44:20] gluing these all together. And that's
[44:21] what Uber wants to do with their
[44:22] platform. So I hop in my autonomous
[44:24] Uber. It takes me without thinking to
[44:26] the right EVTOL site, which takes me to
[44:29] another location 10 km, 20 km away. And
[44:32] then uh I'm in another autonomous
[44:35] vehicle. What I'm missing from all of
[44:37] this, Alex, is is the hyperloop, right?
[44:41] So I actually joined one of the first
[44:44] hyperloop companies. Uh Virgin got
[44:46] involved. Um it was we raised, you know,
[44:50] probably close to a hundred million
[44:52] dollars. It didn't go forward, but the
[44:54] material science of creating hyperloop,
[44:56] and of course, the benefit for hyperloop
[44:57] right now is effectively supersonic
[45:00] travel uh point-to-point inner city and
[45:04] inner city, LA to San Francisco, LA to
[45:06] Las Vegas. That one will be busy. Uh so
[45:10] got to see hyperloop on this list
[45:11] eventually.
[45:13] Peter, which do you think you're going
[45:14] to see first, like in in prac- in
[45:16] practice for, say, New York to Los
[45:18] Angeles? Do you think you're going to
[45:19] see hyperloop first or do you think
[45:21] you're going to see rocket cargo first
[45:23] where you hop on an Elon Starship, go
[45:24] up, go down? You know, I I've thought
[45:28] and looked at uh point-to-point rocket
[45:30] travel and it's a tough
[45:33] it's a tough thing, the energy
[45:35] dissipation cuz you're basically going
[45:38] uh you're going to orbital velocities
[45:40] and you're having to re-enter over or
[45:42] near a city. I guess the version that
[45:44] Elon put forward was offshore landing
[45:47] facilities. That's right. So that
[45:49] you're, you know, a kilometer offshore.
[45:51] >> [snorts]
[45:51] >> Uh I think
[45:53] for one reason, rocket point-to-point
[45:56] travel, because Elon's behind it, uh and
[45:59] because the vehicle exists and they're
[46:01] going to be launching every point 5.3
[46:03] minutes. That's right.
[46:06] Uh and, you know, Elon almost got
[46:09] involved in hyperloop. But like you
[46:10] said, I can't do everything.
[46:13] Anyway, uh I'm curious what I have two
[46:15] thoughts here, quick ones. Yeah, please.
[46:16] Yeah, yeah, please.
[46:17] >> One is I think uh hyperloop will be used
[46:20] largely for commercial and for container
[46:22] loads rather than human beings cuz then
[46:24] you don't have to worry about G-forces
[46:26] and the safety standards can be lower.
[46:28] Uh and the second is remember that all
[46:30] yeah, although it takes us like 3 hours
[46:32] to fly from New York to Miami, um
[46:35] that 3 hours on a plane today is way
[46:37] more productive than it was, say, 10
[46:39] years ago. You've got full internet, you
[46:41] can work. So you can I want to see
[46:42] megabytes, baby, on Starlink.
[46:44] >> we can schedule ourselves now to do
[46:46] things when we largely want to do them.
[46:48] So I think that's a huge opportunity
[46:49] also.
[46:51] So for our listeners, I would love to
[46:53] get your feedback on this format where
[46:55] we're going deep on a topic and having a
[46:57] conversation, trying to educate you
[46:59] about how we think about about in this
[47:01] case transportation or previously
[47:03] TerraFab. Our next conversation is the
[47:06] great reshuffling. Job loss is
[47:08] inevitable. The only question left is
[47:11] what we build on the other side. So
[47:14] here's some of the stats and some of the
[47:16] articles that came out this week that
[47:18] have us thinking about this. Goldman
[47:20] says AI could automate 25% of US work
[47:24] hours.
[47:25] Seems like a low estimate to me. A PwC
[47:28] told its partners, "If you resist AI,
[47:30] you have no place here. AI tool yourself
[47:33] or get out." G42 posted a job listing
[47:36] exclusively for AI agents. Is this sort
[47:40] of a gimmick? Is it real?
[47:42] And I love this one and this come came
[47:45] from sort of a
[47:47] a hit from Jensen. Companies are now
[47:49] tracking individual employee AI token
[47:52] usage. And Jensen came out saying,
[47:55] "If a $500,000 engineer didn't consume
[47:58] at least 250,000 tokens, I'd be deeply
[48:02] alarmed." You know what this reminds me
[48:04] of? This reminds me of De Beers saying,
[48:07] "3 months salary to buy a diamond ring."
[48:09] Right?
[48:11] Doesn't it? I mean, it's like forever,
[48:13] Peter. I mean,
[48:14] >> [laughter]
[48:14] >> token talk forever, that's great. So I
[48:17] mean, literally, Jensen is saying he's
[48:19] taking the total salary, you know, of
[48:22] all engineers on the planet and
[48:25] And then Perplexity AI won in appeals in
[48:28] court uh to continue running shopping
[48:30] agents on Amazon.
[48:35] So uh I'll show one chart here and then
[48:39] let's talk about it. So AI could
[48:41] automate 25% of all work. Uh this is
[48:44] Goldman's chart uh showing
[48:47] uh
[48:47] each of these columns is a different
[48:49] type of work. And I guess the median
[48:51] here is about 25%.
[48:54] We've seen this lots of different places
[48:56] in different versions of it.
[48:59] But let's jump in. So Salim, you work
[49:01] with more consultants than I I do than
[49:04] anybody here does.
[49:06] So what do you think about this uh PwC
[49:08] telling its partners, you know, adapt or
[49:11] die?
[49:12] Yeah, I think that's fine, but I think
[49:13] it doesn't go far enough. And same with
[49:15] the McKinsey's thing. So the the
[49:17] calculations I've been running as I kind
[49:18] of
[49:19] get this paper organizational
[49:22] singularity paper finalizes that you'll
[49:24] be able to run I'm I'm just give me a
[49:26] couple of days, it'll be ready for like
[49:28] draft and review. We're just doing some
[49:29] narrative thing. Um the the
[49:32] uh you'll be able to run a typical
[49:34] company with between 20 or 25% of the
[49:37] employees you have today. Cuz all
[49:39] workflow goes from human to human to
[49:41] agent to agent, right? Now, you could
[49:43] take on the doomer side and go, "Oh my
[49:44] god, 80% job loss." But no, because
[49:46] we're just going to be creating four or
[49:47] five times more companies. And also for
[49:50] bigger companies, that transition to an
[49:53] AI-based workflow is going to take much
[49:55] longer than for a startup or mid-market.
[49:59] Uh and therefore there'll be time for
[50:00] the economy to Uh so so I'm actually
[50:03] suggesting that we want to have no
[50:05] pattern recognition in jobs. Almost
[50:07] zero. Okay? Now, um uh definitely uh
[50:10] partners who resist AI will have no
[50:12] place. There's also some need to say
[50:14] that consulting partners have no place
[50:16] in the future because in the future if
[50:18] you have an AI agent figure out your
[50:20] strategy, why do you need a consultant
[50:21] firm? You're going to need that for more
[50:23] for implementation and if they have
[50:24] better agents than you do, I think
[50:26] that's where we'll end up with that.
[50:28] Alex, uh what are your thoughts on
[50:30] these?
[50:31] You want to pick one?
[50:32] >> difficult. So, on the PwC story, it's
[50:35] very difficult for organizations to
[50:37] self-disrupt. So, if you're a management
[50:39] consultancy or an accountancy, some
[50:41] other
[50:42] uh bill-by-the-hour heavy firm, it's
[50:45] very difficult for for you to willingly
[50:48] voluntarily transition to an
[50:49] outcome-based pricing model versus an
[50:51] input-based pricing model. So, I I I
[50:54] take the you'll have no place comment. I
[50:56] I interpret that as an attempt to
[50:59] self-disrupt. In practice, it's very
[51:01] very hard to to do that. Uh and the the
[51:04] whole point of uh of Schumpeter and
[51:07] disruptive innovation in general is most
[51:10] of the most of the macro replacements
[51:14] for in in this case, for input-based
[51:17] actors in the economy are probably going
[51:19] to come from other firms, not from large
[51:21] firms self-disrupting. On the G42 story,
[51:24] I think it's actually really
[51:25] interesting. I I looked closely at the
[51:27] G42 job listing and it it really is a
[51:31] job listing for AI agents. And
[51:34] one has to wonder, yes, like around the
[51:37] edges, they also ask for details from
[51:40] the developer and what was used to make
[51:42] sure that this was really an AI agent
[51:44] submitting itself for I think it was a
[51:45] marketing job. We are so painfully
[51:49] close, I think, to a near future where
[51:51] there's a sort of reverse discrimination
[51:54] against humans and where humans need not
[51:56] apply
[51:57] >> [laughter]
[51:58] >> and up ends up being an epithet on so
[52:01] many jobs.
[52:03] Well, you have that already with the
[52:04] with the about PwC partners, right? If
[52:06] you don't use AI, get out of here.
[52:08] That's essentially we're getting to that
[52:09] point. We're halfway there already.
[52:12] Mhm. That's that's PwC, though, which is
[52:14] a human-oriented business basically
[52:16] trying to force humans to self-automate,
[52:19] at least from a a unit pricing
[52:21] perspective. Yeah. This is born AI jobs
[52:24] where humans need not apply.
[52:27] Agreed.
[52:28] You know, Saleem, one thing that you and
[52:30] I do for large companies that I think
[52:33] people need to understand, most large
[52:36] companies out there are walking dead.
[52:38] Their business models will be
[52:39] fundamentally disrupted in the next two
[52:41] to five years.
[52:43] And so the question is, how do they
[52:45] disrupt themselves before uh before
[52:47] someone else does? And the answer is,
[52:49] it's really hard, almost impossible. And
[52:52] so, you know, what you and I have done
[52:53] before is invite superb talented young
[52:57] entrepreneurs to come in, hear the
[52:59] company's business model, and say, "This
[53:01] is how I would disrupt you if I, you
[53:04] know, was funded to do it." And then the
[53:06] company should fund the best of them,
[53:09] right? And we've done this. Uh fund the
[53:12] best of them to actually build a
[53:15] adjacent company who's intended to
[53:18] disrupt the primary company. And then
[53:21] literally
[53:22] Yeah.
[53:23] >> I I the company I the design firm the
[53:25] design firm IDEO actually did this. They
[53:27] realized that their methodology would be
[53:29] widely known and they couldn't stop the
[53:32] leakage of that. So, they they picked
[53:34] one of their crazy partners and said,
[53:36] "Go to the edge and build build the
[53:37] disruptor." And he created an open IDEO
[53:40] marketplace of of design ideas. It was
[53:42] amazing. One one caveat to what Peter
[53:44] said there, the private equity guys are
[53:46] having a field day with AI automation
[53:48] and if if a company has great cash flow,
[53:51] even if its business model is doomed in
[53:53] the age of AI, the profitability is
[53:55] going to go through the roof in
[53:56] transition In the in the near term
[53:58] humans In the near term, yeah, cuz you
[54:00] know, then AI can do the job for 10%,
[54:02] soon to be 2% of the cost of the human,
[54:05] you know, with no
[54:07] no labor laws, no overhead, no you know,
[54:10] insurance, no health insurance.
[54:12] >> if you've got good cash flow, there's an
[54:14] entrepreneur looking at that salivating.
[54:17] Coming to eat your lunch.
[54:19] Yeah, so what happens is the private
[54:20] equity guys will come in and say, "Hey,
[54:22] cash flow cash cow with great cash flow,
[54:26] we're going to buy you or buy part of
[54:28] you
[54:29] and then we're going to AI-ify your
[54:31] business
[54:32] uh and that'll drive even more cash to
[54:34] the bottom line and then we're going to
[54:35] use that cash flow either as a vehicle
[54:38] to launch new things like an incubator
[54:40] or to attract that entrepreneur or to
[54:42] just roll up those startups
[54:44] and you know, acquire them back in. So,
[54:46] it becomes kind of a centerpiece. By the
[54:48] way, both Anthropic and OpenAI are
[54:51] setting are partnering with private
[54:52] equity firms to do exactly this. Go buy
[54:55] companies and then AI-enable them cuz
[54:58] you can do it with the owner.
