GPT 5.2 Release, Corporate Collapse in 2026, and 1.1M Job Loss | EP #215 (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) OpenAI releases GPT 5.2. The (00:00:03) capabilities are just shockingly (00:00:05) different than they were a few weeks (00:00:07) prior. (00:00:07) >> OpenAI has just unveiled GBT 5.2, which (00:00:11) it's billing as its most advanced (00:00:12) frontier model yet. (00:00:14) >> The value that we see people getting (00:00:16) from this technology and thus their (00:00:17) willingness to pay makes us confident (00:00:19) that we will be able to significantly (00:00:21) ramp revenue. (00:00:21) >> The fastest scaling consumer platform in (00:00:24) history. We're almost at a billion (00:00:26) users. That just blows my mind. (00:00:28) >> A lot of change is coming rapidly. I (00:00:30) think the the biggest challenge is (00:00:31) people are not projecting properly on (00:00:34) how rapidly this is going to tip. (00:00:36) >> I think 2026 is going to see the biggest (00:00:39) collapse of the corporate world in the (00:00:42) history of business. (00:00:43) >> In 2025, we had 1.1 million layoffs, (00:00:45) which is the most since the 2020 (00:00:47) pandemic. 71% of comparisons between a (00:00:51) human performing this knowledge work and (00:00:54) the machine resulted in the machine (00:00:56) doing a better job at more than 11 times (00:00:58) the speed of the human and at less than (00:01:01) 1% of the cost of the human (00:01:03) professional. So knowledge work is (00:01:06) cooked. (00:01:08) >> Now that's a moonshot ladies and (00:01:10) gentlemen. (00:01:14) Uh, speaking of alien creatures, I was (00:01:17) touring with uh, Colossal yesterday. Ben (00:01:20) Lamb, I'm an adviser, early investor in (00:01:24) this company, and Colossal is amazing. (00:01:27) Uh, they've got something like 12 (00:01:28) different species at different stages of (00:01:32) deextinction, right? They brought back (00:01:34) the direwolf. (00:01:35) >> Uh, they're going to bring be bringing (00:01:37) back the saber-tooth tiger. I can't wait (00:01:39) for that. And, of course, the the woolly (00:01:41) mammoth. uh they created the woolly (00:01:43) mouse, right? So they've been able to (00:01:45) identify the genes that that in (00:01:48) particular are different are different (00:01:50) phenotypes, right? Like length of hair, (00:01:52) length of snout. (00:01:54) >> And it's fascinating what they're doing. (00:01:58) Uh and their ability to actually find (00:02:02) the closest living relative and then (00:02:05) snippets of DNA. So they have DNA going (00:02:07) back as far as 1.2 million years. They (00:02:11) haven't been able to get DNA older than (00:02:13) that, but that's still pretty (00:02:14) incredible. (00:02:15) >> But being able to actually like (00:02:17) >> Yeah. (00:02:17) >> Didn't Ben say that we couldn't uh (00:02:20) restore animals if the DNA was older (00:02:22) than like 10,000 years? (00:02:24) >> Well, for example, the woolly mammoth (00:02:27) DNA that they've gotten uh ranges from (00:02:30) like 10,000 years to 1.2 million years, (00:02:32) right? And (00:02:33) >> Okay. (00:02:33) >> And they've got to identify (00:02:36) that's not a single species, that's a (00:02:38) whole spectrum of a species, (00:02:39) >> right? because there's evolution going (00:02:41) on all that time. And so they're trying (00:02:42) to figure out, okay, what part of the (00:02:45) phenotypes like the tusk and the woolly (00:02:48) mammoth hair and its cold tolerance and (00:02:51) all of those things and they're (00:02:54) reconstructing a single single room, you (00:02:57) know, an approximation of woolly (00:02:59) mammoth. (00:02:59) >> Anyway, the programs are amazing and and (00:03:02) Ben is such an incredibly good CEO. I'm (00:03:04) excited. He's going to be uh one of our (00:03:06) moonshot closing speakers at the (00:03:07) Abundance Summit this year. So we're (00:03:09) going to go deep with how do you how do (00:03:10) you go from 0 to10 billion valuation in (00:03:13) four years and how do you do something (00:03:15) and no bio background at all for Ben (00:03:18) right he was the CEO of Hyper Giant the (00:03:20) software company incredible (00:03:22) >> so your multi-armed robot can shear the (00:03:25) woolly mouse and then we can make make (00:03:27) sweaters in time for the holidays out of (00:03:29) it very (00:03:30) >> we can all wear them on the pod (00:03:31) >> by nonhumanoid robots (00:03:34) >> all right I think uh I think it's time (00:03:38) to to jump in with enthusiasm. (00:03:41) >> Yes. (00:03:43) >> All right. Welcome to Moonshots, another (00:03:45) episode of WTF Just Happened in Tech. Uh (00:03:48) this is the news that hopefully impacts (00:03:50) you, inspires you, gives you moonshot (00:03:54) thoughts, and gets you ready for the (00:03:55) future because that is one of our (00:03:57) primary goals. How do we prepare you for (00:03:59) what's coming next? Uh a lot of AI news. (00:04:03) Uh today is a special episode that we (00:04:06) pulled together uh in order to celebrate (00:04:09) the release of GPT 5.2, but we'll get to (00:04:11) that in just a moment. I wanted to hit (00:04:13) on some of the top sort of like top (00:04:16) level hyperscaler updates and battles. (00:04:20) So, uh just a few headlines here. We'll (00:04:22) be discussing them through the pod here (00:04:24) today. Chat GPT was the most downloaded (00:04:27) app in the iOS app store in 2025. (00:04:30) Congratulations to them. Uh they're (00:04:33) nearing 900 million active users. Gemini (00:04:36) is catching up. Uh Anthropic jumps to (00:04:40) 40% enterprise share. Uh uh amazing. (00:04:44) Accenture is going to be training 30,000 (00:04:46) people on Claude. Elon has let us know (00:04:49) that Grock 4.2 is coming very shortly in (00:04:53) the next few weeks and Grock 5 in the (00:04:56) next few months. (00:04:58) Uh we said in a moment ago, Open AI has (00:05:02) released GPT 5.2. That's going to be uh (00:05:05) coming up in a moment. And interestingly (00:05:07) enough, uh Google launched its deepest (00:05:10) AI research agent the same day that Open (00:05:13) AI dropped GPT 5.2. Uh a little bit of (00:05:17) PR battles going on between them all. (00:05:20) >> Uh all right. Uh one other piece of data (00:05:25) uh on the downloads here to give people (00:05:27) uh a look at the scoreboard. Uh GPT chat (00:05:30) GPT received 92 million downloads. (00:05:34) Gemini is at 103.7 (00:05:37) million downloads and Claude has (00:05:39) received 50 million downloads. Any (00:05:42) comments on on these opening headlines (00:05:44) before we jump into uh GPT 5.2? Well, (00:05:48) I'm I'm in shock this week at the (00:05:51) capabilities. We'll look at the (00:05:52) benchmarks in the in a minute, but the (00:05:54) benchmarks really underell the last two (00:05:57) weeks. The the capabilities are just (00:06:00) shockingly different than they were a (00:06:02) few weeks prior. Uh and we'll get into (00:06:04) it, but uh also the big big change is (00:06:07) the race is on. you know, uh, when, um, (00:06:10) you know, GPT5 kind of disappointed (00:06:12) everybody, the poly market on on Google (00:06:15) running away with the rest of this year (00:06:16) went to like 90 95%. (00:06:19) Uh, now, kind of as Alex predicted, uh, (00:06:23) it's a closer horse race. You know, you (00:06:25) know, Google's still on top of the (00:06:26) stack, but apparently Sam had something (00:06:28) in the tank and who knew. So, we'll (00:06:31) we'll get into that, too. But I'm just (00:06:33) absolutely like, no, no exaggeration. (00:06:37) The things that I got done in the last (00:06:39) week that I couldn't have done three (00:06:40) weeks prior just coding and building (00:06:43) things are it's just I'm I'm in shock. (00:06:46) >> So, um, are they pulling their punches? (00:06:48) We discussed that in the past, right, (00:06:50) where they're releasing this much. They (00:06:52) know that, you know, uh, we're going to (00:06:54) have Grock coming out next. So, let's (00:06:58) then release the next segment to compete (00:07:00) directly there. (00:07:02) they are totally pulling their punches. (00:07:03) They've absolutely been holding back. (00:07:06) >> Uh I think because they're starved of (00:07:08) compute and they're afraid to roll out, (00:07:11) you know, addictive capabilities that (00:07:12) they just can't deliver on. But you (00:07:13) know, Alex experienced this too. Like (00:07:15) yesterday we were, you know, going crazy (00:07:16) with 5.2 trying to see what it can do (00:07:19) and then it's like, "Sorry, you're done (00:07:20) for today. (00:07:22) We're out. We're out of compute. Sold (00:07:23) out. No gas in the tank." And so the (00:07:26) competitive pressure is forcing them to (00:07:27) code red, you know, come out with things (00:07:30) when they normally would want to hold (00:07:31) back and wait until they can find the (00:07:33) data center compute and wait until Chase (00:07:35) Lock Miller finishes Abalene and but (00:07:37) they just don't have that choice with (00:07:38) the competitive pressure on each other. (00:07:40) Yeah, m maybe just to to comment, I I (00:07:43) think there at this point if if you're (00:07:44) open AI and you have your purported code (00:07:47) read and you're in a hurry, you're in a (00:07:49) bind 5.1 GPT 5.1 came out only a month (00:07:53) ago and you need to to rush something to (00:07:56) market to to uh put at ease perceived (00:08:00) competitive pressures. I I think they're (00:08:02) only approximately three levers you (00:08:04) have. So one lever to Dave's point is (00:08:07) compute. you can increase the total (00:08:10) amount of compute allocated to to given (00:08:13) models and that that of course comes at (00:08:15) a cost. It comes at the cost of compute (00:08:17) scarcities. It comes at the cost of (00:08:19) longer response times to prompts. Second (00:08:22) lever that you have is safety. So you (00:08:25) can turn down the safety. You can make (00:08:27) models more sophantic. Uh and that (00:08:30) that's that's a way to improve, (00:08:33) >> right? But can we get a benchmark on (00:08:36) sophantic model? There there are a bunch (00:08:39) of benchmarks for (00:08:40) >> compromising your ideals to win the (00:08:42) market in general. (00:08:43) >> Yeah. (00:08:44) >> Right. So so call it the safety knob is (00:08:46) the second knob that you can turn if if (00:08:48) you're in a pinch. The third knob that (00:08:50) you can turn is the post-raining knob (00:08:52) which can be done on relatively short (00:08:54) notice. So you can pick particular (00:08:56) benchmarks that you want to really (00:08:58) post-rain your models to to do really (00:09:01) well on. and and I I suspect all three (00:09:03) of these more compute maybe maybe not (00:09:07) some some turns of the safety knob uh (00:09:09) and then post-training on select (00:09:11) benchmarks is exactly what we we're (00:09:13) seeing in this cycle now that we have a (00:09:15) real horse race (00:09:17) >> I found it fascinating we've got (00:09:19) probably the most the fastest scaling (00:09:21) consumer platform in history we're (00:09:24) almost at a billion users that just (00:09:25) blows my mind (00:09:27) >> it's starting to eat the operating (00:09:29) system I mean like when you start to to (00:09:31) get order of magnitude a billion (00:09:33) downloads. At some point, you have to (00:09:35) ask the question, is this AI user (00:09:37) interface basically cannibalizing the (00:09:40) entire OS itself? At what point sometime (00:09:42) soon is every pixel that shows up on a (00:09:45) mobile device being AI generated? I (00:09:47) think we're not too far from that. (00:09:49) >> Wow. (00:09:49) >> Well, that was definitely the backstory, (00:09:51) too, when we were at Microsoft uh last (00:09:53) week with Mustafa Solom. I don't Is that (00:09:54) podcast out yet? I'm not I'm not sure (00:09:56) what the order of coming out shortly. (00:09:59) >> Yeah. Well, look forward to that one (00:10:01) because what Alex just said uh is (00:10:03) clearly in the minds of Microsoft. (00:10:05) They're going to do everything and (00:10:07) anything they can to get on this chart (00:10:08) that we're showing right now and they (00:10:10) have a lot of assets that that'll come (00:10:12) up in that pod that that'll give them a (00:10:14) really good chance of getting there. But (00:10:15) it's for exactly the reason Alex said (00:10:17) the the OS you the whole base of (00:10:19) Microsoft the revenue driver for the (00:10:21) last 30 years is at risk now and you you (00:10:23) got to move to the new thing or (00:10:25) >> it's not just it's not just OS, right? (00:10:27) It's the entire app ecosystem. Um I mean (00:10:30) the the end goal here is for these (00:10:34) hyperscalers to capture the user as the (00:10:36) only AI you need to use. So so-called (00:10:39) core subscription and that that (00:10:41) certainly is OpenAI's stated strategy to (00:10:43) become the default core subscription (00:10:46) quote unquote for consumers. Anthropics (00:10:48) strategy apparently is to focus on (00:10:50) enterprise APIs and codegen. XAI (00:10:52) focusing on brute force scaling and (00:10:54) maybe benchmaxing and Google focusing (00:10:57) maybe in a more balanced way on total (00:10:59) stack domination, balanced pre-training, (00:11:02) post- trainining. So I I I think in a (00:11:04) real horse race, which is what we're (00:11:05) finding ourselves in among the the top (00:11:07) four frontier labs, we're starting to (00:11:09) see differentiated strategies coming to (00:11:11) market. Every week, my team and I study (00:11:13) the top 10 technology metat trends that (00:11:15) will transform industries over the (00:11:17) decade ahead. I cover trends ranging (00:11:19) from humanoid robotics, AGI and quantum (00:11:21) computing to transport, energy, (00:11:23) longevity, and more. There's no fluff, (00:11:26) only the most important stuff that (00:11:27) matters that impacts our lives, our (00:11:30) companies, and our careers. If you want (00:11:31) me to share these meta trends with you, (00:11:33) I write a newsletter twice a week, (00:11:35) sending it out as a short two-minute (00:11:37) read via email. And if you want to (00:11:38) discover the most important meta trends (00:11:40) 10 years before anyone else, this (00:11:43) report's for you. Readers include (00:11:44) founders and CEOs from the world's most (00:11:46) disruptive companies and entrepreneurs (00:11:48) building the world's most disruptive (00:11:50) tech. It's not for you if you don't want (00:11:52) to be informed about what's coming, why (00:11:54) it matters, and how you can benefit from (00:11:56) it. To subscribe for free, go to (00:11:59) dmmandis.com/metrends (00:12:01) to gain access to the trends 10 years (00:12:04) before anyone else. All right, now back (00:12:05) to this episode. All right, let's jump (00:12:07) into the core story here today. OpenAI (00:12:10) releases GPT 5.2. We spun up this pod (00:12:13) for our subscribers the day after the (00:12:15) release so we can go into detail. What (00:12:17) does this mean? You know, we heard (00:12:19) OpenAI's red alert. Uh, and here's the (00:12:23) result. (00:12:24) >> Alex, take it away. (00:12:26) >> Yeah, I've been waiting for this all (00:12:27) day. Alex, (00:12:29) >> Dave, you want to lead us or or Alex (00:12:30) here? (00:12:31) >> Oh, no. I I just want to say that the (00:12:33) these numbers when they go from 80 to (00:12:34) 90, uh, it really understates the impact (00:12:38) uh, on what you can do. you know the the (00:12:41) benchmark and when it goes from 10 to 40 (00:12:42) it looks like a big gain on a line chart (00:12:45) but when it goes from 80 to 90 it (00:12:47) doesn't look like a big gain but what (00:12:48) you can do like firsthand is just (00:12:50) mind-blowingly different and I'll I'll (00:12:53) tell you some of the things I've done in (00:12:55) a minute but I've been waiting all day (00:12:56) to hear actually Alexart (00:12:58) who are listening versus uh versus (00:13:01) watching here's a chart of the (00:13:03) benchmarks uh comparing GPT 5.1 thinking (00:13:07) against GPT 5.2 to thinking uh and with (00:13:10) that if you don't mind uh sort of (00:13:12) speaking the percentages as well Alex as (00:13:14) we're going through this that would be (00:13:15) great. (00:13:16) >> Okay sure so maybe some highle comments (00:13:19) and then we can do a detailed (00:13:20) playbyplay. So highle comments. One, (00:13:24) keep in mind what I said a couple of (00:13:26) minutes ago. They're really if if you're (00:13:27) open AI and you need to rush an (00:13:30) impressive model release to market, (00:13:32) they're probably only three knobs you (00:13:34) have. One, you can turn up the compute. (00:13:36) Two, you can play safety games. And (00:13:39) three, you can do post-training on (00:13:41) particular eval particular benchmarks. (00:13:43) So that that story, maybe not the safety (00:13:46) story, but the other two knobs, I I (00:13:48) suspect is what we're seeing here. So (00:13:50) walking through this chart benchmark by (00:13:52) benchmark we have Sweepbench Pro which (00:13:54) is software engineering benchmark. We (00:13:57) see a modest improvement between 5.2 and (00:14:00) 5.1 perhaps attributable to mostly (00:14:03) compute and a little bit of more (00:14:05) post-training andor distillation. We (00:14:07) have Google proof question answering (00:14:10) diamond modest increase from 88.1% with (00:14:13) GPT 5.1 to 92.4 for again so far pretty (00:14:17) modest. We have uh charive reasoning a (00:14:22) larger increase. This is uh scientific (00:14:25) reasoning could be post-training not a (00:14:28) benchmark that I pay super close (00:14:29) attention to. Then we get to frontier (00:14:31) math uh frontier math tiers 1 through (00:14:34) three which are easier math problems. (00:14:37) And then one of my favorite benchmarks (00:14:38) of all time, Frontier Math Tier 4, which (00:14:41) is research grade problems in math that (00:14:44) are supposed to take professional (00:14:46) mathematicians several weeks to (00:14:48) accomplish. I often point to Frontier (00:14:49) Math Tier 4 and progress on Frontier (00:14:52) Math Tier 4 as indicative that uh drink (00:14:55) math is being solved. So, so focus (00:14:59) focusing on Frontier Math Tier 4, we see (00:15:03) Gemini 3 Pro getting approximately 19% (00:15:06) and GPT 5.2 thinking getting 14.6% (00:15:12) and GPT 5.1 thinking getting 12.5%. This (00:15:16) is actually a win in in my mind. This is (00:15:18) a win for Google and a loss for Open AAI (00:15:21) that OpenAI has had a month to to (00:15:24) attempt to to supercale to to beat (00:15:28) Google in this horse race at hard open (00:15:31) or rather hard closed math challenges (00:15:34) but professional mathematician grade (00:15:35) nonetheless couldn't beat Gemini 3 Pro (00:15:38) and it's it's not as if these problems (00:15:41) have been a state secret. In fact, (00:15:43) OpenAI actually sponsored Epic's (00:15:46) creation of the Frontier Math benchmark. (00:15:48) So, OpenAI has had in some sense (00:15:50) privileged access to all of Frontier (00:15:53) Math. Still couldn't beat Gemini. So, I (00:15:55) I think that's pretty instructive. (00:15:58) Moving down the list, Amy uh the (00:16:01) American Invitational Math Exam 2025 (00:16:04) scoring now 100% 5.2 versus 94% (00:16:08) suggestive of post-training. Then we get (00:16:11) to the the second set of benchmarks that (00:16:13) I think are super interesting. ARC AGI 1 (00:16:16) and two. ARC being autonomous research (00:16:19) challenge and of course AGI being AGI. (00:16:22) So for for those who don't pay super (00:16:23) close attention to ARGI, (00:16:26) ARGI sort of a visual reasoning (00:16:29) challenge testing whether problems that (00:16:31) humans find relatively easy sort of a (00:16:34) visual problem solving/program synthesis (00:16:37) challenge but machines historically have (00:16:39) found exceptionally exceptionally (00:16:41) difficult um as sort of an arbitrage (00:16:44) between human minds and machine minds. (00:16:46) We see here some big big differences. Uh (00:16:49) so for ARKGI1 the first version of the (00:16:52) prize we see that's just saturating at (00:16:54) this point 72.8% (00:16:57) with GPT 5.1 86.2% with GPT 5.2 ARGI1 is (00:17:03) cooked at at this point. ARGI 2 is (00:17:07) nearing the point of saturation. So so (00:17:09) huge change from 17.6% 6% with GPT 5.1 (00:17:14) to plus 50% 50 plus percent 52.9% (00:17:19) with GPT 5.2 thinking. So in my mind (00:17:22) this this smacks of post training that's (00:17:26) the obvious strategy. (00:17:28) >> Take a moment and just for those who (00:17:29) don't know what post training is because (00:17:31) I think it's an important uh one of the (00:17:33) three knobs that you you spoke about and (00:17:34) it's important for folks to understand (00:17:36) what that means. (00:17:37) >> Sure. So le let's reason by analogy to (00:17:40) the way humans uh in in sort of a (00:17:43) conventional western upbringing learn. (00:17:46) So you have the the sort of the the the (00:17:49) baby the infantlike learning that that's (00:17:52) approximately pre-training. So the the (00:17:55) the P in GPT stands for pre-trained. (00:17:58) Pre-training is unsupervised training. (00:18:01) you're you're feeding a model just (00:18:03) information about the world and giving (00:18:05) it the goal of predicting what comes (00:18:07) next. There there's not much of a (00:18:09) supervision angle to it and not unlike a (00:18:12) a human newborn where it's just taking (00:18:14) in information via lots of sensory feeds (00:18:17) and trying to make sense with with very (00:18:19) little guidance. Then there's (00:18:22) mid-training and post-training. So think (00:18:25) of these phases of training um as being (00:18:28) not unlike attending primary school, (00:18:31) secondary school where you you receive (00:18:34) explicit supervision. You're receiving (00:18:36) grading. You're being given particular (00:18:38) assignments. Uh and there are many ways (00:18:41) that you could be graded. You could be (00:18:42) graded very granularly like a thumbs up, (00:18:44) thumbs down, grade A, B, C, D, F. And (00:18:47) and there are other ways that you can (00:18:49) grade. for example, you can be given (00:18:51) more of an open-ended assignment and (00:18:54) graded on how well the ultimate final (00:18:56) product of that open assignment is. So (00:18:59) this sort of mid-training post-training (00:19:01) which really became popular with the (00:19:03) Oclass series of reasoning models from (00:19:06) OpenAI and everyone has since adopted (00:19:08) reasoning models and and post-training (00:19:11) not just to make humans happy which is (00:19:13) another form of post-raining like (00:19:14) pleasing your teacher but also showing (00:19:16) that you can via reinforcement learning (00:19:18) via other mechanisms solve hard problems (00:19:21) and reason about hard problems. This is (00:19:23) where post-raining shines. This is where (00:19:25) almost all of the the alpha if you will (00:19:27) in increasing model capabilities over (00:19:29) the past year or so has come from not (00:19:31) from pre-training. So getting back to (00:19:33) the benchmarks RKGI1 RKGI 2 these are (00:19:37) benchmarks the the R in RKGI is (00:19:40) reasoning these are benchmarks designed (00:19:42) to test the reasoning capabilities of (00:19:44) models and we see a huge jump we see (00:19:47) frontier level performance (00:19:48) state-of-the-art performance by GPT 5.2 (00:19:51) to with ARC AGI2 (00:19:53) reasoning is is well on its way to (00:19:56) having been solved at this point and I I (00:19:59) think we'll we'll cover this probably in (00:20:01) the next slide but the costs are (00:20:02) collapsing as well maybe talk about that (00:20:04) in a minute just to wrap up then uh for (00:20:08) purposes of of narrating this chart the (00:20:11) final benchmark here which is perhaps (00:20:13) the most interesting of all is GDP val (00:20:16) so GDP val gross domestic product eval (00:20:20) was created by OpenAI with the idea of (00:20:23) having an eval that measures AI ability (00:20:26) to automate knowledge work in the (00:20:28) general human service economy. So we're (00:20:31) seeing a jump from GPT 5.1 at 38.8% (00:20:36) GPT 5.2 is now at 70.9%. This is the (00:20:40) clearest indicator in my mind that the (00:20:43) human knowledge work economy is cooked. (00:20:46) You you heard it here. It is it's (00:20:48) cooked. It's this is 44 different (00:20:52) occupations that uh that OpenAI and by (00:20:55) the way this is like all open source. (00:20:56) You can go on GitHub and you can read (00:20:57) all of the the tasks for GDP val 44 (00:21:01) different human occupations 1320 (00:21:04) specialized tasks like creating (00:21:06) PowerPoint presentations or Excel (00:21:08) spreadsheets sort of typical (00:21:09) prototypical knowledge work (00:21:11) >> it's cooked it it's automated and 5.2 to (00:21:16) probably again due to elaborate post (00:21:17) train post training can get almost 71% (00:21:20) of these tasks. That's that's 70 the the (00:21:23) what does that actually means? 71% of (00:21:27) comparisons between a human performing (00:21:29) this knowledge work and the machine 5.2 (00:21:33) performing knowledge work resulted in (00:21:34) the machine doing a better job. And that (00:21:36) was by the way at more than 11 times the (00:21:39) speed of the human and at less than 1% (00:21:42) of the cost of the human professional. (00:21:44) So, knowledge work is cooked. (00:21:47) >> Okay. (00:21:47) >> You know, I figured I figured something (00:21:49) out on that last line this week, too. Um (00:21:52) because, you know, I'm I'm you know, (00:21:53) chairman of about a dozen companies and (00:21:55) I'm like, "Guys, what what is holding (00:21:56) you back? Why have you not deployed (00:21:58) this? You can cut costs dramatically. (00:22:00) You can automate. You can expand your (00:22:02) market share." And they're all like, (00:22:03) "Yeah, I don't know. We're really (00:22:05) struggling." Like, oh, it's driving me (00:22:06) nuts. What's going on? So, a couple (00:22:08) things that that I finally figured out. (00:22:10) One of them is uh you know one of the (00:22:11) companies is is working entirely in (00:22:13) Java. And when you turn this loose in (00:22:16) Python where it had a lot more training (00:22:19) data, it can build virtually anything. (00:22:21) It just blows your mind and it really (00:22:23) sucks in C still. And I don't think (00:22:25) they're going to fix it because they (00:22:27) just don't care. You know, we've moved (00:22:28) off of C anyway and there's there's not (00:22:30) enough training data and Java's (00:22:31) somewhere right in the middle. And so (00:22:32) when they benchmark it, they're like, (00:22:34) well, let me try and take my legacy (00:22:35) thing and see if it can just immediately (00:22:37) fix it. And it struggles. But if you (00:22:39) just say no, scrap it. Rebuild it (00:22:41) entirely from scratch in Python. You (00:22:42) come back an hour later and it's done. (00:22:44) >> So they they're stuck there. And also (00:22:46) the other place they're stuck is in (00:22:47) operations. (00:22:49) >> They're saying, "Well, look, the way we (00:22:50) pick up a customer service request is in (00:22:53) an email that's in an Outlook folder and (00:22:55) that has all these security whatevers on (00:22:58) top of it. So it's struggling to open (00:23:00) and read the emails." And so so we're (00:23:03) giving up. like do don't you think you (00:23:07) could maybe fix that front-end interface (00:23:09) in maybe a day and then try it on the (00:23:12) rest of the process and just turn it (00:23:14) loose and it would immediately crush the (00:23:15) problem. So they're they're stuck on (00:23:17) these little edge case issues. And and (00:23:19) I'll tell you it also comes up, you (00:23:20) know, that RKGI benchmark is the one (00:23:23) that was specifically designed to be (00:23:25) things that a human finds relatively (00:23:27) easy and intuitive and the AI is still (00:23:29) struggling with the AGI one and had (00:23:32) countless conversations around, you (00:23:34) know, academia with people who are (00:23:37) desperately want to say there's still (00:23:39) something missing. There's something (00:23:40) fundamentally missing in this great AI (00:23:43) brain and it hasn't been solved yet. and (00:23:45) the proof is ARK AGI1 (00:23:47) and you're like okay boy do you look (00:23:49) foolish now just two two you know three (00:23:51) weeks later five weeks later because (00:23:53) it's going to it's it's basically (00:23:55) saturated but it's going to be (00:23:56) completely saturated imminently (00:23:58) >> and on the GDP val you know if you (00:24:00) remember Elon has spoken about one of (00:24:02) the companies he's going to be starting (00:24:04) is macro hard and his mission is (00:24:07) basically go into a company and simulate (00:24:11) all of your employees and deliver it as (00:24:13) a service back to that company. Uh a lot (00:24:18) of change is coming rapidly. I think the (00:24:20) the biggest challenge is people are not (00:24:21) projecting properly on how rapidly this (00:24:24) is going to tip. Uh our next slide here (00:24:26) is GPT 5.2 ARC AGI update. Uh we spoke (00:24:31) about the numbers in the table just (00:24:33) recently. Here we see it charted out (00:24:36) where GPT has re has had a 390fold (00:24:40) efficiency improvement over uh over 03 (00:24:44) back from 2024. Anything you want to add (00:24:46) to this uh AWG? (00:24:49) >> Yeah. So, we talk on the pod, we've (00:24:51) spoken several times about (00:24:53) hypothetically 40 times 40x (00:24:56) year-over-year hyperdelation. We're (00:24:58) seeing 390x year-over-year hyperdelation (00:25:02) on visual reasoning for ARC AGI. This is (00:25:05) unprecedented. And it will not this (00:25:08) level of hyperdeflation in terms of the (00:25:10) cost of intelligence will not stay (00:25:13) contained to the data centers. It will (00:25:15) not stay contained to to these still (00:25:17) relatively narrow. I I know they they (00:25:19) brand themselves as generally (00:25:21) intelligent benchmarks, but they're (00:25:22) still relatively narrow in the scheme of (00:25:24) things. It's not going to stay (00:25:26) contained. hyperdelation is going to (00:25:28) spread outward from from these sorts of (00:25:30) benchmarks to the rest of the economy. (00:25:31) That that's comment one. Comment two, (00:25:34) just focusing narrowly on ARC AGI. One (00:25:37) of the the lovely things about the ARC (00:25:39) AGI 1 and two benchmarks is they don't (00:25:41) just focus on raw performance. They also (00:25:44) focus on cost. And if if it costs us (00:25:47) hundred trillion to solve a hard (00:25:49) problem, well, if it's if it's larger (00:25:50) than the human economy to solve an (00:25:52) important problem, then it almost (00:25:54) doesn't matter. But if it's incredibly (00:25:56) affordable, you know, to to to your (00:25:58) mantra, Peter, about abundance, if if (00:26:01) abundance is unaffordable, what's the (00:26:03) point? It has to be affordable (00:26:05) abundance. And and the way we get there (00:26:08) is exactly what the ARGI organizers do, (00:26:11) which is you measure on a scatter plot, (00:26:13) performance on the vertical axis, and (00:26:15) cost per task on the horizontal axis. (00:26:18) And that shows you what progress looks (00:26:19) like. You want progress that looks like (00:26:22) points in the scatter plot going up and (00:26:24) to the left. Greater performance at (00:26:26) lower cost. And and in fact, if if going (00:26:30) back to my earlier comments, if you see (00:26:31) a Frontier Lab hypothetically just (00:26:34) increasing compute costs but not (00:26:36) actually making efficiency gains, that (00:26:38) shows up in in these plots too. So you (00:26:40) can see for example if you look at RKGI1 (00:26:44) uh although it's probably a little bit (00:26:46) difficult to read here if you squint you (00:26:48) can see that GPT 5.2 2 is on sort of the (00:26:52) same the same extrapolated slope as GPT5 (00:26:56) mini suggesting that maybe at least as (00:26:59) it pertains to ARGI1 there hasn't (00:27:02) actually been major progress algorithmic (00:27:05) or efficiency progress it's just like (00:27:07) more compute being spent on on the same (00:27:10) tasks and so it feels smarter but it's (00:27:12) actually because you're putting more (00:27:13) work into it as as the apherism goes (00:27:15) you're you're lifting with your back not (00:27:17) with your legs but with ARGI 2 there is (00:27:20) in fact radical improvement. So we're (00:27:22) we're seeing progress. (00:27:24) >> Well, this is a benchmark that I think a (00:27:26) lot of people can relate to. The next (00:27:27) one here, GPT 5.2 writing benchmark (00:27:30) comparison, (00:27:32) long form creative writing and emotional (00:27:34) intelligence. Uh again, we're seeing (00:27:37) improvements across the board. Uh Alex, (00:27:40) one more interpretation here. (00:27:42) >> Spiky. This is very spiky. So