[55:00] As if as if they did anything in the
[55:02] world, just announced a new hundred
[55:03] billion dollar fund to do nothing but
[55:05] this. So, Jeff is doing it. You know
[55:07] that.
[55:07] >> have enough money already.
[55:09] Um Alex, I'm curious or Dave, I'm
[55:11] curious about your thoughts on the
[55:12] fourth bullet here that companies now
[55:13] track individual employee AI token usage
[55:16] and you should have a minimum token
[55:18] usage per employee. Thoughts?
[55:21] Dave, do you want to go first?
[55:22] >> we we already implemented Yeah, sorry.
[55:25] We we already implemented targets across
[55:27] all of our companies on this and we're
[55:29] targeting 80% token, 20% salary. And I
[55:32] think that's a really it's very similar
[55:34] to what Saleem said a minute ago.
[55:36] Uh there's going to be huge amounts of
[55:38] job disruption in the next two or three
[55:39] years and then it'll turn around and by
[55:42] 2030, 2032, things will be good again.
[55:45] But but what you want to do is be one of
[55:49] the 20%
[55:51] that's still there when it's 80% token
[55:53] cost, 20% human cost cuz no employer in
[55:56] the world, including all the companies
[55:58] on the controlling shareholder of, care
[56:01] about cutting the last 20% of payroll.
[56:04] It's not a priority at all cuz at that
[56:06] point, an employee that can improve the
[56:08] efficiency of our AI even 1%
[56:11] is worth a lot more than cost-cutting.
[56:14] And so, we're in this footrace now to
[56:16] 80/20. Jensen's got a stepping stone
[56:19] here of of
[56:21] 2/3, 1/3 uh token cost, but that's going
[56:24] to be very transitory. We're we're
[56:25] racing toward token costs being much
[56:28] much bigger than payroll. So, I have an
[56:30] immediate step at 50/50. But it's coming
[56:33] soon. Sorry, go ahead. Alex, is this the
[56:35] right metric? I mean, cuz you can waste
[56:37] tokens. I mean, it's got to have a a
[56:39] different harness, right? You're you've
[56:40] got to be measuring something else
[56:41] besides just token usage.
[56:42] >> can waste tokens except in this
[56:45] AI-abundant era, you can also ask
[56:47] another AI to look at all the tokens a
[56:49] given employee used and ask, "Was this a
[56:51] good use or not?" Or "Was this just
[56:53] vacuous?"
[56:54] So, exactly.
[56:54] >> It becomes the ultimate self-licking ice
[56:56] cream cone. The the the quote from from
[56:58] Jensen, I think, is interesting and it's
[57:00] it's sort of if a half million dollar
[57:02] engineer didn't at least spend a quarter
[57:06] of a million dollars on on inputs that
[57:09] ultimately flow back to Nvidia, I'm
[57:11] deeply alarmed. So, so there's a little
[57:12] bit of circular [laughter]
[57:13] there's a little bit of circularity
[57:14] there that I I take with a huge grain of
[57:16] salt. No, it's the beers and the diamond
[57:19] ring.
[57:20] >> [laughter]
[57:20] >> Really is.
[57:20] >> That's right. There's another side to
[57:22] this, which is the employee side. So, I
[57:25] I I talked a bit about this in in my
[57:26] newsletter. At some firms, especially
[57:28] the frontier labs, employees are
[57:30] actually competing, so-called token
[57:32] maxing, to max out their their token
[57:34] usage on internal leaderboards to see
[57:36] who can use more tokens than than the
[57:38] other person. So, it's not just sort of
[57:41] Big Brother top-down, it's also
[57:43] bottom-up. I can use more intelligence
[57:46] more superintelligence than you can. And
[57:48] I think this is ultimately
[57:50] probably pretty healthful. Uh to your
[57:52] original question, Peter, about whether
[57:54] tokens are the right unit of
[57:55] productivity, I I think what's
[57:57] interesting is tokens they're they don't
[58:00] even have to be the right measure or
[58:02] right unit of productivity, but they're
[58:04] the first measurable unit of
[58:07] productivity.
[58:07] >> Yes. Hours are are certainly measurable.
[58:10] Like, you can punch clocks And you and
[58:13] you know,
[58:14] Yeah, it's useless. It it's naively
[58:16] measuring inputs, but tokens where you
[58:18] can actually like they're introspectable
[58:20] and they're legible and you can ask you
[58:22] can spend other tokens to look at the
[58:23] tokens the the primary tokens and
[58:26] decide, "Are these valuable tokens or
[58:27] not?" For the first time we have
[58:29] legible, defensible, analyzable inputs
[58:32] for employee productivity and that is a
[58:34] sea change. Yes. Dave, what's the advice
[58:37] here for CEOs? The advice to CEOs here
[58:39] for you. Um Alex completely nailed it.
[58:41] Like, worrying about whether the tokens
[58:43] are being used intelligently or not is
[58:46] not a problem at all in the real world.
[58:48] So, so Jensen's metric is perfect. Just
[58:52] measure the spend on tokens. And then
[58:54] Alex's insight that you must the most
[58:57] important actionable thing is make sure
[58:59] you gather all of the of the prompt
[59:01] strings history Yes.
[59:04] >> for each and every user. Cuz that's the
[59:07] AI can analyze the efficiency
[59:08] >> Analyze that. Yes. Exactly. And you can
[59:10] say, "Hey, you know, I've got a hundred
[59:13] or let's put it realistically, I've got
[59:15] eight direct reports. Evaluate the
[59:17] quality of the prompts and the output
[59:19] they have and give me feedback on which
[59:21] bottom 20% I should cut or or train up."
[59:26] Well, and really practical advice, if
[59:27] you use any of the models on Amazon
[59:29] Bedrock,
[59:31] the grabbing of the prompt history is
[59:33] already built in. It goes right into S3
[59:35] buckets. I'm sure you can do it
[59:36] elsewhere, but our company just be
[59:38] happen to be using it on Bedrock.
[59:39] Uh so it you literally don't have to
[59:41] build anything to start doing this.
[59:44] Uh you just need to grab the data and
[59:45] feed it into another AI, which you can
[59:47] also do on Bedrock or you can do on you
[59:48] know whatever.
[59:49] Uh
[59:50] personally I I like using Cloud 4.6 for
[59:52] the stuff, but uh
[59:55] you just got to close that loop. But but
[59:56] the key is
[59:57] grab the data
[60:00] right now before people get used to
[60:01] using their own home account or or you
[60:04] know something outside of your purview.
[60:05] Do not reimburse people for AI that you
[60:08] can't see. Mhm.
[60:10] Make sure it's on your infrastructure.
[60:12] Welcome to the health section of
[60:13] Moonshots brought to you by Fountain
[60:15] Life. You know, my mission is to help
[60:17] you use the latest technologies
[60:19] including AI to not just do your work at
[60:21] home, teach your kids, but to help you
[60:24] live a long and healthy life. I'm here
[60:27] today with an extraordinary physician,
[60:29] the chief medical officer of Fountain
[60:31] Life, Dr. Don Meusel. Don,
[60:33] let's talk about cancer. Uh you know, I
[60:35] know
[60:36] from the member database that we have at
[60:39] Fountain, our members who come in who
[60:41] think they're healthy, it turns out 3.3%
[60:44] of them have a cancer in their body they
[60:45] don't know about. That's right. You
[60:48] know, the majority of cancers that we
[60:49] screen for, those aren't the ones that
[60:51] are necessarily taking the lives when
[60:53] found at a late stage. We know that when
[60:55] cancer is found early, the chances for
[60:57] cure are much higher. We know it's much
[61:00] easier to treat a cancer when found
[61:01] early versus when found late. What we're
[61:04] finding in our members is over 3.3% were
[61:07] found to have these cancers that were
[61:09] otherwise wouldn't have been found or
[61:10] detected. Yeah, you know, it's
[61:12] interesting people you don't feel the
[61:14] cancer until stage three or stage four.
[61:16] And and if you don't know what's going
[61:17] on inside your body, it's like driving
[61:19] your car with your eyes closed. And you
[61:21] can know. And so when members come
[61:23] through Fountain, how do they detect
[61:25] cancers? So we're doing full body MRI,
[61:28] and we also do early cancer detection
[61:29] screening. This is very very important.
[61:32] These are not typical tools used in the
[61:34] conventional care setting when it comes
[61:36] to prevention. This is a hard thing
[61:38] because currently these are not studies
[61:40] that insurance would yet be covering,
[61:41] but the goal is to collect these
[61:43] numbers, do the research, and work hard
[61:46] to democratize wellness. Yeah. So at the
[61:49] day you can know what's going on inside
[61:51] your body. It's your obligation to know.
[61:53] So check out Fountain Life. You can go
[61:55] to fountainlife.com/peter
[61:57] to get access to the latest technology
[61:59] to help you detect cancer at the very
[62:01] beginning at stage one when it is
[62:03] curable before it gets to stage three or
[62:05] stage four and you're world of hurt. All
[62:08] right, topic number four, the collapse
[62:10] of terminal value. What happens if AI
[62:13] makes every competitive moat temporary?
[62:15] So this is a article posted by Chamath.
[62:18] Uh it's a it's a it's a powerful
[62:21] concept. He argues that AI could
[62:23] compress equity valuations
[62:25] by two to sevenfold of free cash flow uh
[62:29] down from today's S&P average of 22. So
[62:32] today the average S&P companies are
[62:34] getting 22 times
[62:37] free you know forward-looking cash
[62:39] flows. And he's saying we're going to
[62:41] get a massive reduction in that. So AI
[62:44] makes disruption so cheap and fast that
[62:47] no company can project free cash flow
[62:49] beyond five years terminal value.
[62:52] Very true. I mean it used to be all of
[62:54] the SaaS companies were projected
[62:56] forward
[62:58] and you could depend on it. Uh so this
[63:01] can break down investment paradigms,
[63:04] break down VC investing.
[63:07] Uh I'd like to jump into that, but first
[63:08] let me just show this is Chamath's
[63:12] uh sort of image he went along with his
[63:14] his post on X.
[63:15] Um
[63:16] So here we go. There's $58 trillion
[63:20] in the current S&P 500 and this is at
[63:22] the 22x of free cash flow.
[63:26] Uh if we compress it down to sevenfold,
[63:28] it drops, we lose
[63:30] a
[63:31] 2/3 of the value of the of the S&P 500.
[63:35] Uh if we end up driving it down to 2x
[63:38] free cash flow, it's down 90%
[63:41] and we get a lot of disruption of our
[63:44] financial markets.
[63:46] Um here's a chart that's showing the S&P
[63:49] 500 over the last 10 years, actually
[63:51] from 1950 through today and we're seeing
[63:55] it basically uh
[63:57] deviate significantly on on value
[64:01] you know PE ratios.
[64:03] So let's jump into some of the
[64:05] conversations. I've got the article up
[64:06] in front of me as well. I think I'll
[64:08] I'll read uh the opening paragraph here
[64:11] for us.
[64:12] Uh and Chamath said, let's start with
[64:14] first principles. The entire
[64:15] architecture of modern capital markets
[64:17] rest on a single rarely examined
[64:19] assumption that competitive advantages
[64:22] compound over time, moats persist,
[64:24] brands endure.
[64:26] Network effects defend.
[64:29] Strip that assumption away and you
[64:31] aren't just repricing some stocks, you
[64:33] would be dismantling the philosophical
[64:36] foundation of how capital has been
[64:37] allocated over a century.
[64:40] Dave, let's go to you first on this.
[64:42] Absolutely correct.
[64:44] Uh but the conclusion that the S&P is
[64:47] going to collapse is not correct.