spiky. So (00:27:46) we saw that we saw that sky that sort of (00:27:48) interesting three-dimensional plot on uh (00:27:51) when are we going to reach AGI and again (00:27:54) spikiness was was the descriptor for it. (00:27:58) >> That's right. That that that that spider (00:27:59) plot was purportedly comparing humans (00:28:02) with AGIS or strong models in in (00:28:05) general. What we're starting to see here (00:28:07) is increased spikiness and spiky (00:28:10) competition between the different (00:28:11) frontier models. So just a little bit of (00:28:13) context, long- form creative writing (00:28:15) benchmark evaluates model's ability to (00:28:17) basically write a novella uh about (00:28:20) 8,000word novella and as judged by (00:28:23) sonnet 5. And the the emotional (00:28:26) intelligence judge mark benchmark (00:28:28) measures how well a language model or a (00:28:30) model can grade short function. And so (00:28:33) what we're seeing here is no single (00:28:36) model dominating all the benchmarks. (00:28:38) We're seeing, for example, that with (00:28:41) long form creative writing, anthropics (00:28:43) on a 4.5 wins and is is the best job at (00:28:46) writing an 8,000word novella. (00:28:48) >> What do you use? What do you guys use (00:28:50) for for writing? I mean, I've been using (00:28:52) uh, you know, Gemini 3 Pro. Uh, it looks (00:28:56) like Claude, you know, Sonet 4.5 is the (00:28:58) one to go to. Um, are they all (00:29:01) >> I've been using? I've been using Gemini (00:29:03) 3 Pro and I found it to be really (00:29:05) amazing to just craft, but I'm using (00:29:08) mostly business documents. So, that's a (00:29:10) little different. (00:29:12) >> Same for me. I I use 3 Pro for almost (00:29:14) all of my writing. (00:29:16) >> Yeah, I'm using Kimmy K2 for huge (00:29:19) volumes of stuff on my little fleet of (00:29:22) of Nvidia chips that I hijacked. Um, and (00:29:26) then uh but I'm I'm using actually (00:29:28) Gemini to to one despyware it and uh and (00:29:33) to proofread it. Uh and I'm using uh (00:29:36) Claude Opus at extra, you know, my my uh (00:29:39) opus expenses went from 200 bucks a (00:29:41) month to a,000 bucks a month to I'll (00:29:43) easily crack 20 or 30k this month, but (00:29:45) I'll also generate more code this month (00:29:48) than my entire life up to this date. Um, (00:29:52) so it's a bargain at 20k, but my (00:29:54) expenses are going through the roof on (00:29:55) uh on anthropic, and I'm I'm happy with (00:29:58) it, actually. (00:29:59) >> Um, (00:30:00) >> spyware. What's what's despyware? It (00:30:02) mean (00:30:03) >> Well, Alex warned me that when you use a (00:30:05) Chinese open source model, it can inject (00:30:07) evil things into the code that it (00:30:10) returns to you. (00:30:10) >> This is actually publicly information. (00:30:12) We're not breaking news here. Maybe just (00:30:14) to expand on this. Uh, so so two two (00:30:17) comments. One comment is uh there there (00:30:21) have been very well publicized outside (00:30:24) of the pod studies that found for (00:30:26) example prompting certain openweight (00:30:29) models with certain politically (00:30:30) sensitive for certain countries topics (00:30:33) results in those models emitting more (00:30:36) vulnerable code for example that's (00:30:37) something to be wary of so I would say (00:30:40) like more broadly for creative writing (00:30:42) etc like none of these models is so (00:30:45) strong that I can ask them to do a good (00:30:48) job doing uh all writing. What I find (00:30:51) inevitably is I end up like having to do (00:30:54) 80% of the work and models function as (00:30:57) more of like a a junior editor as as it (00:31:00) were and I end up still doing like (00:31:01) majority of of work writing. Similarly (00:31:04) with to Dave's point with with codegen, (00:31:06) I I would certainly not trust codegen (00:31:10) models to not insert vulnerable code. So (00:31:14) >> yeah. Well, when you told me that a week (00:31:16) ago, I was like, you know, Alex, I'm (00:31:17) just I'm going to see the code and, you (00:31:20) know, I'll I'll see if it's injecting (00:31:21) anything evil in there. I'm not super (00:31:23) worried about it. Let's go. So, here we (00:31:25) are a week later and it's generating (00:31:28) volumes that no human being could ever (00:31:30) look at. I'm like, "Oh I was I was (00:31:32) completely wrong." Um, and it worked. (00:31:34) The code just flat out works. I don't (00:31:36) even have to look at it. It's passing (00:31:37) every eval. It's doing it's building (00:31:38) interfaces that I want. It's doing (00:31:40) everything I wanted to do without (00:31:41) needing to look at it. So now I've got (00:31:43) actually GPT 5.2 proofreading right now, (00:31:46) but I think what I need to do is just (00:31:48) turn off Kimmy, pay the 10x higher (00:31:50) price. It's actually 20x higher price to (00:31:53) to run it on GPT2 uh instead. Um 5.2 (00:31:57) instead. (00:31:58) >> Yeah. And but I'm going to have to do (00:32:00) that because I don't I don't know how (00:32:01) else to make sure I don't end up spyw (00:32:03) wearing my entire world. You know, it's (00:32:05) >> this is a real challenge. If if you have (00:32:07) basically intelligence being dumped in (00:32:09) into the world, then there is this (00:32:11) implicit trade-off between do you want (00:32:13) intelligence cheap or do you want it to (00:32:14) be safe? (00:32:16) >> Yeah. I mean, and we've talked about (00:32:17) this as a potential strategy for China (00:32:20) making open- source models available to (00:32:23) the world, it becomes if it becomes the (00:32:26) base on which you've built everything, (00:32:28) uh, then it's it's there from the (00:32:31) beginning. Um, I don't want to impute a (00:32:34) a a dystopian point of view on all the (00:32:37) the Chinese model makers, but it is a (00:32:40) concern. I (00:32:41) >> I think we're going to see a move to (00:32:43) sovereign intelligence. I I I think that (00:32:45) this is the long-term trajectory we find (00:32:46) ourselves on. Every sovereign entity is (00:32:48) going to want their own sovereign (00:32:50) trusted stack. (00:32:52) >> Well, how do you feel about France? So, (00:32:54) Mistl's uh uh Devstral 2 raises the bar (00:32:57) in open source coding tools. Uh, so, uh, (00:33:01) what do you think about about, uh, about (00:33:04) Mistl, Dave? Are you playing with him at (00:33:06) all? (00:33:07) >> You know, it's funny. I saw this chart (00:33:08) and I had kind of forgotten all about (00:33:10) him and, uh, I guess my read on the (00:33:12) chart was, "Oh, it exists." (00:33:15) But, you know, the headline says it (00:33:17) raises the bar, but it's actually below (00:33:20) the I mean, only a notch, but it's below (00:33:22) Kimmy and Deepseig. I guess you could (00:33:24) probably trust it more because Europe is (00:33:25) much very trustworthy. Um but other than (00:33:28) that it was like what's the news here? (00:33:30) >> It's the headlines Europe slow but (00:33:33) trustworthy. Okay. And and also it's not (00:33:36) I mean so so there there's I think this (00:33:38) this sense for a variety of reasons that (00:33:40) Mistral is somehow like the the EU's (00:33:43) sovereign AI stack or sovereign AI (00:33:46) model. But it its roots are all very (00:33:48) much American. Uh all of its early (00:33:50) funding is is from blue chip American (00:33:53) VCs. Its founding team came from deep (00:33:56) mind and meta. it. Yes, it's like raised (00:33:59) a large amount of money from ASML most (00:34:01) recently and I I my understanding is (00:34:03) Europe is very interested in using Mistl (00:34:06) as sort of an AI emissary to the rest of (00:34:08) the world but it its technical roots are (00:34:10) deep deep in the US and sort of this (00:34:13) bizarre world that we find ourselves in (00:34:14) where it's a Parisbased Frontier Lab or (00:34:18) Neolab however they brand themselves (00:34:20) that's the the right now the only and (00:34:22) main counterweight to Chinese openweight (00:34:24) models. There's one thing I thought was (00:34:27) really interesting here. If as it's (00:34:28) getting close once when you once you (00:34:30) have open- source systems being beating (00:34:32) closed systems then you move innovation (00:34:35) to the community level from the lab from (00:34:37) the lab and there's no catching up with (00:34:39) it once you get that flywheel going. So (00:34:42) I thought this was a big deal once they (00:34:44) may need a little bit more improvement (00:34:45) per Dave's point but I think once they (00:34:47) get there (00:34:48) >> is true is that true for AI open source (00:34:51) models? (00:34:53) Um, I know it's true for uh, you know, (00:34:56) for a multitude of of fundamental just (00:34:58) plain software models. We've seen that (00:35:00) before. Alex, do you think (00:35:02) >> it's tricky? It's tricky because like (00:35:05) you have to ask what are the primary (00:35:06) limiting factors to increasing (00:35:08) capabilities and it's it's compute more (00:35:10) than talent. There's lots of talent in (00:35:12) the world, but compute is still pretty (00:35:14) scarce. So the the community has lots of (00:35:17) talent, but I I in my mind (00:35:20) >> they don't have comput. They're comput (00:35:21) starved. This isn't like Linux where you (00:35:23) can sort of say lots of eyeballs make (00:35:25) all bugs shallow. In this case, the the (00:35:27) way you make the bug shallow is by (00:35:28) investing trillions in capex. (00:35:30) >> Well, this conversation is critically (00:35:32) important. And Alex, you can you can (00:35:34) help the world a lot because every (00:35:35) corporate executive in 2026 is going to (00:35:38) need to choose something. And you know, (00:35:41) there's only two types of exec out (00:35:42) there. People that are familiar with (00:35:43) this and they've already kind of got (00:35:45) their their landscape figured out. And (00:35:48) then the other 99% that are going to get (00:35:50) slapped in the face in 2026 and have to (00:35:53) react and they're late to the party. But (00:35:55) you saw the benchmark earlier. (00:35:57) Everything every one of your employees (00:35:58) can do can now be done by AI. What are (00:36:01) you going to do? Just sit there and (00:36:02) ignore that? So in 2026 is the turning (00:36:04) point. But these choices are really (00:36:06) tough on this chart. Like to an (00:36:08) executive saying, "Well, god, I can go (00:36:09) open source at 120th the price, but I (00:36:12) get 72.2 2 ambiguous units of thing or (00:36:16) for 77.9 (00:36:18) like what does that mean? It means a (00:36:20) lot. Anyone looking at the chart would (00:36:22) say oh what's the big deal? It's only (00:36:24) five units. But the reality is the (00:36:27) capability difference in terms of you (00:36:29) know your economic value is massively (00:36:31) massively bigger as this goes up even a (00:36:34) little bit. And so it's a tricky tricky (00:36:36) situation in 2026 for pretty much all of (00:36:38) corporate America corporate world. I I (00:36:41) think it's probably I mean, if I had to (00:36:42) to spitball this one, I think it's going (00:36:44) to take some sort of regulation to to (00:36:47) move the dial on this. Right now, if if (00:36:49) you hang out with with all the Silicon (00:36:51) Valley firms that are using openweight (00:36:53) models, they're just all using Alibaba's (00:36:55) Quen at this point. And Mistrol and (00:36:58) Devstral, that that's great, but it's uh (00:37:01) it's probably in the mind of a typical (00:37:03) Silicon Valley firm that needs to host (00:37:05) their own models, too little too late. (00:37:07) They're all using Quen. They're all (00:37:08) fine-tuning Quinn and it's going to take (00:37:11) an executive order or an act of Congress (00:37:13) or or some sort of regulatory measure to (00:37:16) turn off the the cheap Chinese (00:37:18) openweight intelligence before they're (00:37:20) incentivized to to move over to Mistl or (00:37:22) Devstrol or or GPTOSS. But Dave, I think (00:37:25) one of the points that you made is the (00:37:28) CEO and the board of directors of a (00:37:31) company in uh extremis in in sort of (00:37:35) paralysis not knowing what to do, (00:37:38) >> right? And and their lunch is going to (00:37:40) be eaten by the small startup that says, (00:37:42) "Oh, there's an interesting uh business, (00:37:44) so we should go and enter." and it it (00:37:47) builds a AI native approach uh that (00:37:51) 100th the cost and you know 10x the (00:37:54) innovation uh evolution speed and so (00:37:56) what do they do um you know who do they (00:37:59) turn to to help them reorganize their (00:38:03) company and it's a risky move because if (00:38:06) you brought in an outside consulting (00:38:08) firm right I don't think it's going to (00:38:09) be the biggest the big consultants I (00:38:11) mean they're going to be AI native (00:38:12) companies uh out there we're going to be (00:38:14) having a a pod convers sation with one (00:38:17) company called Invisible that does this (00:38:20) very shortly and there there are others. (00:38:22) Um (00:38:24) the right way to do it, you said it (00:38:26) earlier, is to scrap what you've been (00:38:28) doing and actually start with a fresh (00:38:31) stack and that is so hard for any any (00:38:35) company to do. See? (00:38:37) >> Yeah, this is uh right in our (00:38:39) wheelhouse. Uh essentially we're working (00:38:40) with some very big companies and Dave, (00:38:43) you're exactly right. They're totally (00:38:45) paralyzed. They're flailing. They have (00:38:47) no idea what to do. And if they bring in (00:38:49) one of the traditional consulting firms, (00:38:51) they just push them faster down the old (00:38:53) path, right? And so that doesn't work at (00:38:55) all. And so what needs to happen is they (00:38:58) need to take their capability here, (00:39:00) create a new stack on the edge that's (00:39:03) completely built AI native from the (00:39:05) ground up, and then little by little (00:39:06) deprecate the old and move (00:39:08) functionality, capability, resources to (00:39:10) the new. the political and the uh (00:39:13) emotional uh stress of that is causing (00:39:16) them most of them to do nothing. (00:39:18) >> Yeah. (00:39:18) >> And so out of the say 20 major companies (00:39:22) we're working with uh maybe three are (00:39:25) doing um maybe 50% of the right thing (00:39:29) and and most of them are just like uh (00:39:31) we're going to keep pushing this old (00:39:33) model and seeing where where we get to. (00:39:35) Surely we can catch up because we've (00:39:37) always been able to get there before. (00:39:38) And the answer is you absolutely cannot. (00:39:40) And so this (00:39:41) >> Macy's, it's Blockbusters. And when you (00:39:43) say we, you mean Openex EXO is doing (00:39:45) some work with these companies out (00:39:46) there. (00:39:46) >> Yeah, we have like 42,000 people talking (00:39:48) to companies around the world. And so we (00:39:50) were kind of a aggregating the gathering (00:39:53) the information of all that9. (00:39:56) >> I think 2026 is going to see the biggest (00:39:58) collapse of the corporate world in the (00:40:01) history of business. (00:40:02) >> You've heard that first prediction here. (00:40:03) No doubt because I think this is going (00:40:05) to be and we should have maybe a end of (00:40:07) year perspective and some predictions. (00:40:08) But (00:40:09) >> for all of the madness we've seen in (00:40:11) 2025, it's like this is the slowest it's (00:40:14) ever going to be in 2026 going to be 10x (00:40:16) to 50x to 100x crazier. So I I don't (00:40:20) even know where to start. (00:40:23) And I've got benchmark fatigue right now (00:40:25) >> to dealing with all this. If you hire (00:40:28) Seem to help you with your strategy, one (00:40:30) of the things he'll tell you is read (00:40:32) Klay Christensen, The Innovator's (00:40:34) Dilemma, which exactly addresses this (00:40:36) question. And what that book will tell (00:40:38) you to do and Klay Christensen's (00:40:40) foundation will tell you to do, go find (00:40:43) Link Studio, Y Combinator, Neo, go out (00:40:47) there and find your AI development (00:40:50) partners. (00:40:51) try and do a deal with them where you (00:40:53) either invest in them or you become a (00:40:55) development partner customer for them. (00:40:57) Pull them in, give them revenue because (00:40:59) their market cap will go way up. They'll (00:41:01) all become wealthy, but they'll then (00:41:02) hire the talent. But point them at your (00:41:04) internal problem and have them solve it (00:41:07) inside your organization as an outside (00:41:09) very tightly bounded startup company (00:41:11) that's growing like crazy. That's the (00:41:13) only way you're going to get the talent (00:41:15) focused on your internal problems. You (00:41:16) can't hire the talent directly anymore. (00:41:18) You got billion dollar signing bonuses. (00:41:20) Yeah. you know, all over the place. (00:41:21) >> And by the way, Sem will tell you not (00:41:23) will tell you to go read open exo2 (00:41:26) exponential organizations too, which is (00:41:28) our book, which actually walks through (00:41:30) step by step what to do, (00:41:31) >> how to do this. Yeah. I actually had a (00:41:33) couple of really interesting (00:41:34) conversations with Clay uh before he (00:41:37) passed away. And one of the things he uh (00:41:39) honestly very honestly admitted was the (00:41:42) innovator's dilemma works really well to (00:41:44) identifying the the uh cracks in the (00:41:48) structure but it's not that great at the (00:41:50) prescriptive side or trying to predict (00:41:53) for example in his model Uber is not (00:41:55) very disruptive. And I said well Uber (00:41:57) and he said I said but Uber is very (00:41:59) disruptive. It fits right into the (00:42:01) wheelhouse of our exo thing. And he goes (00:42:03) yeah it means our model's wrong. And the (00:42:05) the when we drilled into it, what we (00:42:07) realized was his the innovator's dilemma (00:42:09) assumes that the verticals like (00:42:11) transportation, energy, healthcare, (00:42:13) education stay in those verticals. So (00:42:16) Uber as a transportation company may (00:42:19) disrupt a little bit of transportation, (00:42:21) but not realizing it, it's also (00:42:22) disrupting healthcare delivery and (00:42:24) restaurant delivery and food delivery (00:42:26) and can go horizontal across a lot of (00:42:28) these. And so there's a the the old (00:42:30) verticals are essentially collapsing of (00:42:32) the old uh newspaper with the printed (00:42:35) places saying utilities and this and (00:42:37) this and this. And to Alex's point, it's (00:42:39) all going to become one category called (00:42:41) compute. And that's where that (00:42:43) >> well if you don't want to do what Selma (00:42:44) is suggesting, the other choice is to do (00:42:46) a $20 billion aqua hire plus 14 billion (00:42:48) of new payroll. And and that's the other (00:42:50) way to solve the problem. (00:42:52) Or I tell you the other thing I'm see (00:42:54) the other thing I'm seeing that's (00:42:56) unbelievable executives at that level (00:42:58) are are go looking at every looking at (00:43:00) the world and going yeah I'm just going (00:43:01) to retire right now and so there's this (00:43:03) unbelievable (00:43:04) >> stop opting out exactly like falling off (00:43:07) the cliff going (00:43:08) >> it's the most fun time in human history (00:43:10) how can you not not diving into the (00:43:13) ground (00:43:15) I actually respect that I tell you why (00:43:17) what they're doing is they're basically (00:43:18) saying I can't navigate this new world (00:43:21) I'm a (00:43:22) and let the younger generation navigate (00:43:24) this because I can't do it. (00:43:28) >> But it's very honest, right? At least (00:43:30) the worst thing in the world is the old (00:43:32) fddy duddies that are running the world (00:43:34) on the old model that can't that won't (00:43:35) get out of the way and we're seeing that (00:43:37) much more in politics to some extent in (00:43:39) the corporate world. This massive change (00:43:42) happening. (00:43:43) >> So talk about talk talk about billion (00:43:45) dollar salaries, talk about innovators (00:43:46) dilemma. Our next story here is Meta's (00:43:49) shifting AI strategy is causing internal (00:43:51) confusion. So Meta is at an inflection (00:43:54) point right after mixed Llama for (00:43:57) results and a reported 14 billion AI (00:44:00) talent spending spree. Uh you know Mark (00:44:04) is looking at considering whether an (00:44:06) open-source strategy can still compete (00:44:09) with closed vertically integrated rivals (00:44:11) like OpenAI and Google. (00:44:14) Dave, what do you think about this? (00:44:16) I I think they're doing exactly the (00:44:18) right thing. Actually, the the other (00:44:19) backstory here, which I guess is (00:44:21) validated, maybe it's more rumor than (00:44:22) validated, but they're they're getting (00:44:24) heavily into distillation of other (00:44:27) people's models to accelerate the (00:44:29) inference time speed. And what's (00:44:31) exciting about that is if you look at (00:44:33) where we are in human history, you know, (00:44:35) intelligence in a box was invented just (00:44:38) days ago, you know, or well, really two (00:44:40) years ago, but but it's brand new in the (00:44:42) world. And now we're in the (00:44:44) hyperexperimentation phase of how do we (00:44:47) make it bigger better by by running many (00:44:50) many agents in parallel by expanding the (00:44:52) context window and dumping in tons more (00:44:55) data um and by iterating it over and you (00:44:58) know chain of thought reasoning it over (00:45:00) and over and over again and we're (00:45:01) getting ridiculous gains but we're brand (00:45:03) new in that game and so what Meta has (00:45:06) realized is look we're behind in the (00:45:08) foundation model race we do need to (00:45:09) rebuild and catch up but that's not (00:45:11) going to happen overnight but where we (00:45:12) can potentially get ahead is by raw (00:45:15) inference time speed and having many (00:45:18) many more agents working on things in (00:45:20) parallel. And I believe that that will (00:45:22) also lead to self-improvement which will (00:45:24) get them back on the map. And so I think (00:45:26) that they're directing all their (00:45:27) research energy now into how do we make (00:45:29) this (00:45:30) >> blazing fast and be the world leader in (00:45:32) distillation. (00:45:34) That's my (00:45:35) >> incredible. I'm blown away by the 14 (00:45:37) billion dollar hiring spree. Just like (00:45:39) that number. I can't I can't process (00:45:41) that number. (00:45:42) >> Well, remember they've got just a (00:45:44) massive cash cow and cash flow and and (00:45:46) cash uh generating engine. And you know, (00:45:50) Mark has basically said this is the race (00:45:53) if we don't spend the money now uh to (00:45:56) get, you know, towards number one. It's (00:45:59) it will just slowly slowly go away. So (00:46:02) Dave, what you're saying (00:46:03) >> and what's cooler than cool is that he's (00:46:04) already decided to use every single (00:46:06) penny of it plus debt on top of that to (00:46:09) try and win this race. And Wall Street (00:46:10) has said that's fine. No damage to the (00:46:12) stock. Go for it. We love We love what (00:46:15) you're saying. That's just a beautiful (00:46:16) thing. (00:46:17) >> So what you're saying is they're moving (00:46:18) from like trying to focus on the open (00:46:20) source of the foundation model to (00:46:22) putting all of their chips on the agent (00:46:24) strategy. (00:46:26) >> Well, so much innovation there, too. (00:46:29) >> Yeah, I think they're in in a bit of a (00:46:30) tricky situation. So, so I I I I know (00:46:32) the key players. Zuck's undergrad (00:46:34) adviser before he dropped out was my (00:46:36) post-docctoral adviser. Natt Freriedman (00:46:38) who's with Alexander Wang helping to to (00:46:41) lead this this new lab was my first (00:46:43) roommate at MIT. I I'm pretty familiar (00:46:45) with the key players in this particular (00:46:47) story. And I I think there are three (00:46:50) strategies that Meta could be pursuing (00:46:52) andor has been pursuing. So one strategy (00:46:56) is that of commodify your compliment (00:46:59) drive the the cost of generative AI to (00:47:02) zero. That that was their llama strategy (00:47:04) that they were pursuing. Problem is (00:47:07) llama 4 was a disaster and the Chinese (00:47:09) openweight models are flooding the (00:47:11) market and doing a much better job. The (00:47:13) second strategy that they could be (00:47:15) pursuing is more conventionally and and (00:47:18) perhaps what Wall Street would expect (00:47:20) out of Meta, use strong AI to improve (00:47:23) like Instagram uh and and other Meta (00:47:26) products. And so uh that I would have to (00:47:30) imagine many executives at Meta would (00:47:32) like to see all of these new AI (00:47:34) resources being used to just improve (00:47:36) Meta's other existing products. Strategy (00:47:39) two. Strategy three is compete directly (00:47:42) with the frontier labs with closed (00:47:44) source API based models to be the first (00:47:47) to super intelligence. So I think what (00:47:49) Meta has to struggle with it's almost (00:47:52) hopefully not like a civil war (00:47:53) internally but what they have to decide (00:47:55) is which of those three three strategies (00:47:58) do they really want to pursue and my (00:48:00) guess is there are constituencies with (00:48:02) different interests within meta that (00:48:03) want to pursue each one of those three. (00:48:05) >> I cannot believe Mark is not all in on (00:48:07) number three. I mean being first to (00:48:09) super intelligence that just feels like (00:48:12) Mark's MMO. (00:48:13) >> Yeah. (00:48:13) >> Yeah. (00:48:14) >> And (00:48:15) >> Yeah. And I think I think very often the (00:48:17) cover story is look we're going to (00:48:18) enhance existing products. We're going (00:48:20) to use our internal data. You know we've (00:48:22) got a huge amount of internal posts that (00:48:23) we can use as training data. That's all (00:48:25) kind of cover story for the real we want (00:48:27) to win the race to AGI and ASI. By the (00:48:30) way, everybody, I want you to realize as (00:48:32) you're hearing these stories about (00:48:34) Google, about Meta, it's all about (00:48:37) business model innovation. Uh, on top of (00:48:40) all of this, right, Google going from an (00:48:42) adbased search company to now an AI (00:48:46) based company that's delivering a whole (00:48:48) slew of different products. Meta is I (00:48:50) mean this is where companies fail when (00:48:52) when Blockbuster did not change their (00:48:54) business model even though they had (00:48:57) twice the opportunity to buy Netflix, (00:48:59) right? So how do you actually disrupt (00:49:02) your own company and and shift its (00:49:04) business model? Um otherwise it's it's (00:49:08) game over. (00:49:09) >> Innovator's dilemma to Dave's point (00:49:11) earlier. (00:49:12) >> Yeah. But it is I I I think also ironic (00:49:14) like Sam Alman has has said publicly (00:49:17) that he'd much rather have a billion (00:49:19) users with not Frontier model than vice (00:49:22) versa. And yet and yet what we see from (00:49:24) Meta is the exact opposite strategy. (00:49:26) Meta already has their billion users (00:49:29) billion plus users but they would much (00:49:31) rather have a frontier model at this (00:49:33) point than no one's ever (00:49:35) >> that grass is always greener at the (00:49:37) other frontier lab. That's funny. (00:49:41) That's a that's a that's a good phrase. (00:49:43) All right. Our next story here is Google (00:49:45) DeepMind to build material science lab (00:49:48) after signing deal with UK. So we've (00:49:51) heard about this as well. Another (00:49:53) company um out of MIT and Harvard called (00:49:56) uh Laya uh is doing something very (00:49:59) similar where you're basically you know (00:50:02) it's all about the data and if you've (00:50:04) consumed all the data you need to go (00:50:05) find new data. So imagine having a you (00:50:09) know lights out robotic capability where (00:50:12) the AI is putting forward a scientific (00:50:14) hypothesis designing experiments and (00:50:17) then at night uh robots in the lab are (00:50:21) running the experiments to get the data (00:50:23) to either confirm or modify your (00:50:25) hypothesis and like let's do that a (00:50:28) thousand times or 10,000 times faster (00:50:30) than humans can do. Uh it's uh I think (00:50:34) we're going to see multiple companies. I (00:50:36) think every frontier lab is going to (00:50:38) need to have this kind of data mining. (00:50:39) We're data mining nature, understanding (00:50:42) what's going on. In particular here, (00:50:43) they're focusing on material sciences. (00:50:45) Uh Laya is looking at biological (00:50:48) sciences. Uh thoughts on this gentleman? (00:50:52) >> I don't know if there's a poly market on (00:50:54) this, but Dennis is really leading the (00:50:55) race to being the coolest guy on earth. (00:50:57) Um he got his Nobel Prize in chemistry. (00:51:00) now he's going to crack computer and (00:51:02) this you you kind of could see this (00:51:03) coming because you know AI can allow you (00:51:06) to be a worldleading expert in anything (00:51:09) and you know he's the master of the (00:51:10) biggest AI you know compute in the world (00:51:14) and algorithms and TP he's got he's got (00:51:16) the (00:51:17) >> and he also isn't uh he's not one of the (00:51:19) corporate you know leaders trapped in (00:51:20) the political prey and (00:51:22) >> beautiful (00:51:25) we're going to have the coolest guy (00:51:27) benchmark okay (00:51:28) >> well you what's great is you want (00:51:31) somebody with that purity at the edge of (00:51:34) this which is fantastic. There's a (00:51:36) couple of things I thought came across (00:51:38) for me having kind of hunkered around in (00:51:40) physics labs during my degree. The if (00:51:43) you have a fully autonomous lab, this is (00:51:46) like the biggest breakthrough in (00:51:47) scientific progress since the scientific (00:51:49) method was invented because we talked (00:51:51) about dark kitchens and dark factories (00:51:53) and now we have dark labs. Holy crap. (00:51:56) >> Yeah. You know, it's funny too. I can (00:51:58) only find like just a handful of people (00:52:00) like Demis, Alex on this pod. There's (00:52:03) there's like 10 or 12 that I could name (00:52:05) that can tell you the implications, you (00:52:08) know, in all these other, you know, in (00:52:09) in biotech, in material science, in (00:52:11) chemistry, in math. You know, Alex is (00:52:13) talking about solving all math. It's (00:52:15) just such a small group of people who (00:52:16) see where this is going to take us and (00:52:18) how short that timeline is. So, it's (00:52:19) good to see Dennis doing material (00:52:20) science. (00:52:21) >> This is AI assisted science and AI (00:52:23) native discovery. Alex, you want to (00:52:26) close us out on the subject? Th this is (00:52:28) what comes after super intelligence. (00:52:29) What comes after super intelligence is (00:52:31) solving math, drink, science, comma, (00:52:34) engineering, comma, and medicine. And (00:52:37) yes, math is being solved. We we've (00:52:39) spoken about that perhaps ad at nauseium (00:52:42) at this point on the pod. We haven't (00:52:44) spoken as much about AI solving all of (00:52:46) material science. And there are like a (00:52:48) dozen companies. It's not just Google. (00:52:50) It's not just Laya. It's not just (00:52:51) periodic. that there are a dozen (00:52:53) companies that are all laser focused on (00:52:56) solving material science and that's (00:52:58) going to give us so many upsides. It's (00:53:00) also when we talk about recursive (00:53:02) self-improvement having better (00:53:03) semiconductors having better (00:53:05) superconductors for science is at the (00:53:08) foundation upon which everything else is (00:53:10) built. (00:53:12) >> The medium here we come (00:53:14) >> and the innermost loop accelerates (00:53:16) again. (00:53:17) >> Yeah. And by the way, uh, for our new (00:53:19) listeners, our new subscribers, if you (00:53:21) hear Alex saying drink, there's been a (00:53:23) bingo game sort of invented for, uh, (00:53:26) terms that are repeated on a regular (00:53:27) basis. You'll you'll be hearing it. All (00:53:29) right, let's move on to our next story (00:53:31) here. Um, (00:53:33) uh, and I don't know how I feel about (00:53:36) this story. I I sort of feel like I (00:53:38) don't want to like overblow over, you (00:53:40) know, overexpose what's been already (00:53:43) overblown, but uh this is a story of an (00:53:46) AI native character called Tilly (00:53:49) Norwood. Uh and she's an AI, you know, (00:53:53) native actress that's freaking out (00:53:54) Hollywood. So, Tilly Norwood is an AI (00:53:57) made actress created by a London