[64:51] Uh
[64:51] if you say look, you know, prior to the
[64:54] computer revolution, my ambition was to
[64:56] build an oil company or a a
[64:59] manufacturing company that would endure
[65:00] for 50 years building the exact same
[65:02] goddamn product or delivering the exact
[65:04] same goddamn oil for 50 years so my
[65:05] great-grandchildren could be as wealthy
[65:07] as a Rockefeller. That's dead forever
[65:09] and good riddance and it should be dead
[65:11] [clears throat] forever. If you said,
[65:13] well look, 22x free cash flow implies
[65:15] that the company will exist for 22 years
[65:17] making about that same amount of money.
[65:20] Well, what company like Apple has has
[65:22] done that? You know, is Apple selling
[65:24] the same products it was 22 years ago?
[65:26] Of course not. So the overall tailwind
[65:29] is 10x
[65:31] over just the next 10 years. There's a
[65:33] massive amount of tailwind coming into
[65:34] the economy, massive amounts of new
[65:36] wealth, more than we've ever seen in our
[65:38] lifetimes is going to come into the
[65:39] economy, but you got to stop looking at
[65:42] 22 times free cash flow from the same
[65:44] product over 22 years. That's nuts. You
[65:47] have to be looking at the management
[65:49] team and the ability to to roll with the
[65:51] innovations ala Elon. And so I think the
[65:54] the overall conclusion is yeah, there's
[65:56] going to be a massive amount of
[65:57] shuffling in the S&P.
[66:00] There's going to be some huge winners
[66:01] like you've never seen before and anyone
[66:03] who's doing the same thing and resting
[66:05] on their laurels like an insurance
[66:06] company or oil company doomed. Yeah,
[66:09] he's right. He's dead right. This
[66:10] analysis is basically exactly the right
[66:12] analysis to show how that stock's going
[66:13] to go down.
[66:15] Yeah.
[66:17] Alex? Yeah, it I mean it's certainly a
[66:20] provocative thesis, but I don't think it
[66:22] holds water. I I think it's the moral
[66:24] maybe the call it the earnings multiple
[66:27] equivalent of friend of the pod Ray
[66:29] Kurzweil's notion that a singularity
[66:31] takes the form of a firewall that you
[66:33] can't see past, but except in earnings
[66:35] multiple form that that as we start to
[66:38] see faster faster more accelerationist
[66:42] innovation that free cash flow just
[66:44] comes to a halt a few years later
[66:46] because everything is disrupted, the the
[66:48] everything disruption if you will.
[66:50] Here's the problem. The the free cash
[66:52] flow does go somewhere. Capital does get
[66:55] allocated somewhere. It may not be
[66:57] allocated to SaaS startups, post-SaaS
[66:59] apocalypse. Maybe it gets allocated to
[67:01] infrastructure. Maybe it gets allocated
[67:03] to lunar mining. But capital does go
[67:07] somewhere. It's it's not actually
[67:09] capital that's being compressed, quite
[67:11] the opposite. Capital is is explosively
[67:13] expanding because now we have so much
[67:15] more infrastructure and so many more
[67:17] capabilities. So I I think the the sort
[67:19] of the the nihilistic take that earnings
[67:23] multiples are compressing because a few
[67:26] years from now there's no moat anywhere,
[67:28] I I think that's relatively narrow or it
[67:31] should be construed relatively narrowly
[67:33] to focus on the areas that are
[67:36] disrupted. In this case Chamath focuses
[67:39] on software and SaaS type businesses,
[67:41] but but but but
[67:43] everything I expect [clears throat] is
[67:45] going to be disrupted. Energy becomes
[67:47] abundant and farmland infrastructure,
[67:49] these all become abundant. So in some
[67:52] sense
[67:53] I want to zoom out and and take his
[67:55] thesis more broadly as sort of almost
[67:57] bemoaning the financial consequences of
[68:00] abundance and it it may just be the case
[68:03] that a number of our existing sectors
[68:06] that are priced based on scarcity and
[68:09] moats when moats arguably are a form of
[68:11] scarcity or at least a way of enforcing
[68:13] scarcity, those go away and we live in a
[68:15] post-moat world and that will be a
[68:18] better world. So you going to start to
[68:20] value companies in a different way. In
[68:23] the old days it was how predictable is
[68:25] their cash flow? I have number of seats
[68:27] in this particular industry and these
[68:28] are many companies I can sell it to
[68:30] number of seats available.
[68:32] And now it sounds like from what you and
[68:34] Dave just said, I'm actually going to be
[68:36] evaluating companies on their agility,
[68:38] on how rapidly they can innovate, how
[68:40] rapidly they can get the next products
[68:43] out the door.
[68:44] Uh Salim, love your thoughts here.
[68:47] Um yeah, well two points. One is you
[68:49] know, we have the CEO index where we
[68:51] ranked the Fortune 100 by their EXO
[68:53] score gauging how flexible and adaptable
[68:55] and purpose-driven are their org
[68:56] structures. And we found the top 10 of
[68:59] the Fortune 100 outperformed the bottom
[69:01] 10 by 40 times in shareholder returns
[69:04] over a seven-year period. So this has
[69:06] been going on for a long time anyway.
[69:08] Okay? I agree with Alex that capital
[69:10] still flows the less towards incumbents
[69:12] and way more towards infrastructure and
[69:13] adaptive platforms. It's a very
[69:15] important point. The only moat I think
[69:18] that's going to be left is a living
[69:19] system that learns faster than your
[69:21] competitors. That's that kind of inner
[69:23] loop that Eric Schmidt was talking
[69:24] about, right? Exactly.
[69:26] >> all the mo all the on that slider are
[69:28] under attack from AI. IP gets copied,
[69:31] switching costs, shrinking scale
[69:32] advantages are weakened, etc. etc. Um,
[69:35] I think but the bigger point I think is
[69:37] that if free cash flow visibility
[69:39] collapses beyond say 5 years, the entire
[69:42] logic of the public market has to be
[69:43] rewritten.
[69:44] And so that's a very big
[69:46] uh thing. You're you're you have to
[69:48] reward renewable renewables and
[69:50] optionality, not scale and stability.
[69:53] Physical assets are going to matter more
[69:54] than because atoms are harder to disrupt
[69:56] than bits, right? Over time. But the
[69:58] entire SaaS business model is broken. So
[70:00] this is I think one way of saying this
[70:02] is we're going to have terminal value
[70:03] collapse.
[70:05] Mhm. Yeah, that's exactly what it the
[70:07] title is actually. So the uh
[70:09] the terminal value collapses. I think if
[70:11] you if you look at the S&P at 22 times
[70:14] free cash flow,
[70:15] uh the mid-market, the non-S&P companies
[70:18] are already down to about seven times
[70:20] free cash flow, most of them. Uh so this
[70:22] is already happened outside of the S&P.
[70:26] What's propping up the S&P is mostly
[70:28] index funds. A huge fraction of the
[70:29] market is passive indexes. Mhm. And and
[70:32] people contributing blindly to 401k
[70:34] plans, which Elon Musk said clearly do
[70:36] not do that right now.
[70:37] >> [laughter]
[70:38] >> Uh but but what will happen next is a
[70:40] lot of those dirt cheap mid-caps and
[70:42] small caps will get a huge tailwind from
[70:45] AI automation. You know, especially the
[70:47] ones that have huge uh payroll and labor
[70:50] components to them. And so that's going
[70:52] to drive record you it's not un-
[70:54] unlikely that you triple your free cash
[70:56] flow while your multiple comes down.
[70:59] And so there's some serious bargains out
[71:01] there uh [clears throat] just from a
[71:03] from a straight cash flow acceleration
[71:05] through AI point of view. I mean
[71:07] shareholder shareholder calls are going
[71:09] to change to this is how we're rapidly
[71:12] iterating our products and services.
[71:15] This is how we're reinventing what we
[71:16] do. Um and our future cash flow and
[71:19] Saleem, you've got to redo the EXO
[71:21] index. It's time to take a shot at that
[71:23] once again. Uh so I'm the part of the
[71:26] paper that we're writing, the reason
[71:27] it's taking a little longer is that
[71:29] I I I hate to say but we breaks the EXO
[71:32] model, right? Community and crowd
[71:34] becomes communities and crowds of
[71:35] agents. So we have to rethink the model
[71:37] from the ground up.
[71:38] >> We're kind of mostly we've done that cuz
[71:40] then you evaluate based on those new
[71:41] criteria. For example, what's your
[71:43] intelligence that? What's your MTP
[71:45] architecture? What's your trust
[71:47] framework? There's a bunch of different
[71:48] elements that are new here that we have
[71:50] to take into account cuz the the concept
[71:53] of an organization where you did things
[71:55] inside the organization is completely
[71:56] gone. We're going to be doing API calls
[71:59] to get uh various things done, legal
[72:01] work, etc. etc. It's all going to be
[72:03] agentic. And then the firm, which used
[72:06] to be coordination costs and transaction
[72:08] costs with a bit of legal liability, now
[72:10] becomes only legal liability risk uh
[72:13] purpose container and liability
[72:15] container.
[72:17] Yeah. Well, also uh uh
[72:18] >> Amazing. And this is not super
[72:19] mainstream, so I'll get off the high
[72:21] horse quickly here but but if a private
[72:23] equity firm like Advent they must have
[72:25] heard comes in and brings either Saleem
[72:27] or Alex along and says, "Hey, we want to
[72:29] take this non-AI company with huge free
[72:31] cash flow and we want to retool it for
[72:33] the age of AI.
[72:35] Triple the cash flow and retool the
[72:37] business plan for AI." Alex or Saleem
[72:40] can tear down that company now using
[72:42] agents in 1/1000th the time that it
[72:45] would have taken a year ago or 2 years
[72:47] ago. And so whatever private equity firm
[72:49] mechanizes that is going to have no
[72:51] trouble retooling all of these entities
[72:54] cuz cuz you know exactly what the
[72:56] company's assets are, you know, whether
[72:57] it's a regulatory framework or whether
[72:59] it's a bunch of data. You can rip
[73:01] through that with Gemini or with Claude
[73:03] age or with OpenAI agents in light speed
[73:06] now. It's a transformation wave. It's
[73:08] fully automated. Yeah. It's a it's a
[73:09] transformation wave. I remember there
[73:11] was a one of the Star Trek episodes was
[73:13] uh the was it the Genesis uh machine
[73:16] that like you
[73:17] >> project. The Genesis project had like
[73:19] this wave forming agent, yeah.
[73:20] >> Yes. And this this this this uh this
[73:23] wave that that went over a planet and
[73:25] transformed everything. We're going to
[73:26] have the same thing. You're going to
[73:27] have cherry you know, you have teams
[73:29] cherry-picking companies and reinventing
[73:31] them. And disrupting. Not just not just
[73:33] private equity. I would argue this
[73:34] applies equally well to public equity.
[73:36] So something I would like I'll I'll
[73:38] broadcast a a request for startups if I
[73:40] may to the audience. Please. I would
[73:41] love to see activist investing disrupted
[73:44] by AI. I'd like AIs to write open
[73:47] letters to public firms telling the
[73:49] firms what they're doing wrong to
[73:51] disrupt them. If if you're if you're
[73:53] building an AI activist investor, write
[73:56] to me, please. Would would love to find
[73:57] a way to support you.
[73:58] >> That's a great idea, Alex. And what is
[74:00] it's a service to the CEO of the
[74:02] companies who need prompting or need
[74:05] sort of a forcing function to transform
[74:08] their business models. And the And if
[74:10] you're a board member of any of these
[74:11] companies, your job as a board member is
[74:15] to give your CEO top cover and to say,
[74:18] you know, you must get on the disruption
[74:20] uh
[74:21] uh you know, band here. You've got to
[74:23] reinvent your
[74:24] >> also a sort of a stealth way for AI to
[74:26] become a manager of the entire economy
[74:28] and not just picking off mom and pop
[74:30] businesses on the margins.
[74:32] All right. On to story number five, one
[74:35] of my favorites. Hopefully Alex, one of
[74:37] yours too. It's the new great space
[74:38] race. NASA picks SpaceX for the moon.