studio (00:54:00) uh to star in films and social media. Uh (00:54:03) built over six months with GPT. Uh Tilly (00:54:07) went through 2,000 design versions and (00:54:10) YouTube videos have garnered over (00:54:11) 700,000 views in October. We're we're (00:54:14) see we saw this also in the music (00:54:17) business where fully AI native bands and (00:54:20) and music tracks have been created and (00:54:22) people don't even realize they're (00:54:24) listening to something that's just fully (00:54:26) AI generated. (00:54:27) >> Uh (00:54:28) >> she has her own agent. (00:54:30) >> Yeah. (00:54:31) >> And reportedly like 40 different (00:54:33) contracts for for movies and other (00:54:35) development projects. This is I I would (00:54:37) say like this is consistent with my my (00:54:39) modal hypothesis that over the next 10 (00:54:41) years we're going to live out the plot (00:54:43) of every sci-fi movie ever made. In this (00:54:46) case, this is actually I don't know if (00:54:48) you saw the movie Simone. Uh this was (00:54:51) the the plot of the sci-fi movie Simone (00:54:53) where an an AI actress develops a life (00:54:56) of her own, takes over. It has Alpuccino (00:54:58) in it. It's it's a fun movie, but like (00:55:01) we're we're going to see AI actors and (00:55:03) and actresses take over potentially, or (00:55:06) at least we'll we'll discover how soon (00:55:07) humans crave authenticity in their (00:55:09) entertainment. (00:55:10) >> There's no doubt in my mind that that (00:55:12) humans do not crave authenticity as much (00:55:15) as we think we do and we will just watch (00:55:17) whatever is interesting and (00:55:18) entertaining. And I was at the (00:55:20) Washington Post when you know every (00:55:21) reporter there was saying (00:55:23) >> you know the post will be fine because (00:55:25) people will want genuine great reporting (00:55:27) from great reporters who are struggling (00:55:29) in the field to find the stories. That (00:55:32) was right before Yeah. Guess again. (00:55:34) Gone. Just gone in and in in just a (00:55:36) couple years too. The timeline was so (00:55:38) much shorter than than they ever would (00:55:40) have thought. From from top newspaper in (00:55:43) the world, multigenerational been in the (00:55:44) family for three generations to gone. (00:55:47) Jeff Bezos bought it for cents on the (00:55:49) dollar in just what, three years, four (00:55:52) years. So, that's going to happen here, (00:55:54) too. Uh, and no doubt in my mind, it's (00:55:55) going to happen with music. It's going (00:55:56) to happen with movies. It's going to (00:55:58) >> Yeah, it's inevitable. (00:55:59) >> This is This is an AI performer working (00:56:02) 24/7, (00:56:03) uh, appearing on in unlimited projects, (00:56:06) never aging, never burning out, uh, (00:56:08) never needing to renegotiate contracts. (00:56:11) I mean, this is the Screen Actors Guild (00:56:13) worst nightmare. I had dinner a couple (00:56:16) of nights ago with a dear friend on my (00:56:18) ex-prise board who used to be the head (00:56:20) of two of the major studios and then uh (00:56:24) an actress uh who's another dear friend (00:56:26) and we were talking about this and it it (00:56:30) is scaring the daylights out of the (00:56:32) industry uh and the I mean it's it's (00:56:36) >> well no good because they'll react and I (00:56:38) don't I'm not I I wish nothing but good (00:56:40) to happen to the people that are in the (00:56:42) industry but good that they're scared (00:56:43) because then they'll react as opposed to (00:56:45) getting crushed. I didn't mean to. (00:56:46) >> Well, well, the question become the (00:56:48) question becomes then what's the (00:56:50) response, right? Are you as an actor (00:56:53) going to license (00:56:55) your persona because that's the way (00:56:56) you're going to make money in the final (00:56:58) result? Because if you don't then then (00:57:00) the industry will simply or you know the (00:57:02) next generation industry will simply (00:57:04) create a Tilly Norwood who actually is (00:57:08) cuter than you or more handsome than (00:57:09) you. Uh able to (00:57:11) >> doesn't age. (00:57:12) >> Doesn't age. (00:57:13) >> Oh yeah, there you go. Doesn't age. (00:57:15) That's a huge one. Um, I'll tell you one (00:57:17) thing. (00:57:20) >> I wonder when you'll have one of these (00:57:22) winning the Oscar, (00:57:23) >> right? Because in theory, in theory, (00:57:25) they should be the best. (00:57:26) >> We have a lot of those benchmarks. When (00:57:28) will the first AI win a Nobel Prize, (00:57:30) right? When will the first AI, you know, (00:57:32) build a (00:57:32) >> billion (00:57:34) already did it cuz he's kind of half (00:57:35) anyway. (00:57:36) >> That's done. (00:57:36) >> It's squishy. Also, there have been, by (00:57:38) my count, at least two Nobel prizes. (00:57:39) There was Demis with with Alpha Fold in (00:57:42) and chemistry and then there was also (00:57:43) Jeff at all with restricted Boltzman (00:57:46) machines for physics. I the the squishy (00:57:49) thing here is it you can always do a (00:57:52) secret cyborg as as some would say and (00:57:54) wrap AI talent inside a a human meat (00:57:57) body and the human claims the credit for (00:57:59) it. So I I it's unclear again like how (00:58:02) much humans crave authenticity. Does (00:58:05) this become a separate category in the (00:58:07) in the Oscars like animation? Is this (00:58:10) sort of an an increment on top of (00:58:12) animation that's real life animation or (00:58:14) is this an actual labor substitute? I (00:58:17) don't know yet. (00:58:18) >> I think a lot of that thinking though a (00:58:20) lot of that thinking is is a little bit (00:58:21) misguided in that what what the actors (00:58:24) will be looking for is a featurelength (00:58:26) movie in a theater where it's all AI and (00:58:29) that's what they're going to use as (00:58:30) their bellweather for the threat. But (00:58:31) that's not what's going to happen. And (00:58:32) if you look in the data, short form (00:58:34) video is taking over the movies anyway. (00:58:37) And video games are already miles ahead (00:58:39) of movies. (00:58:39) >> We had these conversations. Kids don't (00:58:41) go to the movies. They watch YouTube (00:58:42) videos. It's all (00:58:43) >> Exactly. So Tilly Tilly will end up (00:58:45) being a star in every video game and (00:58:47) also Tik Tok clip (00:58:49) >> across (00:58:49) >> and they'll say, "Well, that's not a (00:58:50) threat. That's not a threat to Yeah. (00:58:52) across platforms." And the actors will (00:58:53) say, "Well, that's not a threat to me. (00:58:54) I'm a real actor. I do Shakespeare and (00:58:56) you know, whatever." Like, well, no, it (00:58:57) is a threat to you because the audience (00:58:59) has moved and the budget has moved and (00:59:01) that'll undercut you. So they're looking (00:59:02) at the wrong bell weather. When when (00:59:04) Tilly shows up in five billion Tik Tok (00:59:07) posts, that's when you know you're dead (00:59:10) long before it hits you in your long (00:59:12) form movies. So you just got to look at (00:59:15) the video games, too. (00:59:16) >> A related story of this, which is OpenAI (00:59:18) is working with Disney uh to bring (00:59:21) Disney characters into Sora 2, (00:59:23) >> right? So that (00:59:25) >> Yeah, they just announced that. (00:59:26) >> Yeah, it's a fascinating. So (00:59:28) >> a billion dollar investment and and (00:59:30) licensing. I I I I think there's going (00:59:33) to be a certain fungeability between (00:59:35) classic IP assets and and generative (00:59:39) everything and and so for maybe in the (00:59:41) short to medium-term it's a three-year (00:59:43) reportedly licensing agreement that (00:59:45) OpenAI and Disney struck. Maybe in the (00:59:47) short term the the short-term remedy is (00:59:50) existing actors can license their Visage (00:59:53) out as an asset to customers who want to (00:59:56) do sort of fan pickics. But they're (00:59:58) really if you're like a really popular (01:00:00) star like a Peter Diamandis, you know (01:00:02) what's the thing you should do right (01:00:03) away? (01:00:03) >> Signed Sign my rights already. (01:00:06) >> Yeah. (01:00:06) >> Get your avatar out there. Get it built (01:00:09) and out there right away. Get your Tilly (01:00:10) Tilly Norwood equivalent, Peter or (01:00:12) whoever out there right away so that (01:00:14) personality can grab before you know the (01:00:17) true synthetics take over. (01:00:19) >> Yeah, it really is going to be a race (01:00:21) for neurons, right? if if you're looking (01:00:25) you're going to you know the general (01:00:27) public you know Dunar's number only (01:00:29) really cares about 150 people and holds (01:00:31) them close and so the question is are (01:00:34) one of those or 10 of those going to be (01:00:36) synthetic actors um and once you get to (01:00:41) a point of popularity uh it's going to (01:00:44) be hard to replace you (01:00:47) >> for for for what it's worth uh at maybe (01:00:49) to tie a bow on this also the Dunar (01:00:51) limit of 150 people that that was like (01:00:54) in the ancestral environment if if the (01:00:57) number is is valid at all in in the post (01:00:59) social media era you can maintain light (01:01:01) casual associations with thousands of (01:01:04) people and (01:01:05) >> but Dunar's number is is basically sort (01:01:08) of the human tribe and I've done this (01:01:09) when I was running Singular University (01:01:11) when it's it's the number of people you (01:01:13) can actually uh remember their names go (01:01:16) deep with and so forth sure you can have (01:01:18) a rolodex of 22,000 people but Dunar's (01:01:22) number in terms of who you feel (01:01:24) connected to closely is is a real (01:01:26) number. (01:01:27) >> I I'm with Alex on this one. What I (01:01:29) noticed was once you have Facebook and (01:01:31) you could essentially Facebook acted as (01:01:33) your RAM for Dunar, you could move (01:01:35) people in and out of that spectrum very (01:01:37) easily without really noticing. And you (01:01:39) have the opposite effect also where once (01:01:41) you kind of start to connect with enough (01:01:43) people. Peter, you've probably had this. (01:01:44) I remember walking down University (01:01:46) Avenue in Palo Alto uh right after one (01:01:49) of our exec oneweek executive programs (01:01:51) and this guy stops me and he goes hey (01:01:52) Sim nice to see you and I'm like have we (01:01:55) have we met he said we I just spent the (01:01:58) week in the classroom with you right and (01:01:59) I wow like it's like our our brains are (01:02:02) to blown up now with the limits of that (01:02:04) we need technology to expand that (01:02:06) capability and it's start it's already (01:02:08) done that to one extent and we can move (01:02:11) things in and out the question is what (01:02:12) do we do when we have all these (01:02:13) synthetic AI levels (01:02:15) going going through that. So, (01:02:17) >> this episode is brought to you by (01:02:18) Blitzy, autonomous software development (01:02:21) with infinite code context. Blitzy uses (01:02:24) thousands of specialized AI agents that (01:02:27) think for hours to understand (01:02:29) enterprisecale code bases with millions (01:02:32) of lines of code. Engineers start every (01:02:35) development sprint with the Blitzy (01:02:37) platform, bringing in their development (01:02:39) requirements. The Blitzy platform (01:02:41) provides a plan, then generates and (01:02:44) pre-ompiles code for each task. Blitzy (01:02:46) delivers 80% or more of the development (01:02:49) work autonomously while providing a (01:02:51) guide for the final 20% of human (01:02:54) development work required to complete (01:02:56) the sprint. Enterprises are achieving a (01:02:59) 5x engineering velocity increase when (01:03:01) incorporating Blitzy as their preIDE (01:03:04) development tool, pairing it with their (01:03:06) coding co-pilot of choice to bring an AI (01:03:08) native SDLC into their org. Ready to 5x (01:03:12) your engineering velocity? Visit (01:03:14) blitzy.com to schedule a demo and start (01:03:16) building with Blitzy today. (01:03:21) Our next story here comes out of the (01:03:22) White House. Trump signed an executive (01:03:24) order curbing state AI rules. So, this (01:03:28) is a decisive federal power grab uh over (01:03:31) AI regulations. Trump's one rule (01:03:34) executive order is going to preempt (01:03:36) state level AI laws. Uh it's like, nope, (01:03:39) it's not going to be, you know, (01:03:42) Washington uh Washington DC is going to (01:03:45) win over everybody. It's not California (01:03:47) laws or Texas laws. It's Washington DC. (01:03:50) I mean ultimately I think this is what (01:03:53) the EU needs as well. Um it needs top (01:03:57) level direction. It's going to be harder (01:03:58) there. Any particular thoughts on on the (01:04:01) one rule (01:04:02) >> here? It's absolutely positively (01:04:04) necessary. I hate I hate it when this (01:04:07) happens but we got to do it. Um because (01:04:09) you know variety across states is one of (01:04:11) our best assets. On the other hand, New (01:04:14) York just passed a law that says you (01:04:16) can't use the likeness in an AI of (01:04:18) somebody who's deceased without going to (01:04:19) their ancestors. Like, what what about (01:04:21) all these Einsteins floating around (01:04:23) already? Like, how are you going to keep (01:04:24) it out of New York? There's no way to (01:04:26) just launch it across the country and (01:04:28) then New York users get blocked somehow. (01:04:30) I mean, it's just it's just unworkable. (01:04:32) Uh, so (01:04:33) >> I'm going to I'm going to claim I'm one (01:04:34) of Aristotle's ancestors and you can't (01:04:36) use his like I mean how far back (01:04:39) >> down Aristotle's (01:04:41) >> I had two I had two thoughts when I saw (01:04:42) this one was when I saw one rule I very (01:04:45) quickly thought about one ring to rule (01:04:47) them all. Um uh and I just love the (01:04:50) politics of this where the a huge amount (01:04:52) of the effort for Trump was saying let's (01:04:55) push all the thing down to states rights (01:04:57) and now we're going totally the opposite (01:04:59) direction and I think it's a necessary (01:05:01) thing. I agree with uh Dave here. It has (01:05:03) to be done because if we don't get um (01:05:05) uniform AI treatment, uh where the hell (01:05:08) are we going to get to? (01:05:09) >> I also I mean there's an interstate (01:05:11) commerce angle here. Models are being (01:05:13) trained in one state and inferenced in (01:05:15) in other states. That this in my mind (01:05:18) and I I I read the executive order um in (01:05:20) the past 24 hours. The the EO is (01:05:22) ensuring a national policy framework for (01:05:24) artificial intelligence. it. I I I think (01:05:27) this is it's it's both reasonable under (01:05:31) the interstate commerce clause and also (01:05:33) necessary for for international (01:05:36) competition. It's not at all obvious how (01:05:38) a patchwork of state-based regulations (01:05:40) results in anything other than total (01:05:42) chaos. (01:05:43) >> I mean, this is this is a piece of the (01:05:45) overall White House strategy on energy, (01:05:48) on data centers, on chips. uh it's all (01:05:52) aligning everybody to make the US as (01:05:55) competitive as possible on the global (01:05:57) stage and to accelerate as fast as (01:05:59) possible. It is a race to super (01:06:01) intelligence. Um and this is just part (01:06:03) of the uh (01:06:05) >> can I make a radical prediction here? (01:06:07) >> Yeah, of course. (01:06:08) >> This is this the over the next 5 years (01:06:10) the entire US constitution will (01:06:12) evaporate. Uh every clause is starting (01:06:15) to just melt away. I look the right to (01:06:17) privacy fourth amendment gone right. Um, (01:06:20) we're we're we're going to see the whole (01:06:22) thing. It needs to be rewritten from the (01:06:24) ground up and it's going to be (01:06:25) interesting to see how that happens. And (01:06:26) I will move that. (01:06:28) >> Boom. That's what you need. (01:06:30) >> The instead of the founding fathers, (01:06:31) it's the founding models. Um, (01:06:35) >> for for the record, I I don't buy that (01:06:36) prediction for one second. (01:06:38) >> Good. We can put some money on it. (01:06:41) >> That's poly markets, baby. All right. (01:06:43) Uh, let's move to a conversation on the (01:06:45) economy. Uh and you know this is data (01:06:49) just to support what we