[74:42] Potatoes are growing in lunar dust and
[74:44] asteroids are carrying the code of life.
[74:47] All right. So here we go. I mean,
[74:49] listen, Boeing has been building uh the
[74:53] you know, the Artemis 2 vehicle. It's
[74:56] going to be launching on April 1st. Uh
[74:58] it had its uh its readiness review on on
[75:01] March 12th. And if all goes well, April
[75:04] 1st, we're going to be going back to the
[75:06] moon. Not to land but to do basically an
[75:09] Apollo 8 style circumlunar orbit. Uh
[75:12] very cool but you know, the new NASA
[75:14] administrator, uh an amazing individual.
[75:17] Uh I'm very happy and proud to call him
[75:19] a friend. We're going to have him on the
[75:20] pod. He's agreed to do it. Just needs to
[75:22] get scheduled.
[75:23] Um is elevating SpaceX into the Artemis
[75:26] program. So Starship is going to be
[75:29] taking
[75:30] uh I think astronauts more safely, more
[75:33] economically. We'll see those numbers in
[75:34] a moment. But just to be clear, this is
[75:37] not the US story only. Uh China has
[75:40] confirmed their intention to land on the
[75:42] moon by 2030. Let's play it back again.
[75:45] History repeating itself. 1961, we're on
[75:48] the moon before end of the decade.
[75:50] China's saying they're on the moon
[75:52] before the end of the decade. Um and so
[75:55] that's going to be a a beautiful
[75:57] competition. Uh we'll get to the idea
[75:59] that you know, you can grow potatoes
[76:02] >> [laughter]
[76:02] >> in lunar soil. They're going back to the
[76:05] Martian, you know, uh uh again an
[76:07] incredible movie now that uh Project
[76:09] Hail Mary is out. I can't wait to go see
[76:11] it in IMAX theater later this week.
[76:13] And we just saw that on the asteroid uh
[76:17] Ryugu Ryugu, um
[76:19] we found the five nucleo nucleo bases
[76:23] nucleic bases for DNA, which has four
[76:26] ATCG, and RNA, which has uh U for
[76:29] uracil.
[76:30] And we found all five of those bases on
[76:33] that asteroid. This strengthens the
[76:35] panspermia theory that life on Earth
[76:39] originated elsewhere in our galaxy in
[76:41] the universe and it rained on Earth and
[76:44] gave us the starting components uh for
[76:47] that. Let's take a look at uh this
[76:49] chart. I put on the left here uh NASA's
[76:52] SLS. And now to be clear, when I say
[76:55] NASA's SLS, NASA's the prime contractor
[76:59] and it has probably uh aerospace
[77:02] companies in every single congressional
[77:04] district building that vehicle. It is an
[77:07] expendable vehicle in in a time when
[77:10] everybody's going reusable. There you
[77:13] have SpaceX um with Starship. And just
[77:16] for comparison of size, here's the
[77:18] Saturn uh Saturn 5 that got us to the
[77:20] moon. If you look at these two charts,
[77:23] these two bar charts on the left here,
[77:26] uh we're seeing uh well, let me just go
[77:28] to the center ones. We're seeing the uh
[77:31] the delivered mass uh to orbit. That
[77:35] that uh is it
[77:38] is Starship over, you know, twice as
[77:41] much as we're getting with with Artemis
[77:44] with um the SLS vehicle. And if you look
[77:47] at uh to translunar injection TLI,
[77:50] getting out of Earth orbit to the moon,
[77:52] we're seeing twice as much mass going on
[77:54] a Starship compared to um uh to the SLS.
[77:59] But where the rubber really hits the
[78:00] road is launch costs and mission costs.
[78:03] Um it's expensive to be running the SLS
[78:06] system. It's like the space shuttle. The
[78:08] space shuttle used to cost if you did
[78:11] four launches a year, it was a billion
[78:12] dollars a launch. If you did one launch
[78:15] a year, it was four billion dollars a
[78:16] launch. It wasn't the cost of the
[78:18] vehicle. It was the standing army of
[78:20] 20,000 humans that were used to to
[78:23] operate the space shuttle. So I I
[78:26] honestly don't know why the SLS has
[78:29] existed as long as it has. I think
[78:31] Starship's going to do a clean sweep of
[78:33] this. And of course, we've got Blue
[78:34] Origin as well.
[78:36] Alex, comments.
[78:38] Well, uh do you want to place bets as to
[78:40] how long before the United Launch
[78:42] Alliance, which is the prime contractor
[78:44] for SLS, gets acquired by Jeff Bezos or
[78:47] someone else?
[78:48] But why would you acquire it? I guess
[78:50] for the contracts? For the contracts,
[78:53] for the expertise. Uh I'm familiar with
[78:56] again all the cliches in the space space
[78:58] about how SLS was a a make jobs program
[79:02] or a way to keep alive in civilian form
[79:05] certain capabilities that were useful
[79:07] for defense or other say intelligence
[79:10] purposes. But I think at the end of the
[79:12] day, we're so painfully close to finally
[79:15] relaunching a second space race. And I
[79:18] think Starship is is the obvious
[79:20] incumbent there, not the SLS. So, I
[79:23] hopefully we have humans on landing on
[79:26] the moon again in the next two to three
[79:28] years and we get humans eventually on
[79:31] Mars and all of this plays out exactly
[79:34] as For All Mankind has foreseen except
[79:37] decades late.
[79:38] Yeah.
[79:39] I don't know if you if Dave or Salim you
[79:41] want to play on this on this
[79:42] conversation. I just think, you know,
[79:45] we're we're building the
[79:47] uh I don't know what your best
[79:49] historical analogy, the you know, the
[79:51] covered wagons, the railroads Wagon
[79:54] train to the stars, Gene Roddenberry
[79:55] called it. It's it's all currently on on
[80:00] Starship.
[80:01] Uh Starship is the only economical.
[80:04] Yeah. Go ahead.
[80:05] >> I've I've thought two or three things.
[80:06] One is we've gone from kind of
[80:08] government space theater to commercial
[80:10] space evolution. I think that's really
[80:12] powerful. For me, the really exciting
[80:14] thing was finding all the nuclear bases
[80:16] on reusable. Um you know, it takes life
[80:18] from scarcity to abundance.
[80:21] Um Yeah, let's let's go there. So,
[80:22] here's
[80:24] Here's the Here's the graphic if you
[80:26] would. Again, adenine, guanine,
[80:29] cytosine, thymine and uracil, the five
[80:32] components of DNA and RNA
[80:35] found on these on these vehicles. I do
[80:37] believe that as we get to Mars, as we
[80:40] get to Europa, as we get to the all of
[80:42] the planets and moons, that we're going
[80:44] to find at least microbial life uh
[80:47] you know, ubiquitous on all of these.
[80:50] Did you see Peter Garrett's prediction
[80:52] about microbial life on Mars?
[80:54] No, what did he say? Garrett's our NASA
[80:56] administrator, yes. Yeah,
[80:58] uh
[80:59] NASA administrator
[81:01] said that he he predicted more than 90%
[81:04] plus probability that NASA will
[81:07] eminently find evidence of microbial
[81:10] life in some form on Mars, which is a
[81:12] sea change in in terms of NASA's
[81:14] official position on life on Mars. It
[81:16] was always, well, we found water, frozen
[81:19] water. Now we found liquid water. We
[81:20] hope to find signs of life. Signs are
[81:22] ambiguous. Now for the first time we
[81:24] have a NASA administrator who's saying
[81:26] 90% probability we're going to find
[81:28] microbial life.
[81:30] And the exciting thing is how related is
[81:33] it going to be to microbial life on
[81:34] Earth. One of the theories, of course,
[81:36] Mars cooled first, which probably means
[81:40] life evolved on Mars first. And of
[81:42] course that we know when aster a large
[81:45] asteroids impacted Mars, the ejecta, the
[81:48] rocks that flew out, some of them
[81:50] reached Mars escape velocity and landed
[81:52] on Earth. And so we have Martian
[81:55] meteorites
[81:57] in museums today.
[81:59] Uh and did those did those meteorites
[82:02] carry life with them from Mars to the
[82:04] Earth? Are we going to find
[82:06] basically even genes that are common
[82:09] between Martian life and life here? I
[82:11] mean, the real exciting thing is if we
[82:13] go to Europa or someplace like that and
[82:15] we find completely independent life
[82:17] forms that don't connect with life on
[82:19] Earth, that will be amazing.
[82:22] That's really cool.
[82:23] So, I got a question for you since since
[82:25] you guys are experts on this and I'm
[82:26] not. Um
[82:28] is it in the scenario where lo and
[82:31] behold it turns out
[82:33] that everything we learned in biology
[82:34] should have said life started on Mars or
[82:37] actually started farther out in the
[82:38] solar system
[82:40] and then asteroids knock chunks off and
[82:42] then they transport to other chunks
[82:44] and then life starts over again
[82:46] and then it ends up on Earth through
[82:48] that mechanism. Is that all going to be
[82:50] then bounded to the solar system or is
[82:53] it more likely in your mind that hey,
[82:55] wait, this this propagates through deep
[82:57] space?
[82:58] >> when I was a freshman in school, I did a
[83:01] paper on the interstellar medium and you
[83:03] can actually look at the interstellar
[83:05] medium and you can find the building
[83:07] blocks of life out in the in you know,
[83:11] in the medium between stars in our in
[83:13] our in our galaxy.
[83:15] These components are everywhere.
[83:19] And and the galaxy is relatively
[83:21] constant on the time scale of a billion
[83:23] years. So, I think the statistic is Mars
[83:25] cooled about a billion years or so plus
[83:27] or minus earlier than Earth. That's a a
[83:30] lot of So, so the galaxy first of all is
[83:32] not rigid. We're we're constantly We
[83:34] have different stars at different
[83:36] velocities passing by each other, close
[83:38] passes, all of that. So, on a time scale
[83:41] of a billion years, that buys an
[83:42] enormous amount of time for panspermia
[83:45] at potentially a galactic scale, not
[83:47] even necessarily at a an interstellar
[83:50] neighborhood scale. And we're we're
[83:53] we're several generations in as well.
[83:55] We're, you know, born from from several
[83:58] generations of of stars
[84:00] exploding and then forming new stars.
[84:03] There's been a lot of nebular mixing in
[84:05] our interstellar neighborhood. There
[84:07] there's one other of relevance story um
[84:10] Please.
[84:11] >> folks can find it if if if they Google
[84:12] it.
[84:13] This is from a few years back attempting
[84:16] to extrapolate based on genetic
[84:17] complexity when the the last common
[84:20] ancestor actually would have been
[84:23] and finding that if you just take
[84:25] genetic complexity as I don't forget how
[84:28] exactly it's measured, but you come up
[84:29] with some appropriate parameterization
[84:31] of genetic complexity of life on Earth,
[84:34] extrapolate backwards, you find that the
[84:37] time when you get the first base pair
[84:39] happens approximately a billion years
[84:42] before life is thought to have observed
[84:44] on Earth. So, that's sort of an
[84:46] independent measure of when in principle
[84:48] life as we know it, DNA RNA based life
[84:51] could have emerged. Maybe it started on
[84:53] Mars, but we'll find out I suspect soon
[84:55] enough. Exciting times. Salim, you want
[84:57] to add something?
[85:00] No, I just remember my favorite thing
[85:02] around all this is the Drake equation.
[85:04] Where you where I'm I'm just going
[85:06] through the
[85:07] where you
[85:09] where you calculate all the
[85:11] factors that led to the probabilities of
[85:14] binary stars and life appearing. And you
[85:16] And the when you add it all up, you end
[85:18] up with 100%. It's
[85:20] So, the panspermia thing, but I think
[85:22] what we
[85:23] mentioned earlier, if we could find
[85:24] something that's non-carbon based, that
[85:26] would be truly exciting.
[85:28] You know, what's really cool to me is
[85:29] this this idea that the the dinosaurs
[85:32] >> [laughter]
[85:33] >> The The dinosaurs were extinguished by a
[85:35] meteorite
[85:37] and or meteor. And
[85:39] uh
[85:40] the propagation of the DNA or the base
[85:43] pairs is also via asteroids and meteors.
[85:47] And early in the universe history, you
[85:51] know, this may be popping up all the
[85:52] There might be life popping up
[85:53] constantly everywhere and propagating
[85:56] through all these projectiles flying
[85:58] around. But it always gets extinguished
[86:01] by another meteor, you know, just like
[86:02] the dinosaurs were.
[86:04] And it's not until everything cools and
[86:06] settles that you can have enough time to
[86:08] evolve human
[86:10] intelligence or other intelligences out
[86:12] there in the universe. So, it's just a
[86:14] big system dynamics settling problem,
[86:16] which is just
[86:18] just really cool to me to think about.
[86:20] I hope it turns out to be right. Yeah.
[86:23] All right. Uh
[86:24] story number six, the model wars go
[86:26] underground. The AI frontier is
[86:28] fracturing into a stealth arms race
[86:31] where anonymity is the new moat. And
[86:34] here's the story. There two stories here
[86:36] to focus on. One, OpenAI launches
[86:38] GPT-5.4 mini and nano
[86:41] that runs twice as fast and approaches
[86:44] the full GPT-5.4 on coding benchmarks.
[86:47] So, these models are getting smaller and
[86:49] faster.
[86:51] And the second story, which I think is
[86:52] most of our conversation here, is there
[86:55] was a mystery model. So, a one trillion
[86:58] parameter model called Hunter Alpha
[87:01] appeared on open router with no
[87:03] attribution. It was secret. It a million
[87:06] token context window. It was free. There
[87:09] was no developer announced, no press
[87:10] release, no origin story and it
[87:12] processed 160 billion plus tokens.
[87:16] Everyone thought it was DeepSeek V4.
[87:19] Right? Cuz DeepSeek had been the main
[87:21] player here, but it turned out to be
[87:23] Xiaomi's AI team and when that was
[87:25] announced, their stock went up 5.8%.
[87:29] I remember meeting the team at Xiaomi
[87:31] when they came out with their first
[87:32] mobile [music] phone, like three young
[87:34] founders. They've since gone beyond just
[87:37] mobile phones to electric cars and now
[87:39] they've got a killer model.
[87:42] Points, gentlemen.
[87:45] I I think there are at least
[87:46] >> Point one is proliferation of models is
[87:49] very hard to contain because the the
[87:51] existence of the prior model gives you a
[87:53] complete roadmap on how to build the
[87:54] next model and it helps you build the
[87:56] next model. Mhm. So, I at this stage, I
[87:58] think it's a fair bet that trillion
[88:00] parameter models are going to propagate
[88:02] all over the world with anyone who has
[88:04] about 50 to 100 million dollars that
[88:06] they're willing to invest.
[88:08] Uh and that'll come down, too, as Alex
[88:10] is pointing out. Many times uh the
[88:13] the algorithmic improvements are driving
[88:14] that down constantly.
[88:16] Alex
[88:16] >> Alex. Sorry, I cut you off.
[88:17] >> Yeah, maybe a couple points. So, first
[88:19] on the the 5.4 story, distillation
[88:23] continues to work and I find that
[88:25] completely remarkable. On the one hand
[88:28] >> you explain distillation for our
[88:30] listeners who don't know? Yeah. Sure.
[88:32] So, the
[88:33] I think the reasonable expectation for
[88:35] say
[88:37] OpenAI as well as other firms launching
[88:39] a big model first with lots of weights,
[88:42] a high parameter count, and then
[88:44] subsequently launching a mini or nano
[88:46] version, and by the way, Anthropic does
[88:48] the same thing and DeepMind does the
[88:50] same thing. They all launch smaller
[88:52] models later is that they're using the
[88:54] larger models to generate lots of data,
[88:57] synthetic data, and then using those
[88:59] synthetic data to train a a smaller
[89:02] model that can be faster and less
[89:04] expensive. So, that that's sort of a a
[89:07] caricatured way of describing the
[89:08] distillation process of in some sense
[89:10] squeezing down or compressing the larger
[89:13] model down to a a simpler student model.
[89:16] And the fact of this continues to work
[89:19] is I think borderline magic. The the
[89:22] amount of complexity that's already in
[89:25] the full 5.4 model that and and moreover
[89:29] that 5.4 has likely been the result of
[89:31] so-called iterated amplification and
[89:34] distillation over many cycles where 5.4
[89:37] was likely in large part trained off of
[89:39] synthetic data generated by distilled
[89:42] models from earlier generations that we
[89:44] can keep playing this magic trick over
[89:46] and over again. It's borderline magic
[89:49] that that it continues to work and that
[89:51] we continue to be able to distill down
[89:53] models while retaining a large fraction
[89:55] of their capabilities. It it again
[89:58] makes me think that there has to be an
[90:00] end to the story, but hopefully it's a
[90:02] very satisfying end where at the end of
[90:05] the the distillation rainbow we get like
[90:07] the the distilled black hole of a model
[90:10] or a neutron star or something, the
[90:12] ultimate phase change where it's maybe
[90:14] like a few million parameters. It's the
[90:16] the end state of the game.
[90:17] >> A one kil- a one kilobyte file on your
[90:19] on your phone is all you need.
[90:21] >> file that's like the master equation for
[90:23] superintelligence after all of this
[90:25] distillation. It it we we showed on a
[90:27] previous podcast uh a gentleman on his
[90:29] iPhone using a distilled model on
[90:32] airplane mode uh being able to basically
[90:35] answer every question. So, imagine if on
[90:38] all of your devices without having to
[90:40] have, you know, Wi-Fi internet access,
[90:43] you have the distilled knowledge of
[90:44] humanity there to serve you. It's inside
[90:47] your kids teddy bear. It's in your, you
[90:49] know, Thomas train set. Um
[90:51] it's becomes magical. Here's my question
[90:53] for you uh
[90:55] Alex and and and Dave. You know,
[90:58] now that we're seeing this,
[91:00] uh we're seeing a basically uh a mystery
[91:05] trillion parameter model announced
[91:07] without any attribution.
[91:09] Uh
[91:10] it used to be that the traditional moat
[91:13] [clears throat] for these models was
[91:15] their brand, their capitalization,
[91:18] who they were, you know, is there any
[91:21] defendability or are we just going to
[91:22] see newcomers rushing in with new models
[91:26] um that you're going to just utilizing
[91:29] new model, you're going to be no longer
[91:31] dependent upon AI uh or Gemini. Thoughts
[91:35] on that? Well, you you've said it a
[91:36] million times, Peter. Data is actually
[91:39] the great moat, not the model itself.
[91:42] And many, many people are accumulating
[91:44] phenomenal data for, you know, for sur-
[91:46] brain surgery, for material science, for
[91:48] chemistry, for all of these use cases.
[91:51] And, you know, if you create the next
[91:54] great, great, great model using that
[91:55] proprietary data, the parameters are out
[91:58] there in the world, but the the data
[91:59] that trained it is not, and it's very
[92:01] hard.
[92:02] You people can use the model, but they
[92:03] can't compete with you by creating a a
[92:05] rip-off model cuz they don't have the
[92:06] underlying data. Now, you can generate
[92:08] synthetic data using the prior model.
[92:10] Uh Alex is dead right about that, but I
[92:12] don't think that it's
[92:14] I don't think it's all these companies
[92:15] killing each other. I think it's the
[92:16] whole all the boats rising with the
[92:18] tide. I also think that if you take what
[92:20] Alex said a minute ago,
[92:22] and you know, so many college seniors
[92:24] ask me, "What should I do? What should I
[92:25] do? How many How do I you know, what
[92:27] Just replay 10 times what Alex just said
[92:30] slowly
[92:31] until you fully understand everything he
[92:33] just said. And then ask your favorite AI
[92:36] to generalize on it and find as many
[92:38] documents as you can around the internet
[92:39] to read.
[92:40] At the end of that process, you'll be
[92:43] able to build a distilled
[92:46] focused model that solves some problem
[92:48] better than anyone else on the planet.
[92:50] And that's that's instant business,
[92:51] instant value add, instant success. So,
[92:53] just just really In fact, the other
[92:55] thing you can do is take your Open Claw
[92:58] and have it look for every episode of
[92:59] this podcast where Alex said something
[93:01] related to what he just said. And have
[93:03] it also synthesize that and bring it
[93:05] back and feed it into your machine. I
[93:07] guarantee that's a good move. There's
[93:09] something good good spring project for
[93:11] anyone listening. If Sam Altman were in
[93:13] this discussion, he might point in terms
[93:15] of the moat question that you were
[93:16] asking, Peter, to well, Open AI is
[93:20] building up its own data centers.
[93:21] Although, that's no longer really true.
[93:23] Stargate is being pivoted to now renting
[93:26] servers. So, maybe less of a moat there.
[93:29] He might point to having the best
[93:31] research team in the world generating
[93:34] the best models.
[93:35] But, they've been hemorrhaging
[93:37] researchers, and those are becoming a
[93:39] commodity. Then he might point to being
[93:41] the becoming the core subscription
[93:43] having as as he said, a billion plus
[93:46] users. I'd much rather have a more than
[93:48] a billion users than I would have
[93:51] state-of-the-art model because models
[93:52] walk out the door every day. There's a
[93:54] lot of fungibility in terms of research
[93:56] employees.
[93:58] Only one problem. That the billion user
[94:00] distribution advantage may be a a little
[94:03] bit tenuous at the edges because you see
[94:05] maybe enterprises are more valuable as
[94:08] customers than individuals. So, maybe
[94:10] the billion users a little bit less
[94:12] valuable on margin. And then maybe also
[94:14] you see other labs that are able to use
[94:16] cheap Chinese open-weight models maybe
[94:19] fine-tuned legally or otherwise with
[94:22] clawed outputs are able to put out
[94:24] seemingly miraculous results. So, I I I
[94:26] do think we're seeing the baseline
[94:28] models for the moment become something
[94:30] of a commodity, and the value then
[94:32] migrates up the stack to Open Claw or
[94:34] other higher-level frameworks.
[94:36] This episode is brought to you by
[94:38] Blitzcy, autonomous software development
[94:40] with infinite code context. [music]
[94:43] Blitzcy uses thousands of specialized AI
[94:46] agents that think for hours to
[94:48] understand enterprise-scale code bases
[94:51] with millions of lines of code.
[94:53] Engineers start every development sprint
[94:56] with the Blitzcy platform bringing in
[94:58] their development requirements.
[94:59] >> [music]
[95:00] >> The Blitzcy platform provides a plan,
[95:02] then generates and precompiles code for
[95:04] each task. Blitzcy delivers 80% or more
[95:08] of the development work autonomously
[95:10] while providing a guide for the final
[95:12] 20% of human development work required
[95:15] to complete the sprint. Enterprises are
[95:17] achieving a 5x engineering velocity
[95:20] increase when incorporating Blitzcy as
[95:22] their pre-IDE development tool pairing
[95:25] it with their coding copilot of choice
[95:27] to bring an AI native SDLC into their
[95:30] org. Ready to 5x your engineering
[95:32] velocity? Visit blitzcy.com to schedule
[95:35] a demo and start building with Blitzcy
[95:37] today. [music]
[95:41] If we could, I'm going to move on to
[95:42] number seven, machines that build
[95:44] machines. AI designed a CPU in half a
[95:47] day, and now it wants to put data
[95:49] centers in orbit. So, here's the article
[95:53] um that uh you know, prompted me to have
[95:55] this conversation about, you know,
[95:57] machines building machines. We're seeing
[95:59] recursive self-improvement
[96:01] happening at a faster and faster and a
[96:04] more fundamental rate than ever before.
[96:06] So, an AI agent called Design Conductor
[96:09] uh by Vector autonomously built a 1.5
[96:12] GHz Linux-capable RISC uh RISC-V CPU
[96:18] uh from concept to tapeout in 12 hours
[96:21] compressing a quarterly engineering
[96:22] cycle into a lunch break.