already know. (01:06:51) OpenAI finds AI saves workers nearly an (01:06:54) hour a day on average. Uh so workers (01:06:58) using OpenAI tools have saved between 40 (01:07:00) to 60 minutes a day. The survey of 9,000 (01:07:03) people in 100 companies found that 75% (01:07:06) say AI makes work faster or better. The (01:07:09) biggest time saver over a million (01:07:11) businesses today are using open AI (01:07:14) tools. I'm going to couple this story (01:07:16) with our next one, which is layoffs uh (01:07:20) announced, you know, 2025, we had 1.1 (01:07:23) million layoffs, which is the most since (01:07:25) the 2020 pandemic. (01:07:28) All right. U Dave, you want to jump on (01:07:30) on this? I was talking to Scott Perry, (01:07:33) the CEO of Tree Lending Tree, a public (01:07:35) company yesterday actually, and he said (01:07:38) 20,000 incredibly talented people in (01:07:40) Seattle are now cut loose from Microsoft (01:07:43) and Amazon, and it's the best hiring (01:07:46) opportunity for tech talent he's ever (01:07:48) seen in his life. But these are really, (01:07:50) really solid great people that the mega (01:07:53) tech companies have just cut out because (01:07:55) AI is automating, improving, enhancing. (01:07:59) you know, coding is one of the biggest (01:08:01) early beneficiaries and you know, my my (01:08:03) top coders are 10 times more productive, (01:08:05) so I don't need nearly as many. So, (01:08:07) that's where the layoffs are coming (01:08:08) from. But this is just uh you know, (01:08:09) we'll look back on this and say, "Wait, (01:08:11) what? That was a bell weather. I why did (01:08:14) I not notice this little thing, you (01:08:16) know, and but when you see what happens (01:08:17) in 2026, you'll say, "When did this all (01:08:19) start?" Well, right now, this is when (01:08:20) it's (01:08:21) >> Now, what do you predict for 26, Dave? (01:08:24) >> Continued. (01:08:26) >> Yeah. the the capabilities (01:08:28) will be you know able to eliminate on (01:08:31) the order of 80 90% of all jobs (01:08:34) but then the roll out and the (01:08:36) percolation is dependent on regulation (01:08:38) and also corporate bureaucracy (01:08:41) and so it's it's tough to predict how (01:08:44) quickly people will react. My my guess (01:08:46) is that it'll get a very slow start. (01:08:48) Everybody's very stodgy. Um, but then (01:08:51) everyone's a sheep. And when somebody in (01:08:53) your industry is an early adopter and (01:08:55) their stock goes up 10x just because (01:08:57) they're an early adopter, then your (01:08:58) board beats you up like crazy and says, (01:09:00) "What's what about us?" And then the (01:09:02) sheep effect flips in 2026. So by the (01:09:05) end of 2026, everyone's in absolute (01:09:08) panic mode and then they're wishing they (01:09:10) started at the beginning of 2026. You (01:09:12) know, I I think there's going to be, (01:09:14) this is one of my predictions, I think (01:09:15) there's going to be a absolute need for (01:09:19) all the medium-siz and large companies (01:09:21) to bring in a reskilling uh consultancy (01:09:26) uh some type of a program could be a (01:09:29) fully AI based, but uh that provides (01:09:32) some kind of a safety net for your (01:09:34) employees that you're going to reskill (01:09:37) people before you fire them, and if they (01:09:39) aren't able to be reskilled, then (01:09:41) they're let (01:09:42) I also think that's a huge business (01:09:44) opportunity for an entrepreneur out (01:09:46) there to build that kind of capability. (01:09:48) >> Totally. Totally right. In fact, you (01:09:49) know, if we look in our portfolio, the (01:09:51) companies that are quote unquote floor (01:09:53) deployed (01:09:54) >> um are killing it. And you know, if you (01:09:57) couple that with what we just said, (01:09:58) there's 20,000 highly talented people in (01:10:00) Seattle that just got cut loose. If (01:10:02) you're if you're growing your business, (01:10:04) a lot of the younger companies uh you (01:10:07) know 22 23 year old leaders are afraid (01:10:11) to be forward deployed because they've (01:10:12) never they've never done it before. They (01:10:14) don't have any management experience. (01:10:15) They don't have any enterprise sales (01:10:16) experience. (01:10:17) >> Well, get hire hire those 20,000 people, (01:10:20) train them on how to be AI forward (01:10:23) deployed consultants or delivery people (01:10:25) and then get them embedded back into (01:10:27) corporate America at State Street Bank, (01:10:28) at at JP Morgan, at Walmart. they'll (01:10:31) they'll hire your people instantly to (01:10:34) get AI deployed inside their (01:10:36) organization because they can't get that (01:10:37) talent. But if you grab grab those (01:10:40) people, retrain them very very quickly (01:10:42) on your own AI training platform and (01:10:44) then get them redeployed into corporate (01:10:46) America, your growth rate, you you'll be (01:10:48) sold out every time you have a meeting, (01:10:50) you'll you'll generate a sale. (01:10:51) >> So the founders, the really young (01:10:54) founders are afraid to do it. They want (01:10:55) they want to just like launch their (01:10:57) software on hacker news and hope that (01:10:59) the world sucks it up and it there's (01:11:02) just this big gap between there and (01:11:04) where corporate America starts and it's (01:11:05) it's just never going to fill if you (01:11:07) don't get forward deployed. (01:11:08) >> Um I don't think this is a skills issue. (01:11:10) This is a cultural problem. The problem (01:11:12) is in corporate America with all the (01:11:14) structural impediments in a big company, (01:11:16) you need a mindset shift at scale (01:11:20) company to even adopt this. I think I (01:11:22) think the large companies and the (01:11:24) medium-sized companies to be very (01:11:25) specific about my prediction here are (01:11:27) going to need to hire a very specific (01:11:31) kind of consultancy, right? A company (01:11:34) that comes in and their job inside your (01:11:37) company and I think every company's (01:11:39) going to have a version of this is (01:11:41) reskilling. And so that when you go to (01:11:43) work for a company, you know, there's a (01:11:46) a reskilling (01:11:48) um you know safety net there for you. (01:11:51) Exo. Yeah. Um, (01:11:55) but what I'm saying is it's not just (01:11:56) reskilling. It's a mindset shift. Real (01:12:00) change. It's a cultural change that has (01:12:02) to take place. And that's actually much (01:12:04) harder. And I want to say two things. (01:12:07) >> There's cultural and mindset shift at (01:12:09) the CEO, at the executive level, and at (01:12:11) the employee. (01:12:12) >> All of them. (01:12:13) >> It goes through it goes through the (01:12:14) organization. Uh, and we've actually (01:12:17) been working on this for several years (01:12:19) now. And I want to tell a quick story. (01:12:21) our second ever client when we finished (01:12:23) one of our 10-week sprints realized that (01:12:25) they had to lay off a thousand people in (01:12:27) the company and they decided what are we (01:12:29) going to do because we're a family-owned (01:12:31) business. We have uh we want to really (01:12:32) provide for these folks. What do we do? (01:12:34) We actually got them to give them a (01:12:36) one-year UBI so that they could find (01:12:39) their own passion, find their own work, (01:12:41) and if they didn't by the end of the (01:12:42) year, they would try and hire them back. (01:12:44) And it was an incredibly successful (01:12:46) program. I think we're going to see a (01:12:47) lot more of that as we kind of transform (01:12:50) the workforce. (01:12:50) >> All right, let's get into data centers, (01:12:52) chips, and energy. Um, we're seeing data (01:12:55) centers begin to pop up in countries (01:12:57) around the world. I don't want to spend (01:12:59) too much time on this, but Qar uh you (01:13:02) know, QIA, the uh the sovereign fund (01:13:05) there is investing 20 billion to launch (01:13:07) a data center in Qatar or Qatar or (01:13:11) however you want to pronounce it as a (01:13:12) Middle East hub. uh we're seeing (01:13:14) Microsoft uh and Sat and Satya just (01:13:17) coming back from India meeting with (01:13:19) Prime Minister Modi there committing (01:13:21) 17.5 billion in India to expand an AI (01:13:24) ready cloud uh there in the region so (01:13:28) we've got uh I mean this is going to be (01:13:30) the case in all major nations these (01:13:34) partnerships taking place the real (01:13:37) >> this is Alex's comment about tiling the (01:13:39) world with data centers and everyone (01:13:41) >> drink tile the earth with sovereign (01:13:43) Inference time compute. Drink drink. (01:13:46) >> Okay. But we're drinking coffee this (01:13:49) morning, ladies and gentlemen. (01:13:50) >> Drinking water. (01:13:51) >> Alcohol. All right. So, uh here's here's (01:13:55) the story I want to dig into. You know, (01:13:57) in our last pod, we talked about China's (01:14:00) uh sort of incredibly expanding role. (01:14:02) So, China is set to limit access to (01:14:05) Nvidia's H200 (01:14:07) chips despite uh Trump's export (01:14:10) approval. So, you know, President Trump (01:14:13) says to Nvidia, "Okay, you can export (01:14:15) these." And now the China leadership is, (01:14:17) "No, no, no, you can't buy them. You (01:14:19) need to buy Chinese-made uh and you (01:14:22) know, GPUs." Uh, fascinating, right? (01:14:26) It's this is propping up its own chip (01:14:29) economy. I think it's a smart move on (01:14:32) China's behalf. (01:14:33) >> This is so fun and annoying at the same (01:14:36) time to watch. You know, this is pure (01:14:38) protectionism. The US never did it (01:14:40) before and now we're now we're playing (01:14:41) the game. But you know what happens is a (01:14:44) country invents something like an LCD TV (01:14:46) or a car or you know whatever and (01:14:48) another country says okay what we're (01:14:50) going to do is we're going to protect (01:14:51) the home market. We're going to (01:14:52) manufacture our own. Then we're going to (01:14:53) dump it on your market cheaply and we're (01:14:56) going to dump it until your companies (01:14:58) collapse and the venture capitalists all (01:15:00) run away and then we're going to price (01:15:02) it up. So what we did is we embargoed (01:15:05) the chips from China and they're like, (01:15:07) "Oh we need to build our own whole (01:15:10) supply chain." And as soon as they get (01:15:11) it up and running, we're going to say, (01:15:13) "Oh, no, no, it's okay. Now we're going (01:15:15) to actually allow you to buy the H200's (01:15:18) and that entire thing you just built (01:15:20) makes no economic sense." And so China's (01:15:22) saying, "All right, I I see what you're (01:15:24) doing here. I've played this game for a (01:15:25) long time. We're not going to we're not (01:15:27) going to buy them." Like, but why? You (01:15:29) know, it's an incredible buy. Why why (01:15:31) would you not allow us to buy them? cuz (01:15:32) we we already made a massive investment (01:15:34) in our own fabs. We're going to have to (01:15:36) keep subsidizing that to get this up and (01:15:38) running cuz we know what you're doing (01:15:39) here. You're going to let us buy them (01:15:41) right up until our stuff collapses and (01:15:43) then you're going to cut it off again. (01:15:45) >> This is it's a trust issue. (01:15:47) >> Big trust issue. (01:15:48) >> There's no trust at all between the US (01:15:50) and China right now. (01:15:51) >> Well, this the same thing happened, (01:15:52) right? The Japanese came over during (01:15:55) Trump's first administration and spent a (01:15:57) lot of time negotiating a trade deal and (01:16:00) and then in just a few months ago, Trump (01:16:03) um the administration cancelled that (01:16:05) trade deal. And the Japanese are like, (01:16:07) "We're not negotiating another one (01:16:09) because we don't know which way up is (01:16:11) anymore." And every single time it (01:16:13) changes completely. So there's no trade (01:16:15) deal. And this is really a a big problem (01:16:18) going forward. And I think what China is (01:16:20) saying is we don't want to play that (01:16:21) game. (01:16:22) Well, there's no doubt that the outcome (01:16:24) is look, two completely separate (01:16:25) ecosystems. You Europe is kind of a wild (01:16:28) card. It's interesting and and so is (01:16:30) India is kind of a wild card right now, (01:16:31) but there's no doubt the US ecosystem is (01:16:33) going to grow completely independent of (01:16:35) the China ecosystem because there's no (01:16:37) chance of reestablishing trust after (01:16:39) that chip embargo. (01:16:40) >> Yeah. There's like no way that that's (01:16:42) going to get get mended. (01:16:44) >> That's right. So, sovereign data center (01:16:47) AI compute to Alex's point, (01:16:50) >> it's it's a new It's almost like a (01:16:52) second cold war. It it's it's a a world (01:16:54) that we move to where there are spheres (01:16:56) of influence and spheres of fab and (01:16:58) spheres of compute and the decoupling (01:17:00) happened. (01:17:02) >> Yeah. (01:17:03) >> I f Okay, move on to power generation. (01:17:06) Uh there's a company called Boom. Uh (01:17:09) many years ago, it set out to build the (01:17:12) first supersonic uh passenger airliner (01:17:16) to replace the Concord. And I was so (01:17:18) impressed by the the founder and CEO, (01:17:20) his hutzbah, if you would, to take on (01:17:22) this moonshot to build a supersonic (01:17:26) consumer airplane. And I was like, I (01:17:28) don't know how you get there. How much (01:17:30) money is going to be required uh to to (01:17:32) build this. So, it's a fascinating backs (01:17:36) stop that Boom had been developing, you (01:17:39) know, supersonic uh engines and now (01:17:43) they've unveiled a supersonic super (01:17:45) power turbine uh that's able to provide (01:17:50) 42 megawatts of natural gas turbine (01:17:53) capabilities (01:17:54) uh to data centers. Um and so this is, (01:17:58) you know, a backstop business model uh (01:18:01) for Boom. uh and it's and it's huge, (01:18:04) right? So, uh this is moving power to (01:18:11) the data centers, right? It's uh it's a (01:18:14) gas turbine strategy and we've heard (01:18:16) before all the gas turbines have been (01:18:18) sold out for some time. Uh Alex, you (01:18:21) want to jump on this? (01:18:22) >> Yeah, I mean the as you were were (01:18:24) gesturing, Peter, that the wait times (01:18:26) right now for gas fired turbines for AI (01:18:29) data centers are seven years in some (01:18:31) cases. So I I think this is a brilliant (01:18:33) strategic pivot by by boom, it also to (01:18:36) the extent referencing comments from a (01:18:39) minute ago to the extent we're in almost (01:18:41) a quasi second cold war. This is is (01:18:44) almost like a self-directed defense (01:18:46) production act type move pivoting (01:18:48) resources perhaps from turbines for (01:18:52) supersonic consumer jets to turbines for (01:18:55) AI data centers. And of course there (01:18:57) there are synergies there, but this is I (01:18:59) think it's a brilliant pivot. And the (01:19:00) the irony is there's probably a much (01:19:02) much larger addressable market for gas (01:19:05) turbines for AI data centers than there (01:19:06) is for consumer supersonic jets at this (01:19:09) point. I I just hope for the sake of (01:19:10) boom that that they retain at least some (01:19:13) semblance of the original supersonic (01:19:14) vision and just don't get overwhelmed by (01:19:16) the AI data center business. (01:19:17) >> I just love that audio clip. Hey, hey, (01:19:20) behind the scenes, I need that audio (01:19:22) clip like right away. That that is (01:19:24) >> because you know there's so many (01:19:25) companies including Vesmark you know one (01:19:27) of the ones I founded preai (01:19:29) >> uh you know manages $2 trillion of (01:19:31) assets 20 million lines of code (01:19:33) profitable great business and I'm like (01:19:36) guys you got to be an AI company like (01:19:37) tomorrow (01:19:38) >> pivot pivot (01:19:39) >> pivot pivot pivot pivot we've got (01:19:41) >> you know so this is a great case study (01:19:43) like you you you wouldn't think that a (01:19:47) jet engine (01:19:49) is company is culturally going to pivot (01:19:51) and become a power generation company, (01:19:53) but when you look under the cover, it's (01:19:54) like, well, what are our assets here? (01:19:55) Well, we've got the blades, we've got (01:19:57) the manufacturing, we've