[96:25] Um
[96:27] pretty extraordinary. And so, uh you
[96:30] know, here's the here's the actual
[96:32] numbers. Uh Vector AI did this
[96:35] particular design task in 12 hours. You
[96:38] can see in the chart here while the
[96:39] traditional engineering team would have
[96:41] normally taken 90 days. Now, maybe this
[96:45] is a little bit overplayed. Uh I'm sure
[96:48] it's not just 90 days. I'm sure that
[96:50] they were saving time along the way.
[96:52] But, what we're seeing over and over
[96:54] again is AI being able to do, you know,
[96:57] .0 to completion on its own
[97:01] uh iterating faster than any humans.
[97:04] Um Alex,
[97:06] this is recursive self-improvement
[97:07] starting to break out of the software
[97:10] loop. Uh this is the innermost at least
[97:12] portion of maybe a rivulet uh of the the
[97:15] innermost loop where it's You see this
[97:17] recursive self-improvement which would
[97:19] otherwise be software optimizing
[97:20] software starting to eat through its
[97:22] container. It's eating down to the the
[97:25] EDA uh electronic design automation
[97:27] level of designing RISC-V cores. And
[97:30] then it I think it's going to eat
[97:32] further out and redesign the data
[97:34] centers and the energy supplies and then
[97:37] the entire economy. So, it is in one
[97:40] sense very satisfying to to see this
[97:42] happening. In another sense, maybe a
[97:45] person who's slightly more skeptical
[97:48] that that this is that this represents a
[97:50] broader trend would say, "Well, of
[97:51] course it was able to to automatically
[97:54] design a RISC-V core. RISC-V has all of
[97:56] these unit tests, so it's it's easy to
[97:58] define verifiable rewards, and then you
[98:00] can do RL and all of these other things.
[98:03] You can
[98:04] iterate uh and put react loops on it
[98:07] because you you have an easy way of
[98:08] knowing whether a given architecture uh
[98:11] a given floor plan for the chip works or
[98:13] not because it's such a common
[98:14] architecture." But, I I think that
[98:16] cynical perspective completely overlooks
[98:18] how remarkable it is that we're now at
[98:20] the stage of recursive self-improvement
[98:22] where the thing is designing its own
[98:24] chips.
[98:25] And not only here, it's going to be
[98:26] robots building robots. It's going to be
[98:29] everything. Um so, here's the here's a
[98:31] couple of questions, you know, so if a
[98:32] senior design engineer earns $400,000 uh
[98:36] and a full tapeout team costs, you know,
[98:39] millions over course of a year, if AI
[98:41] collapses it to a lunch break, what
[98:44] happens to the 50,000 hardware engineers
[98:46] currently working?
[98:48] Uh where do they get applied?
[98:50] Oh, I don't think that I think that
[98:51] vision is flawed in a huge way in that
[98:54] right now because the cost of
[98:56] engineering a new chip is so high, we
[98:59] all use the same GPU and CPU for every
[99:01] single task even though it's nowhere
[99:03] near optimal for that task. What this
[99:05] unlocks is chip designs that are
[99:07] specific to the use case that are
[99:09] probably about a factor of 10 more
[99:11] efficient and maybe more. And if you
[99:14] think, well, we're going to spend two,
[99:16] three trillion dollars
[99:18] on these chips over the next couple of
[99:20] years on these data centers, if you can
[99:22] unlock a 10x performance improvement for
[99:24] a use case,
[99:25] that has hundreds of billions of dollars
[99:27] of implications. So, all 50,000 of those
[99:29] engineers are going to be useful using
[99:31] the AI for all the different use cases,
[99:32] for all the different chip designs.
[99:34] Also, the fab doesn't care a whit. Like,
[99:36] you can change the mask every day for a
[99:40] different design and still get the same
[99:42] throughput through the factory. So, it's
[99:44] it the the fabs don't care a whit if
[99:46] there are tens of thousands of different
[99:47] designs instead of us all using the
[99:48] exact same CPU for everything. So, this
[99:51] is just a huge unlock. I I love the way
[99:53] we're using trillions and trillions now
[99:55] on a regular basis. We're just, you
[99:58] know, a couple years ago it was billions
[99:59] and billions.
[100:00] It's more fun and you can feel it.
[100:02] >> you can feel the acceleration. We need a
[100:04] new TV series that's called Trillions,
[100:06] Not Billions. Uh for sure.
[100:09] Um I I want to hit a few other stories
[100:12] quickly before we get to our AMA
[100:13] segment, and these are stories that
[100:16] didn't fit in the other categories. And
[100:17] again, please give us your feedback on
[100:19] the format here today. Are you enjoying
[100:20] it more? Um we'd love to know. So, in
[100:24] other news, here we go. Um
[100:28] Uh the DOE announced about 300 million
[100:31] dollars for the Genesis project, uh
[100:33] inviting teams to leverage AI across 20
[100:35] national challenges spanning
[100:37] manufacturing, biotech, and energy.
[100:39] Of course, the Genesis project is about
[100:42] the US actually, uh you know,
[100:45] using its national labs and the data
[100:48] contained within national labs uh to
[100:50] accelerate and expand uh
[100:53] uh the US in its AI and scientific
[100:56] pursuits.
[100:57] Alex.
[100:59] I think it's a generally a good thing
[101:01] for the US to have an industrial policy,
[101:03] and I think Genesis mission to the
[101:05] extent that for the first time, at least
[101:07] from the Department of Energy's vantage
[101:09] point, it is starting to articulate
[101:12] grand challenges that are in
[101:14] collectively part of a broader
[101:15] industrial policy, which the US hasn't
[101:17] had for decades. I think it's very
[101:20] important. So, fusion, obviously, one uh
[101:24] one of the grand challenges. I I think
[101:26] it's so important for to the extent we
[101:29] have a federal government that has a
[101:31] budget to to fund progress, to put money
[101:34] behind grand challenges in general. So,
[101:37] I'm I'm I'm I'm in the weeds from a
[101:40] bunch of different dimensions with the
[101:41] Genesis mission. Actually, uh Dario,
[101:44] who's the uh
[101:46] leading uh the relevant portions of DOE
[101:49] on this, I worked with him as an
[101:50] undergrad at MIT and as an undergraduate
[101:53] researcher. So, so some fond memories.
[101:56] >> [laughter]
[101:57] >> It's a what a tangled web we weave. What
[101:59] can I say? But, I'm I'm I'm generally a
[102:01] big fan of of what Genesis is doing.
[102:02] >> what concerns me here, Alex, is that um
[102:05] this is great, right? These are like X
[102:07] prizes in one sense that the
[102:09] government's going to be running. But,
[102:11] and it's moving us in the direction that
[102:14] China has been doing for a while now.
[102:16] Yes.
[102:17] >> is deploying hundreds of billions of
[102:20] dollars into state-directed AI
[102:22] investments and saying, you know, we
[102:25] want fully development in uh in the
[102:27] architectures around robotics, around
[102:29] these AI models, and so forth. This is a
[102:31] relatively [clears throat] small amount
[102:33] of money for the US government.
[102:34] Hopefully, it's just a first toe in.
[102:37] It's true, but on the other hand,
[102:39] I I would argue China distorts its
[102:42] markets so much relative to to what
[102:46] if you compare US industrial policy
[102:49] distortion versus Chinese industrial
[102:51] policy distortion, they're they're not
[102:53] in the same league, and we have much
[102:55] deeper private capital markets that
[102:57] China lacks. I I like our odds on
[103:00] balance much more than China's.
[103:02] Mhm.
[103:03] >> [clears throat]
[103:03] >> Um our next article here is
[103:06] the rural Ohio Ohioans
[103:10] seek a constitutional amendment uh to
[103:13] ban data centers over 25 megawatts in
[103:15] the state. And, you know, this is the
[103:18] ultimate NIMBY, not my backyard. And
[103:20] it's pretty extreme. I mean, to go after
[103:24] an you know, a constitutional amendment.
[103:26] This is a genuine grassroots revolt at
[103:29] the end of the day.
[103:30] Um We we need I think, Peter, we need a
[103:32] new acronym. Like, what what was it?
[103:35] >> N- not not not in my backyard. Like, yes
[103:38] in my orbital plane. Yeah. Okay. Or
[103:41] something. [laughter] This is just going
[103:42] to drive all these data centers to
[103:44] orbit. But, this is this is crazy. I
[103:46] mean, these communities don't realize
[103:48] the amount of wealth these data centers
[103:50] are going to create for them. I think
[103:52] it's about 10 billion dollars per
[103:54] gigawatt of invested power.
[103:57] Or in in invested.
[103:59] They'll miss them when they see them in
[104:00] the night sky.
[104:02] >> [laughter]
[104:04] >> Yeah, obviously utterly insane and
[104:07] utterly insane to use a constitutional
[104:08] amendment for this purpose.
[104:11] I mean, to point to point out the
[104:11] obvious, the data centers are are tiny
[104:14] as a footprint on land. They're
[104:15] absolutely tiny. Yeah. And the wealth
[104:17] that they create is astronomical for the
[104:19] neighborhood they're in. So, there's got
[104:20] to be a much better way to make a
[104:22] win-win than to ban something that's
[104:23] obviously going to benefit your state
[104:25] tremendously. But, let put that aside.
[104:28] You're in California.
[104:30] Alex and I are in Massachusetts.
[104:32] The way we make decisions through
[104:34] legislatures is so messed up.
[104:38] Yeah.
[104:38] >> Like, that something like this could
[104:39] even get proposed is ludicrous. And
[104:42] that's what really needs to change cuz
[104:43] when you talk to the governors, they're
[104:45] like, I don't want this.
[104:46] Like, okay. Well, you we're
[104:48] representative democracy. There's
[104:49] supposed to be very, very smart people
[104:51] thinking about complex issues and then
[104:52] deciding what happens. You don't throw
[104:54] things like this out to a referendum of
[104:57] people who just got laid off.
[104:59] And it's and it's people saying, you
[105:00] know, my access to clean water and
[105:03] energy and my you know, my consumer
[105:06] price index of energy is going through
[105:07] the roof, and there's other ways to deal
[105:09] with this. Um then instead of banning it
[105:12] uh by constitutional amendment. That's
[105:15] for me, that's insane. All right.
[105:17] >> I've never seen a data center that
[105:19] affected the water supply. That's like
[105:20] it's so I hear it all the time. It's
[105:22] utterly ludicrous. The data center needs
[105:24] a fixed amount of water to cool itself.
[105:26] It doesn't drink the water. It just goes
[105:27] around in a circle. It's
[105:29] It's nuts.
[105:31] Uh
[105:32] All right. Uh another story worth
[105:34] mentioning is Nvidia
[105:36] won approval to sell uh its H200 chips,
[105:40] its most advanced chips, uh to Beijing.
[105:43] That's a big deal.
[105:44] Um I'm surprised.
[105:47] >> the Well, the realization is the ban
[105:49] didn't work.
[105:50] Uh China both was getting access to chip
[105:53] through third parties, and China was
[105:55] developing its own competitive. And this
[105:57] cost Nvidia tens of billions of dollars.
[106:00] And since it's not working, in fact,
[106:02] it's stimulating a homegrown homegrown,
[106:04] you know, uh
[106:05] equivalent of Nvidia in China, they said
[106:08] uh
[106:09] let's reverse policy. The question is,
[106:11] is it too late? Yeah. That everything
[106:13] that you just said is correct except the
[106:15] one part where when Jensen complains he
[106:17] lost tens of billions of dollars, every
[106:19] single thing he's manufactured is sold
[106:21] out for years to come. Uh
[106:24] So, that's true.
[106:24] >> that it didn't go to China, it
[106:26] definitely sold. Even if it was like a
[106:28] you know, a a dysfunctional 8080 design
[106:31] or whatever that did the you know,
[106:32] Chinese design, it still got sold.
[106:35] Everyone everyone in the world wants
[106:36] these things. So, that he didn't
[106:38] actually lose any money. I'm surprised
[106:40] though because I think
[106:41] I think the embargo or the ban didn't
[106:43] work. You're dead right. Uh China is
[106:45] doing its own thing now. But, I also
[106:47] think that if you say, well, let's start
[106:49] selling them again, maybe they'll stop.
[106:51] No, that's not going to
[106:53] you cut them off. They're not going to
[106:55] forget. Like, that that isn't going to
[106:57] happen. So, I was really surprised that
[107:00] you know, that they reversed course on
[107:01] this. I don't know if I want to say
[107:03] anything further.
[107:04] >> Yeah, if if if I'm Beijing, I mean, on
[107:06] the one hand, I I read the same stories,
[107:08] and of course, variety of Chinese
[107:10] frontier AI labs are all slurping up as
[107:13] many H200s as they can get. And of
[107:15] course, it's borderline obvious that the
[107:18] Blackwills are are now the frontier. So,
[107:20] in some sense, Beijing is being kept a
[107:22] half step or two behind the frontier
[107:24] chips available to US AI labs.
[107:27] I think the story behind the story, not
[107:29] not to be overly speculative, but if I
[107:32] were the Chinese Communist Party, I'd be
[107:34] doing the moral equivalent of having
[107:36] people taste my water at this point in
[107:38] terms of these chips. Nvidia's been very
[107:41] public about how there are a variety of
[107:44] countermeasures that can be put in place
[107:46] to prevent the wrong chips from ending
[107:48] up in the wrong locations.
[107:50] I would
[107:52] and this is based on stories that I've
[107:53] read, stories where the Chinese
[107:55] government is suspicious at the at the
[107:58] circuit level of uh American chips. Uh
[108:01] I I have to imagine that they're looking
[108:03] now at our chips with renewed scrutiny
[108:06] to see what else is in these chips that
[108:08] we're shipping to them.
[108:09] >> What what what algorithms are embedded
[108:10] deep inside. We've seen this uh in the
[108:12] opposite direction. All right. Here's a
[108:14] story, Alex, that you and I have enjoyed
[108:16] talking about. Scientists successfully
[108:18] froze an entire pig brain
[108:20] while locking in the cellular activity
[108:22] with minimal damage. Uh this is
[108:24] cryogenics, and it's happening in a
[108:27] large mammal. Of course, the pig has
[108:30] organs, uh heart, liver, lung, kidney,
[108:32] and brain on the order of human organs.
[108:36] So, this is this is significant. Um
[108:40] So, actually, I I played a minor role in
[108:43] the story, and I'm not subject to
[108:44] confidentiality on the story, so I can
[108:46] tell the story. This is from This is
[108:48] from a company I informally advise named
[108:51] Nectome. Uh and uh I have a another
[108:54] company with uh that where the the
[108:57] founder of Nectome is also involved.
[108:59] This is Eon, focusing on whole brain
[109:02] uploading and emulation. Nectome, which
[109:04] I'm not formally involved with, is
[109:06] focused on just the preservation side.
[109:09] And I I'd been nudging them like they
[109:12] have these amazing results, publish the
[109:14] results. They published the results and
[109:16] it it is so wonderful to see now for the
[109:19] first time real competition in call it
[109:22] the cryonics space or the preservation
[109:24] space since the way Nectome works isn't
[109:26] quite the same way as say the way 21st
[109:29] century medicine, which we've spoken
[109:31] about previously on on the pod works.
[109:32] 21CM is more focused on vitrification.
[109:36] Nectome is more focused on a type of
[109:39] chemical preservation, but nonetheless
[109:42] >> the cryopreservant is? I mean this is
[109:44] about and just to describe it to the to
[109:46] our listenership, you're basically at or
[109:49] near death, you're replacing the blood
[109:50] supply with something that goes and
[109:53] fundamentally
[109:55] you know, infiltrates the cells and
[109:57] keeps the water in the cells from
[109:58] crystallizing and destroying the
[110:00] structure in the cells.
[110:04] That's right. So you're you're you're
[110:06] searching your latest model to find the
[110:08] answer.
[110:08] >> I'm I'm I'm double checking to see how
[110:10] much they've made public. Uh-huh. So so
[110:13] maybe let me just talk about it at a
[110:16] high level. So
[110:18] so it's it's a chemical technique. It's
[110:20] it's a little bit less focused on
[110:22] vitrification with the whole point of
[110:24] vitrification on the 21CM side is
[110:28] is basically in ensuring that ice
[110:30] crystals don't form and that there isn't
[110:32] strong osmotic pressure reverse osmotic
[110:35] pressure that that cause cells to
[110:37] explode. On the Nectome side, it's a
[110:39] chemical process. I'm I'll be cautious
[110:42] with what I say because I haven't
[110:44] I need to check to see what what's in
[110:45] the public and and what isn't about the
[110:47] process.
[110:49] But the more important I think results
[110:52] here in addition to Nectome putting out
[110:55] I think their their first bio archive
[110:57] paper in years since the original paper
[110:59] that won the brain preservation
[111:01] Foundation award for demonstrating local
[111:05] preservation of the structure of of
[111:08] neurons is that now they've demonstrated
[111:10] in in full public view scaling this
[111:13] process up to an entire
[111:15] mammalian brain, a large mammalian
[111:17] brain, not even just a mouse brain. So I
[111:20] I think we're we're finding ourselves in
[111:22] in a near future slash present where
[111:25] finally we have enough data to to be
[111:27] confident that entire mammalian brains
[111:31] are being preserved and this immediately
[111:33] raises the question, which is the
[111:34] question I ask almost everyone, why
[111:37] where are all of the cryonics patients?
[111:40] Why don't we have billions of people now
[111:43] that we have a growing body of evidence
[111:46] that brain structure can be preserved by
[111:48] whichever technique, whether it's
[111:49] Nectome on the one side, 21CM on the
[111:52] other, why don't we have a billion
[111:53] people signing up for cryonics? And I
[111:56] would again to the audience, sign up for
[111:58] cryonics. Like just do it.
[112:00] >> way
[112:01] this is the way you get to see the 23rd
[112:03] century. Yes. It's a I got
[112:06] I heard from from the head of Alcor
[112:09] after my last call to action to to do
[112:11] cryonics. Apparently lots of people
[112:13] flooded into Alcor. It's a nonprofit. I
[112:15] make no money off of saying this, no
[112:17] financial interest. Just get yourself a
[112:20] cryonics plan as part of a portfolio for
[112:23] longevity. Love it. Love it. I want to
[112:26] take a second and just say thank you to
[112:28] Nick Singh and
[112:30] and Dana Khan our producers for
[112:32] supporting us on this new format. I
[112:34] enjoyed it. Did you guys enjoy it?
[112:37] I love it. It just feels organized.
[112:40] Yeah, well it feels organized and fun.
[112:42] It actually feels fun to think through
[112:43] the topics with you guys. It's like an
[112:45] entire episode worth of AMA. Yeah, with
[112:48] ourselves. Yeah, it actually changes the
[112:50] stories we cover too, you know, cuz we
[112:52] we normally go through the most
[112:53] important stories to change your life.
[112:56] But here when you put it into themes,
[112:58] you actually dig up other stories that
[112:59] are related to the primary topic that
[113:01] you otherwise would have missed. So I
[113:02] love that. All right,
[113:04] here we go. Let's pick one each from
[113:07] page one and one from page two.
[113:10] Alex, would you go first?
[113:13] Oh, so many good options.
[113:15] Yes. We'll start with number two. Where
[113:19] should entrepreneurs actually run their
[113:21] AI compute? Local hardware, AWS cloud,
[113:24] or iPhone? And that comes from Frank
[113:27] Gerard Marketing.
[113:29] There isn't a good answer. There are at
[113:32] least no single good answer, lots of
[113:34] decent answers. There are benefits to
[113:36] each. So with local hardware, you have
[113:39] greater control, greater confidentiality
[113:42] and data privacy. You're going to on the
[113:44] other hand end up maintaining said local
[113:47] hardware. You have to worry about your
[113:48] own backups. Can be a pain in the neck
[113:50] from variety of different perspectives.
[113:53] With AWS or one of the many other public
[113:56] clouds, you don't have to worry about
[113:57] that. That's abstracted away. On the
[113:59] other hand, you might have to compete
[114:02] ferociously for access to say GPU
[114:05] resources. You're competing with other
[114:08] tenants for common resources. You might
[114:11] have to worry on margin depending on how
[114:13] familiar you or your organization are
[114:15] with with opsec and and cybersecurity.
[114:18] You might have greater surface for
[114:20] attack. On the other hand, you have more
[114:23] more scalability.
[114:25] With the iPhone, you have it it's sort
[114:27] of the ultimate edge device until all of
[114:30] us and not just some of us are running
[114:32] foundation models on our watches and our
[114:34] smart glasses, which is already
[114:36] happening and is going to be more evenly
[114:37] distributed. You have even greater
[114:40] privacy. So I I don't think this is
[114:43] maybe this goes without saying. I I
[114:45] don't think this should be viewed as a
[114:47] black or white or binary trade-off.
[114:49] There is a spectrum from edge
[114:52] edge compute to data centers at the
[114:55] core. I I think the best answer actually
[114:58] is I want to run my AI compute in the
[115:00] Dyson swarm. And that Dyson swarm will
[115:03] be perfect blend when fully realized in
[115:06] a few years
[115:07] of of data centers. We'll have lots of
[115:10] maybe if if Elon's statistics are to be
[115:12] believed, 100 kilowatt nodes filling the
[115:15] sky.
[115:16] But also it'll be incredibly elastic. If
[115:19] if we're disassembling the moon to fire
[115:20] off new 100 kilowatt nodes in in the
[115:24] the stellar or Dyson swarm fabric,
[115:26] [clears throat] it perfect blend.
[115:28] Okay Dave, which one you going to
[115:31] choose?
[115:31] >> one thing on this topic only because I
[115:33] spent the whole weekend
[115:35] dorking around on Amazon AWS Bedrock,
[115:38] which is a great choice by the way, even
[115:41] though
[115:42] if my bed was made of rock, it would
[115:44] feel like getting started on Amazon
[115:45] Bedrock. I mean that's that's it's
[115:47] it's a brutal get up and running process
[115:50] on Bedrock. But it does the critical
[115:53] thing that you need, which is it
[115:55] captures all of your prompt history for
[115:57] you and any teammates that you have into
[116:00] easy to manage S3 buckets. So your AI
[116:02] can analyze everything that you've done,
[116:05] which is a critically important
[116:06] function. So that may be available
[116:07] elsewhere, too,
[116:09] but it's it's probably as good a choice
[116:12] as any. But whatever you do, don't just
[116:13] start running on some random hardware
[116:15] and then lose all the prompt history.
[116:17] So this is just an easy way to capture
[116:18] it. Pick a number.
[116:20] I'll take number three. Are humanoid
[116:22] robots over-engineered? Would it be more
[116:24] efficient to isolate basic needs like
[116:26] food, water, and clothes and automate
[116:28] those directly instead from Go Unite
[116:31] FB3GN.
[116:33] Short answer, yes, absolutely. So why
[116:37] are we putting all this energy into
[116:38] humanoid robots?
[116:40] The reason we're putting all the energy
[116:41] into humanoid robots is because AI kind
[116:43] of came into the world almost overnight
[116:46] and we're in a race to capital right
[116:48] now. And so what's critical for all
[116:50] these
[116:51] projects and startups is getting funded.
[116:54] The humanoids are so much more visually
[116:56] appealing that they're easier to fund.
[116:58] They're also easier to recruit into. And
[117:01] that'll unlock the supply chain of all
[117:02] the parts and that'll unlock all of the
[117:04] other robots that you know, that farm
[117:06] and create clothes and whatever, which
[117:08] will probably not look all that
[117:10] humanoid. But you know, when you look at
[117:12] the gigafactory like Peter and I did,
[117:14] vast majority of the automation there is
[117:16] not humanoid robots. It's machines that
[117:18] look like you know, machines doing their
[117:20] jobs. Yep.