got metal, you (01:19:59) know, like that's all it takes. The age (01:20:00) of AI has so much opportunity that (01:20:02) didn't exist the day before. And you (01:20:05) don't have to be that close to the (01:20:06) center point. You have to be adjacent (01:20:08) and just pivot quickly and you and (01:20:10) you'll succeed wildly. And so I I hope (01:20:12) these guys just crush in fact I know (01:20:13) they'll crush it cuz cuz like you said, (01:20:16) Alex, they I I know personally (01:20:19) >> data center operators that yeah they (01:20:20) they'll spend anything and they're and (01:20:22) they're pre- buying too. They'll pay you (01:20:24) upfront for something that you're going (01:20:25) to make next year (01:20:27) billion dollar backlog. Uh and it's a (01:20:30) product they can deliver immediately, (01:20:33) right? This is on premise power (01:20:35) generation for data centers which is so (01:20:37) critical. You know, they've been (01:20:39) working, Boom's been working on this (01:20:41) for, I don't know, six, seven, eight (01:20:43) years, and they've built the scale model (01:20:45) of their supersonic airplane, and (01:20:47) they're trying to get advanced orders (01:20:48) from all of the airlines. But to get (01:20:52) through the FA thicket is so difficult, (01:20:54) decade, that's decades, (01:20:55) >> it will kill you. But if you've got a an (01:20:58) actual business model delivering revenue (01:21:00) right now, I mean, I I agree with you, (01:21:02) Alex. I hope Boom actually delivers on (01:21:05) their original idea. I think this (01:21:06) increases the probability a huge amount. (01:21:09) Right? Then and this is the equivalent (01:21:10) of uh of Amazon realizing with uh Amazon (01:21:15) web services, it's got something that it (01:21:17) can offer uh to everybody else that (01:21:21) makes you know very strong near-term (01:21:22) profits. (01:21:26) >> Elon or Elon like delivering Starlink (01:21:29) now and Mars colony in 10 years. (01:21:31) >> Yeah, it's (01:21:32) >> that's the sexiest looking gas turbine (01:21:34) I've ever seen, by the way. (01:21:36) beautiful looking thing. (01:21:37) >> I'm sure after you run it, it gets (01:21:39) dirtier. (01:21:40) >> 1.25 billion in backlog. Congratulations (01:21:42) to the team at Boom for that strategic (01:21:45) pivot. And everybody else, (01:21:46) >> everybody learned from this story. Like (01:21:48) we should track this uh you know in a (01:21:49) few weeks or a few months. (01:21:51) >> What do you have? What do you what are (01:21:52) you building right now that's a cost (01:21:54) center for you that could become a (01:21:56) profit center for you in the AI (01:21:59) ecosystem? That's the question. All (01:22:02) right. On the energy side, China builds (01:22:04) nuclear reactors at $2 per watt versus (01:22:08) the US at $15 per watt. Uh, again, (01:22:12) what's going on here? Why is that why is (01:22:15) that happening? Alex, do you have a (01:22:16) thought? (01:22:17) >> Yeah. Well, China does have more people (01:22:19) than the US. China does have a need for (01:22:22) more energy. If if there if AI were not (01:22:24) part of this equation and and China were (01:22:27) to attain US per capita energy footprint (01:22:30) standards, China would need more energy (01:22:34) than in in a total sense in an absolute (01:22:36) sense than the US. That that part makes (01:22:39) sense. What doesn't make sense if if you (01:22:41) look at the permitting processes (01:22:43) required for nuclear energy in the US, (01:22:46) it's a very different beast. There are (01:22:48) obviously the the the NRC regulates US (01:22:52) nuclear power deployments at the (01:22:53) national scale, but then on top of that, (01:22:55) you have some states that de facto ban (01:22:58) nuclear power entirely. We have a (01:23:00) patchwork of state and local regulations (01:23:02) that make it extremely difficult to to (01:23:04) deploy nuclear energy. Here in (01:23:06) Cambridge, Massachusetts, many people (01:23:08) not may or may not be aware of this. (01:23:10) Cambridge has a nuclear reactor. It's (01:23:12) it's not very well advertised. It's on (01:23:14) Massachusetts AB. on the the MIT campus, (01:23:17) but we have a working nuclear reactor (01:23:19) and and have had one since I think the (01:23:21) the late '60s, early '7s, but that (01:23:24) that's very much like not par for the (01:23:27) course in the US. I wouldn't be (01:23:28) surprised if sometime in the next 2 to 3 (01:23:32) years, we see some equivalent for (01:23:34) nuclear energy of of what we just saw (01:23:36) with the White House's executive (01:23:38) >> to see it in the next few months. I mean (01:23:40) the bottleneck is not physics, it's (01:23:44) permitting and execution and that's got (01:23:46) to be cleared. (01:23:47) >> Yeah, (01:23:49) >> I'll give you a little uh side story (01:23:50) related to this. Um you the MIT brand, (01:23:53) here's the MIT brand. The MIT brand is (01:23:55) absolutely skyrocketing in this AI (01:23:57) revolution. But we found out that that (01:23:59) MIT nuclear reactor is going to be (01:24:00) exothermic and powering the campus. And (01:24:03) I'm like, wow. Because we don't have a (01:24:05) single nuclear reactor in the state, you (01:24:07) know, we can't get that approved. We buy (01:24:08) our nuclear power from New Hampshire, (01:24:10) but MIT can actually get stuff like that (01:24:13) done now. Just crazy how how that brand (01:24:16) has skyrocketed in impact with this AI (01:24:19) revolution. All right, want to jump into (01:24:21) robotics. A special uh you know hat (01:24:23) tipping here to Sem. This is Sem's (01:24:27) perfect robot. It's got something like (01:24:29) 14 different arms on it. See, are you (01:24:31) happy with this robot? (01:24:33) >> This looks awesome. Look at all the (01:24:35) chickens that can move around very (01:24:36) quickly. Um, this this is this is Yeah, (01:24:40) I love it. Just love it. (01:24:41) >> For those of you new to the pod, See is (01:24:44) having a running debate about, okay, why (01:24:46) humanoid robots? Why just two why just (01:24:48) two arms? Well, Seem, you've got all the (01:24:50) arms you could possibly put on a body (01:24:52) here. (01:24:53) >> I just love all the wires sticking out (01:24:55) of it. Also, like it looks (01:24:56) >> I mean there there is a serious story (01:24:58) here too, like in in China there's an (01:25:01) image doing doing (01:25:04) >> I can't wait for that. (01:25:05) >> Yeah. doing doing the rounds with six (01:25:08) arms that there I don't think there's (01:25:10) anything like super Yeah. (01:25:11) >> Yeah. I was going to bring that I was (01:25:13) going to bring that article forward as (01:25:14) well. (01:25:14) >> Yeah. There is (01:25:16) not about six armed robots. Yes. Coming (01:25:19) out of (01:25:19) >> China is not about having a humanoid (01:25:21) robot. It's about mimic it's about (01:25:22) integrating into human spaces and and (01:25:25) kind of moving around where humans have (01:25:27) been. And so there there's some case for (01:25:29) it. But in general there's it's very (01:25:31) easy to be 10x more efficient than a (01:25:33) human being. We're we're very very (01:25:35) inefficient in most of the things that (01:25:37) we do. (01:25:38) >> Yeah. I think evolution has done (01:25:40) evolution has over billions of years or (01:25:42) maybe order of magnitude a billion years (01:25:44) done a search through body space. And (01:25:46) there are lots of body shapes that (01:25:48) aren't anthropomorphic humanoid bodies. (01:25:51) You know, more arms, more legs, more (01:25:52) heads, uh lots of different formats. And (01:25:54) I I do suspect we'll we'll see to to (01:25:57) See, I'm not sure if this is your dream (01:25:59) or your nightmare, but we will see lots (01:26:01) of different Cambrian explosion, lots of (01:26:03) different body shapes tested. (01:26:05) >> All right, listeners call dream or (01:26:08) nightmare. It's just the most effective (01:26:10) use case for trying to get something (01:26:12) done. (01:26:14) >> Call call out to our listeners. I made (01:26:16) that on Nano Banana. Somebody make, now (01:26:18) that we know about the woolly mouse, (01:26:20) make Salem's perfect robot for turning (01:26:22) the woolly mouse hair into sweaters for (01:26:24) us and then send it to us. We'll put it (01:26:25) on the next pod. Okay, that's a hell of (01:26:28) a prompt. All right. Uh, another form of (01:26:31) robots are drones. And I just found this (01:26:33) anti-gravity drone. That's the the the (01:26:37) name of this drone. It's manufactured by (01:26:39) a company called Insta 360 in Shenzen. (01:26:42) For those you who don't know, Shenzen is (01:26:43) really sort of the entrepreneurial (01:26:45) hotbed in China. U I've visited many (01:26:48) times. You can go there and every part (01:26:51) and component you need uh is there uh to (01:26:54) be manufactured. So check out this check (01:26:57) out this video uh of an 8K 360 degree (01:27:01) drone uh talk about marketing genius. (01:27:07) So, (01:27:09) this drone user is using it with VR (01:27:14) goggles and he's on a platform suspended (01:27:18) by a balloon at 5,000 ft altitude and (01:27:23) the drone is just flying a beautiful uh (01:27:27) you know 360 view of him. (01:27:30) >> The dude standing on a platform (01:27:32) suspended by a hotter balloon. That's (01:27:33) way more interesting than the drone. (01:27:36) That's ridiculous. (01:27:37) >> Well, it's it's like what are you going (01:27:39) to do to capture someone's uh eyeballs, (01:27:41) their attention, right? (01:27:42) >> You know, I think Sem is on to to (01:27:44) something here. Drones are a commodity, (01:27:46) but the the experience of being on a hot (01:27:49) air balloon at altitude in a VR headset (01:27:52) controlling a 3D drone, that that's got (01:27:54) to be some sort of consumer experience (01:27:56) that one could build an enormous (01:27:57) business out of. Maybe that's more (01:27:58) interesting than the drone itself. (01:28:00) >> Yeah. (01:28:03) All right. (01:28:04) Well, all right. Let's move on to our (01:28:06) next uh story in the robot. (01:28:08) >> You have the VR headset. Why do you need (01:28:10) to be suspended up at 5,000 ft? That (01:28:12) makes no sense. (01:28:13) >> Well, for latency, right? (01:28:16) >> You want to see yourself suspended on (01:28:18) the balloon at altitude. It's more (01:28:20) exciting or something. (01:28:21) >> All right, let's go to our next robot (01:28:22) story. Uh, and this is robotically uh (01:28:26) automated vertical farms, which is an (01:28:28) important part of our future food chain. (01:28:32) So, of course, out of China once again, (01:28:36) and uh what we're going to see here are (01:28:39) these massive vertical farms uh that are (01:28:41) operating 24/7. (01:28:44) um (01:28:46) basically growing at the perfect uh you (01:28:49) know light frequency at the perfect soil (01:28:52) and and uh drip irrigation pH and it's (01:28:56) being you know the AI is checking to see (01:28:59) if it's ripe if it's ready for (01:29:01) harvesting and the robot arms are (01:29:02) harvesting and this is going basically (01:29:05) 24/7 (01:29:06) uh in a city near you. I mean this is (01:29:09) one of the futures you know stem cell (01:29:11) grown meats and vertical farming that (01:29:13) helps us bring food to the individuals. (01:29:16) I don't know if you realize this guys (01:29:18) but like half the cost of a meal that (01:29:21) you have is food miles transporting the (01:29:23) food uh from you know sort of (01:29:26) Argentinian beef or Chilean red wine or (01:29:30) >> the average the average meal in the US (01:29:32) travels 2400 miles to get to your table. (01:29:35) >> Yeah. Um this is something really this (01:29:37) is something kind of incredible. We've (01:29:39) been tracking this for a while. Um you (01:29:41) know we've crossed over into um economic (01:29:44) efficiency for uh farming and (01:29:48) agriculture and food production. This (01:29:50) calculation I've seen that's the most (01:29:52) startling is if you took 35 skyscrapers (01:29:54) in Manhattan turn them into vertical (01:29:56) farms that would feed the entire city (01:29:58) sustainably. So you think about the food (01:30:00) security u logistics trucking all of (01:30:04) that stuff and when you can automate the (01:30:06) entire farm the yield is something like (01:30:08) 7 to n times what you can get with (01:30:10) horizontal farming because you can give (01:30:12) exactly the right frequency of light uh (01:30:14) that you can d by the way uh you save (01:30:17) 99% of fresh water and 70% of our our (01:30:20) fresh water goes to agriculture so you (01:30:22) don't need a lot (01:30:23) >> and no pesticides no fertilizer all of (01:30:25) this stuff the benefits are kind of (01:30:27) incredible so we're going to see (01:30:28) vertical farms next to every restaurant (01:30:31) uh over time just feeding the (01:30:34) restaurant. This is amazing stuff. (01:30:36) >> Yeah, (01:30:37) >> it's probably also just quickly worth (01:30:39) pointing out that video to to my (01:30:41) knowledge was actually put out by the (01:30:43) Chinese government and this is a a new (01:30:45) form of soft power, soft influence (01:30:48) broadcasting these these visions (01:30:50) presumably ground truth accurate but (01:30:53) presumably of radical forms of (01:30:55) automation. I think we're going to see (01:30:57) many forms of propaganda, soft influence (01:31:00) as showing these amazing tech (01:31:03) demonstrations of robotics in action (01:31:05) start to hit the internet. (01:31:07) >> And by the way, a humanoid robot makes (01:31:09) no sense in that factory. Just (01:31:12) >> agreed. But a humanoid robot does make (01:31:14) sense in this next story again out of (01:31:16) China. Uh China is testing retail (01:31:19) automation with humanoid robots running (01:31:21) the shops. Right. So what do we have (01:31:23) here? you know, you're walking by, you (01:31:25) look inside, you don't see humans, you (01:31:27) see a robot behind the table, behind the (01:31:29) desk, and you know, I want to go in and (01:31:31) check it out. So, um, this is the rise (01:31:34) of the robotrun convenience store, uh, (01:31:37) taking humans out of the loop. Uh we've (01:31:40) seen Amazon do a version of this, right, (01:31:42) with their Amazon Go where you walk into (01:31:45) the shop and you just pick up anything (01:31:48) off the shelf and there's cameras, you (01:31:50) know, noticing what you took and (01:31:52) noticing what you put back on the shelf (01:31:54) and then you're automatically rung up as (01:31:56) you walk out. Uh but here we've got a (01:32:00) twoarmed, two-legged humanoid robot (01:32:02) doing the the store clerking. Um, I I do (01:32:06) think that this is going to be viewed as (01:32:08) sort of like the atomic vacuum cleaner (01:32:09) moment of 2025. Like, do do you really (01:32:12) need a humanoid robot in a convenience (01:32:14) store? No. Probably there's more (01:32:15) ergonomic solution like as you say, (01:32:17) Peter, Amazon's just walk out technology (01:32:19) on the one hand. On the other hand, I (01:32:21) would love to to live in a world where (01:32:24) every convenience store is filled with (01:32:25) humanoid robots in the US doing this as (01:32:27) well. (01:32:28) >> I I think it's fun. I mean, I'm sure (01:32:30) we'll see this I'm sure we'll see this (01:32:31) this year as soon as uh as soon as 1x (01:32:34) with their Neo Gamma or Figure. And (01:32:36) we'll be visiting Figure at the end of (01:32:39) January to record our next podcast with (01:32:41) Brett Adcock. I just spoke to him (01:32:43) yesterday. (01:32:44) >> Uh super excited about going and seeing (01:32:46) behind the scenes there. (01:32:47) >> Two two counter predictions. One is I (01:32:49) think this takes at least 5 years to (01:32:51) have a convenience store operator with a (01:32:53) humanoid robot. And by the time that (01:32:55) five years arrives that we won't need (01:32:56) convenience stores anymore for various (01:32:58) other reasons. (01:32:59) >> Ah, interesting. Everything is being (01:33:00) conveniently (01:33:02) taken to you by a drone. (01:33:03) >> Drone delivered. (01:33:04) >> Yeah. (01:33:05) >> You know, with Brett Edcock, maybe he'll (01:33:07) let us go behind the scenes for real, (01:33:08) like into the factory because with 1X, (01:33:10) you know, there's too much proprietary (01:33:11) stuff. They wouldn't let us do it. But (01:33:14) if they cleaned up a little bit, maybe (01:33:15) we could have done it. But