[117:21] >> And then the humanoids just operate
[117:22] those machines. So I think they are
[117:25] over-engineered and over-invested
[117:27] relative to where we'll end up, but for
[117:28] a very good reason you should think
[117:29] about like visual appeal and capital
[117:32] raising are a core core part of this
[117:34] step function we're living in right now.
[117:35] So I'm going to go with number four. How
[117:37] can AI be used to end a war, not as a
[117:40] weapon, but as an impartial negotiator
[117:43] that all parties could trust? This is
[117:45] from at
[117:46] JN kind five. So I find that absolutely
[117:51] fascinating and I do think it's a
[117:53] powerful tool. If you haven't used a
[117:56] large language model for negotiations,
[117:59] one of the things is we don't know how
[118:01] to think other than the way that we we
[118:03] know how to think. So being able to put
[118:05] yourself in the mindset of another
[118:07] individual is extraordinarily powerful.
[118:11] If you haven't said listen, I'm you
[118:13] know, I'm anti-guns, my my neighbor, my
[118:16] friend, my spouse loves guns, you know,
[118:19] could you please help me
[118:21] explain to them my feelings in a way
[118:23] that lands with them and isn't viewed as
[118:27] offensive,
[118:28] you can get some you know, extreme
[118:32] you know, support on your negotiation
[118:34] skills. And at the end of the day, I
[118:37] think this is one of the most exciting
[118:38] unexplored applications for AI because
[118:41] the system can ingest also every peace
[118:44] treaty, every negotiation script, every
[118:46] conflict resolution framework that's
[118:49] ever been had and can you know, model
[118:51] outcomes with no tribal allegiance. One
[118:54] of the biggest challenges we have as
[118:56] humans is we have these cognitive biases
[118:59] and these tribal biases that are driving
[119:02] us. So, you know, can we use this for
[119:05] negotiation? Absolutely. Um
[119:08] and I think both sides, if you set as an
[119:11] objective function that you want to
[119:14] reach a balanced solution that both
[119:16] sides have and both sides are using AI,
[119:20] I mean, it could be different models, I
[119:21] think the probability of getting to a
[119:23] solution is much, much higher. Um
[119:27] we are biased when we're dealing with
[119:29] humans. Uh and one of the things that
[119:31] goes on when you're talking, for
[119:33] example, to an AI model for, uh you
[119:36] know, psychological therapy,
[119:38] um
[119:39] when you realize you're, you know,
[119:41] you're telling your innermost thoughts
[119:43] to a human, you feel like you're going
[119:45] to be judged, but you don't feel judged
[119:47] when you're talking to an AI model. And
[119:49] so, I think there's real value to be had
[119:51] here.
[119:52] Um I don't know if Salim's back online
[119:54] or not, but
[119:55] >> He's not, but if if if if if I might add
[119:57] just one thing to to your point, Peter.
[119:59] Something that I'm seeing more and more
[120:01] in the past few months, not for for war,
[120:03] but for commercial negotiation, I'm
[120:05] seeing this all the time. Two parties
[120:07] that are at loggerheads in a commercial
[120:10] negotiation, one of the parties will
[120:12] bring in a frontier model and ask the
[120:13] frontier model what the commercially
[120:16] reasonable outcome is, bring it to the
[120:18] other side, the other side will will
[120:20] consult their model, and they'll come to
[120:22] rapid agreement. Yeah.
[120:24] >> I'm seeing this happen now over and over
[120:26] again.
[120:27] All right, here's the next eight
[120:28] questions from our AMA. Uh Alex, over to
[120:31] you.
[120:32] Sure. Uh again, there are so many fun
[120:35] questions here. Uh
[120:37] rather than choose eight, which would
[120:38] require me to give implicit investment
[120:40] advice, number seven,
[120:43] I'll avoid that one. Number six, uh
[120:46] less slightly less interesting. I'll
[120:48] tackle number five since since I've been
[120:50] beating the drum a bit for for solve
[120:52] everything including disease.
[120:54] So, the question is, once AI solves most
[120:57] diseases, how soon will treatments be
[120:58] available to everyone? Will access lag?
[121:01] And this is from Katisse 896.
[121:05] So, maybe the sub question first is,
[121:08] when do I think
[121:09] AI has a decent chance of solving most
[121:12] diseases? My timeline, and this is not
[121:14] specific to me, I I think if you ask the
[121:16] the more optimistic elements of CZI, the
[121:19] Chan Zuckerberg Initiative, Biohub,
[121:22] maybe Arc Institute, and some other
[121:24] organizations, maybe Anthropic on a good
[121:27] day, I think they'll say something like
[121:30] five years from now.
[121:31] So, five years is
[121:33] pretty rapid time scale. It's more rapid
[121:36] in in many cases than what historically
[121:39] has been the the clinical trial process
[121:41] end to end through three phases.
[121:44] So, second sub question here is, how
[121:47] soon will treatments be available to
[121:49] everyone? If if say tomorrow one of the
[121:51] frontier labs says, all right, here here
[121:54] are the cures
[121:55] to the top 5,000 unsolved or untreatable
[121:58] diseases, we have vast computational
[122:02] experiments
[122:04] demonstrating to to the satisfaction of
[122:06] all experts that these are the cures or
[122:09] at least the treatments for these
[122:11] diseases, how soon would those
[122:12] treatments be broadly available?
[122:14] Under the present regime, which by the
[122:16] way is not the same as the regime even
[122:18] one year ago, there would still be
[122:20] probably a multi-year process. The FDA
[122:23] has recently announced two major
[122:25] developments that are, I think, relevant
[122:27] here. One, the FDA under this
[122:30] administration has decided to adopt a
[122:32] Bayesian perspective as opposed to a
[122:35] frequentist perspective, meaning that
[122:37] they're willing to incorporate for the
[122:39] first time in history evidence in terms
[122:41] of clinical approvals from outside a
[122:43] particular drug. That's that's a huge
[122:46] sea change. It means that in in
[122:48] principle, drug approvals can operate
[122:51] much more quickly because they can take
[122:52] into account lots of pre-existing
[122:53] information that predated the particular
[122:56] drug. Second big development, a move
[122:58] from, and this was again relatively
[123:00] recently announced by this FDA, from a
[123:02] two clinical trial process to a one
[123:04] clinical trial process for certain
[123:07] cases, expediting the the approval
[123:09] process.
[123:10] So, I I think, fast forwarding to as
[123:13] long-winded answer to how soon will
[123:15] treatments be available to everyone, I
[123:16] think if tomorrow or five years from now
[123:19] a frontier lab or multiple frontier labs
[123:21] said, here are the very well-motivated
[123:23] top 5,000 cures to everything, I think
[123:25] we would see similar developments from
[123:28] the FDA to go to a zero clinical trial
[123:31] model given enough Bayesian evidence and
[123:34] given enough computational evidence,
[123:36] which is to say a zero trial model. I I
[123:39] think there would be so much political
[123:40] pressure that we would probably, barring
[123:43] some exceptional circumstance, see
[123:45] relatively fast availability.
[123:47] >> All right, so Dave, let's go to you.
[123:48] Okay. Uh well, I'll take number eight
[123:51] since Alex couldn't touch it and we've
[123:53] lost Salim. Um if you had to choose one
[123:55] public company to bet on in the age of
[123:57] AI, which one and why? From Matthew
[123:59] Johnson 6525.
[124:01] Uh so, we can't give investment advice,
[124:03] obviously, but I will tell you, I've
[124:06] said a bunch of times on the pod, go to
[124:08] 13f.info.
[124:11] Look up the Situational Awareness Fund,
[124:13] which is Leopold Aschenbrenner,
[124:15] and every quarter he has to file his his
[124:18] holdings. He's killing it, and the
[124:20] reason he's killing it is cuz he listens
[124:21] to exactly what Alex is always saying,
[124:23] look for the innermost loop.
[124:25] Find you the tailwinds in equities and
[124:28] assets are like nothing you've ever
[124:29] seen.
[124:30] But you have to be in the AI loop to be
[124:33] relevant. And so, you'll see all
[124:35] Leopold's holdings are somehow in the
[124:37] centerpiece of the innermost loop. And
[124:39] so, those are the things you want to
[124:40] own. So, those include things that are
[124:42] chip fabs, things that have power,
[124:44] things that are related to chip design,
[124:47] things that are algorithmic that are
[124:49] directly deriving use cases. Those are
[124:51] all in that fund. So, that's your road
[124:53] map. So, look at his holdings and then
[124:55] generalize from there, and you'll find
[124:57] lots of great stuff that you should be
[124:58] you should be buying. Also, you know,
[125:00] W-2 income is going to get pummeled in
[125:02] this next three years, but assets, you
[125:05] know, holdings and ownership and things
[125:06] is going to go through the roof. So, so
[125:08] buy stuff, you know, whether it's
[125:10] equities, public public or private, real
[125:13] estate, you know, things that'll
[125:14] appreciate. That's what you need uh in
[125:17] this next three-year window.
[125:19] Amazing. I am under time pressure
[125:21] myself. I've got to jump on a film
[125:24] recording. I'm going to go to our outro
[125:27] music here,
[125:28] uh which is uh brought to us by CJ
[125:31] Trueheart. Um it was a piece he
[125:34] developed for the Abundance Summit
[125:36] called Moonshot Minds. Uh take a listen.
[125:39] Love the lyrics. Uh here we are. Thank
[125:42] you to CJ for Moonshot Minds.
[125:48] >> [music]
[125:57] [music]
[127:47] >> All right, gentlemen.
[127:48] >> gets better and better all the time.
[127:50] >> all the time. AWG, DB2, uh this was fun.
[127:54] Wishing you guys an extraordinary day.
[127:57] Uh Salim, who's airborn to Brazil, safe
[128:01] travels, buddy. Uh soon we'll be putting
[128:04] you on rocket rides to get you there.
[128:06] Anyway, uh thank you everybody for
[128:08] listening. Uh please tell
[128:10] >> hyperloop. Yeah, yeah, I guess we can do
[128:12] hyperloop under the under the Gulf.
[128:15] Anyway, long story short, thank you for
[128:17] listening. Uh please join us. If you've
[128:19] not subscribed, please do. We're putting
[128:21] out these podcasts on
[128:23] cadence that you want to get alerts, so
[128:26] uh our mission, help you see an abundant
[128:29] world. The world is getting better at an
[128:31] extraordinary rate. The technologies to
[128:33] solve the world's biggest problems. If
[128:35] you're an entrepreneur, thank you for
[128:36] being an entrepreneur. Entrepreneurs are
[128:38] individuals who find juicy problems and
[128:40] solve problems. The more entrepreneurs
[128:41] on the planet, the better Earth and
[128:44] humanity is.
[128:46] Gentlemen,
[128:47] till I see you next time. Be well. Peter
[128:49] Diamandis, your host, signing off.
[128:52] See you guys.
[128:53] Take care.
[128:53] >> Have a good movie, Peter.
[128:55] Thanks, pal. If you made it to the end
[128:56] of this episode, which you obviously
[128:58] did, I consider you a moonshot mate.
[129:00] Every week my moonshot mates and I spend
[129:02] a lot of energy and time to really
[129:04] deliver you the news that [music]
[129:06] matters. If you're a subscriber, thank
[129:07] you. If you're not a subscriber yet,
[129:09] please consider subscribing so you get
[129:11] the news as it comes out. I also want to
[129:13] invite you to join me on my weekly
[129:16] newsletter called Metatrends. I have a
[129:18] research team. You may not know this,
[129:19] but we spend the entire week looking at
[129:22] the Metatrends that are impacting your
[129:24] family, your company, your industry,
[129:25] [music]
[129:26] your nation. And I put this into a
[129:28] two-minute read every week. If you'd
[129:30] like to get access to the Metatrends
[129:32] newsletter every week, go to
[129:33] dmandis.com/metatrends.
[129:35] That's dmandis.com/metatrends. [music]
[129:39] Thank you again for joining us today.
[129:41] It's a blast [music] for us to put this
[129:42] together every week.