it's (01:33:16) incredible when you go back and see the (01:33:18) the actual robot construction. It's h (01:33:21) god if we can get footage. (01:33:22) >> We went back we went back and saw it but (01:33:24) we couldn't bring the cameras back there (01:33:25) is what you were saying. (01:33:26) >> Yeah. Yeah. Too many secrets. (01:33:28) >> Another story here back in the US. (01:33:30) Boston Dynamics announces its plan to (01:33:32) ship automotive volumes of humanoids. Uh (01:33:35) and this is uh from their lead uh their (01:33:38) product. I actually interviewed the CEO (01:33:41) uh at FII. So we're owned by Honda (01:33:45) Hyundai for a reason. We can ship (01:33:47) automotive volumes of humanoids. So (01:33:50) there's a billion cars uh right now out (01:33:53) there and these are being manufactured (01:33:56) at you know tens of millions. Uh imagine (01:33:59) well we've talked about this Elon plans (01:34:01) to do this Brett Adcock plans to do (01:34:03) this. We've heard this from Brent Borick (01:34:06) uh now we're hearing this from Atlas (01:34:07) right the ability to manufacture uh at (01:34:10) the millions and tens of millions robots (01:34:13) building robots. (01:34:14) >> We don't need billions of cars. We do (01:34:16) need billions of humanoids. Yeah. Two (01:34:19) armed humanoids, Sem. Two armed (01:34:21) humanoids. (01:34:22) >> Okay. Well, (01:34:23) >> don't be arrested. (01:34:24) >> I'm staying silent on this one. (01:34:27) >> Uh uh here's a story that's fun. Um (01:34:30) years ago, uh I had the pleasure of (01:34:32) meeting an extraordinary entrepreneur, (01:34:34) Eric Mijigovski, who built the Pebble (01:34:37) Watch. And uh he did this on uh on a (01:34:41) crowdfunding platform. Remind me which (01:34:44) one it was. Um it was Kickstarter. Yeah. (01:34:46) He built he was running out of money. (01:34:49) >> Yeah. He was running out of money and he (01:34:51) had like 3 months of cash in the bank. (01:34:53) He was able to get funding for his (01:34:55) Pebble watch. (01:34:56) >> And so he goes on Kickstarter and he (01:34:58) says, "Hey, if you want one of these (01:35:00) watches, uh, fund me." And he went from (01:35:04) uh from one problem of not having enough (01:35:06) money to another problem. I forget how (01:35:09) many orders he had. I (01:35:10) >> I'll I'll So Eric's a fellow Waterlue (01:35:13) grad. Um and he uh was running out of (01:35:17) money as you say even coming through Y (01:35:19) cominator no investor in Silicon Valley (01:35:21) he talked about 20 plus and now nobody (01:35:23) would fund it because hardware was kind (01:35:25) of a bad word back then so he puts it up (01:35:27) on Kickstarter trying to raise a hundred (01:35:29) grand to build a prototype of his watch (01:35:31) gets $10 million worth of orders. (01:35:33) >> That's right. (01:35:33) >> U and it's an important point because it (01:35:36) tells you two or three things. One, the (01:35:38) investor is wrong. Fine. Secondly, if (01:35:40) you can do this, why do you need the (01:35:41) investor at all? But the third thing (01:35:43) that I think is the most powerful and (01:35:45) one of the big inflection points, we (01:35:47) talk a lot about this in exponential (01:35:49) organizations is that now that you can (01:35:51) do this type of Kickstarter type thing, (01:35:53) you can actually get market validation (01:35:55) for a product without build before you (01:35:57) build a product. (01:35:58) >> And we've never have been able to do (01:35:59) that before in consumer uh hardware or (01:36:02) consumer products. So this is an amazing (01:36:04) inflection point. Sony is actually (01:36:06) launching anonymous Kickstarter (01:36:08) campaigns and then funding the winners (01:36:10) because it's their product development (01:36:11) has not been the greatest over the last (01:36:13) couple of decades. So they're kind of (01:36:15) tapping into this modality which is (01:36:17) really powerful. So Eric goes from (01:36:19) having one problem of not having money (01:36:20) to another problem which he's got to (01:36:22) deliver now on $10 million worth of (01:36:24) orders. So he literally takes the first (01:36:26) plane out of the US to Shenzen and and (01:36:30) basically builds the manufacturing chain (01:36:32) in China uh to deliver this. Uh and it (01:36:35) was a great watch. Remember having I (01:36:37) gave it out at Abundance 360 years ago (01:36:40) when it a decade ago, but then Apple (01:36:43) Watch came out and sort of crushed the (01:36:45) marketplace. Well, uh Eric's come back (01:36:48) and he's got something called (01:36:49) >> pivoting to AI. (01:36:50) >> Yeah. The Pebble uh smart ring. And for (01:36:54) 75 bucks, you wear a ring that's got one (01:36:56) purpose. It's got a small little (01:36:58) physical button on it. And when you (01:37:00) press the button, a microphone records (01:37:04) whatever you want. So this is, you know, (01:37:06) you remember like waking up in the (01:37:07) middle of the night like remembering (01:37:08) something. You just push your ring and (01:37:10) you whisper into your ring. Or you're (01:37:12) meeting with somebody, you walk away (01:37:13) from your meeting and say, "Okay, I need (01:37:15) to call, you know, XYZ as soon as this (01:37:18) is over." And it's sort of uh, you know, (01:37:21) reminders. uh and it's notes that go (01:37:23) into your AI model. It has one purpose, (01:37:26) right? This is is not, you know, (01:37:27) tracking your heart rate or your sleep. (01:37:29) It's tracking uh sort of uh bits that (01:37:32) dribble out of your out of your thought (01:37:34) during the course of a day. (01:37:36) >> I I love and critically like the where (01:37:39) does the voice go? The voice goes from (01:37:41) the ring to an ondevice on your phone (01:37:45) hosted large language model that then (01:37:47) transcribes and analyzes. So what is (01:37:49) this really doing? This is really to to (01:37:51) the extent that a a ring stays on you (01:37:53) almost all the time. This is about (01:37:55) adding a button to the human body that (01:37:57) enables you to speak to a large to a (01:38:00) foundation model that's also on your (01:38:02) body. And so question to uh to to the (01:38:06) moonshot mates here. How long until it's (01:38:09) not just a button on your body that (01:38:11) enables you to talk to a foundation (01:38:12) model, but you're you're swallowing (01:38:15) foundation models? How long to the first (01:38:17) edible foundation model? Well, (01:38:20) injectable (01:38:21) or sub subdermal. (01:38:23) >> You think it'll be injectable versus (01:38:24) edible first? (01:38:26) >> Uh, well, yeah. I mean, if you're if (01:38:27) it's edible, it's going to pass through (01:38:28) your elementary canal all the way out to (01:38:30) the other end. (01:38:31) >> So, I I want this, you know, there's (01:38:33) interesting. There's part of the skull, (01:38:35) right, the mastoid bone in the back (01:38:36) behind your ear. That's this hollow area (01:38:39) of uh of of of (01:38:42) bone. I think it's a great place to (01:38:44) implant a a permanent um uh you know (01:38:48) microphone and speaker. Uh yeah, that's (01:38:51) my prediction. We're gonna be implanting (01:38:53) a microphone speaker at the back of your (01:38:54) head. (01:38:55) >> That was directly on Shark that exact (01:38:57) thing was on Shark Tank and Mark Cuban (01:38:59) vomited. (01:39:00) >> Really? (01:39:03) >> You can iterate hardware much faster (01:39:04) outside the body than inside the body. I (01:39:07) don't think it'll be invasive for a (01:39:08) while. Yeah, I think we'll see (01:39:10) swallowable swallowable foundation (01:39:12) models in the next two years. (01:39:14) >> Bluetooth like just Bluetooth in and out (01:39:15) of your uh body to your phone. (01:39:18) >> Bluetooth but critically locally hosted. (01:39:20) Very locally hosted. (01:39:21) >> Okay. (01:39:23) >> All right. A few subjects, a few a few (01:39:25) topics on space here. Let's move us (01:39:27) along, guys. Chile becomes the first uh (01:39:30) Latin America country to enable Starlink (01:39:32) direct to sell. Uh so I mean listen, (01:39:36) Starlink is such the killer app uh for (01:39:40) for SpaceX and the ability for him to (01:39:43) potentially bypass the current phone (01:39:45) industry which I mean tens and hundreds (01:39:48) of billions of dollars has been put down (01:39:51) in terms of uh of uh you know G4 and G5 (01:39:56) level distribution networks now to be (01:39:58) bypassed by Starlink. Crazy. Um, but (01:40:02) this is what I find this next story. (01:40:04) Take a listen. I mean, can you (01:40:06) >> can I just go back to that? Can I just (01:40:08) go back to that just for a sec, Peter? I (01:40:10) think this is something a very big deal (01:40:12) because, you know, throughout history, (01:40:14) this is the failure of government. The (01:40:16) UN should have launched something like (01:40:18) Starlink. You know, they should be (01:40:21) launch. (01:40:22) But they're fundamentally unable to and (01:40:24) it needs private sector to do this type (01:40:27) of stuff. What I find incredible is the (01:40:29) demonetization and the dematerialization (01:40:32) of technology allows now a private (01:40:34) individual to do something like this (01:40:36) that changes the world completely uh in (01:40:39) a such a powerful way and you kind of (01:40:41) can say well governments just step out (01:40:43) of the way and let private sector do (01:40:45) everything going forward right because (01:40:46) it'll navigate most of this with light (01:40:49) regulation uh we can navigate most of (01:40:51) this stuff now so I'm really really (01:40:53) excited by this (01:40:54) >> okay can I ask you guys a question (01:40:56) because I was trying to look at the data (01:40:57) behind this. You know, the idea of (01:41:00) orbital data centers wasn't in the (01:41:03) conversation how long ago. I mean, we (01:41:06) weren't talking about this a year ago. (01:41:07) We weren't talking about it 9 months (01:41:09) ago. (01:41:10) >> It's the last guy the guy at Abundance (01:41:13) 360 (01:41:14) >> uh March published a paper on this about (01:41:16) 14 years ago and if you were reading (01:41:19) Incelerondo in which case you had the (01:41:21) blueprint for everything we're seeing (01:41:23) now. (01:41:23) >> Sure. But it wasn't. (01:41:25) >> But no, but no. March a year ago, one of (01:41:26) your guy, one of your abundance 360 guys (01:41:28) was talking about it and he was going to (01:41:30) do Bitcoin mining in space at that point (01:41:31) in time and everybody thought he was (01:41:33) insane. And we also thought we couldn't (01:41:34) do the cooling. So that was only March a (01:41:37) year ago. So that's nine months. (01:41:38) >> But there's a (01:41:39) >> So I know at that point it was nothing. (01:41:41) >> Yeah. But the last 6 months, really the (01:41:43) last four months, all of a sudden, every (01:41:46) single player, we've got companies out (01:41:48) of China. we saw at the last pod. We (01:41:50) have now a company out of Europe and we (01:41:52) have a dozen companies in the US. And (01:41:54) then I found this video clip which I (01:41:57) found fascinating because Google was not (01:41:59) discussing it a few months ago but here (01:42:01) we are. Listen to Sundar. (01:42:02) >> Yeah. (01:42:02) >> How do we one day have data centers in (01:42:04) space so that we can better harness the (01:42:07) energy from the sun. You know that is (01:42:09) 100 trillion times uh more energy than (01:42:12) what we produce in all of Earth today. (01:42:14) So we want to put these data centers in (01:42:17) space closer to the sun uh and and I (01:42:20) think we are taking our first step in (01:42:21) 27. We'll send tiny uh tiny racks of uh (01:42:25) machines uh and and have them in (01:42:28) satellites, test them out and then start (01:42:30) scaling from there. But there's no doubt (01:42:32) to me that a decade or so away we'll (01:42:34) we'll we'll be viewing it as a more (01:42:37) normal way to build data centers. (01:42:39) >> I never thought I'd hear Sundai Sundar (01:42:41) say tiny racks of machines. That's (01:42:43) hilarious to me. (01:42:44) >> I just love the school boy level (01:42:46) excitement he's got there. You can see (01:42:47) him actually grinning. He's like, "Oh, (01:42:49) data centers in space. This is amazing." (01:42:51) >> I I love the AI AI generated. The big (01:42:54) banner on top of that video is AI (01:42:57) generated. It's like we're going to (01:42:58) we're going to always tell you that this (01:43:00) scene in deep space is AI generated as (01:43:03) if as if it was not. Um the the reason (01:43:06) the reason Peter why you know I mean (01:43:08) even though I I maybe a little bit glib (01:43:10) saying well if you had read accelerando (01:43:12) this would have been obvious to you (01:43:13) almost 30 years ago on the one hand the (01:43:15) reason you know that this is a sudden (01:43:17) phase change in in the way the industry (01:43:19) works is Google's plans this is public (01:43:22) information the Google plan to launch (01:43:24) these so it's TPUs first of all Google's (01:43:27) launching TPU based data centers (01:43:29) obviously are on planet satellites (01:43:32) planet labs it's not Google's own (01:43:33) satellites it's planet labs. So, so you (01:43:36) know, if Google's hitching a ride via (01:43:38) SpaceX on planet satellites, this is all (01:43:42) of a sudden. I I I'll say that second (01:43:44) point. Sun-synchronous orbit is about to (01:43:46) become very very crowded. (01:43:48) Sun-synchronous orbit is is is a a low (01:43:50) Earth orbit that satellites that want to (01:43:52) always have sun exposure, never pass (01:43:54) behind the Earth, never be in the (01:43:56) shadow, always have solar power for (01:43:58) their panels. It's going to be very (01:43:59) crowded. (01:44:00) >> It's a real estate. It's a limitation. (01:44:02) And there, you know, there currently is (01:44:05) limits on how close you can get to other (01:44:07) satellites. Um, that's going to be a (01:44:09) real it's going to be a real challenge (01:44:10) because we've got, you know, a dozen (01:44:12) companies all wanting to do this at the (01:44:14) same time. It's going to be a race and (01:44:17) how the FAA, which governs this, is (01:44:20) going to decide who gets the territory, (01:44:22) who doesn't. In geostationary orbit, uh (01:44:25) there's a very clear demarcation of I (01:44:29) own these orbital slots over my country, (01:44:32) but low Earth orbit doesn't have that (01:44:34) situation. (01:44:35) >> Peter, you're making the the case for (01:44:37) the Dyson swarm. Again, the Dyson swarm. (01:44:40) So, we move out of geo, we move out of (01:44:42) LEO, and Sundar himself in in this clip (01:44:45) was saying, we want to get closer to the (01:44:47) sun. So, we're we're sleepwalking (01:44:50) straight into the Dyson swarm. Well, (01:44:52) Peter, to your prior point too, this was (01:44:54) science fiction a year ago and now (01:44:55) suddenly it's mainstream among the top (01:44:57) CEOs in the country. How does that (01:45:00) happen? But, you know, you look at Elon (01:45:01) and his credibility. You look at, you (01:45:03) know, Alex, your credibility. A lot of (01:45:05) things that were impossible a year ago (01:45:07) are going to be very easy a year from (01:45:09) today. And if your track record of (01:45:11) predicting them is is near perfect, (01:45:14) then, you know, the credibility of these (01:45:16) crazy sounding ideas immediately catches (01:45:19) on. And you're going to see a lot more (01:45:20) of that I think because the the you know (01:45:22) the capabilities are are exponentially (01:45:25) growing but you know some of these (01:45:26) things are truly hairbrained and some of (01:45:28) them actually are (01:45:30) >> is there line of sight on solving the (01:45:32) heat dissipation problem for these (01:45:33) satellite data center? (01:45:34) >> Yeah and for radiate in the direction of (01:45:37) the cosmic microwave background. So (01:45:39) >> yeah, the final answer shocked me, but (01:45:42) for every square meter of solar panel, (01:45:44) it only takes one same square meter of (01:45:46) radiant cooling, radiative cooling, (01:45:48) which really surprised me. I thought it (01:45:50) would be we we estimated on Gemini, (01:45:52) which was wrong. Uh at 10x, uh you need (01:45:55) a 10x more, you know, area. And it was (01:45:58) just wrong. It's it's cooling at 1x and (01:46:01) I don't know how they and it's all (01:46:02) aluminum based, so it's not weird weird (01:46:04) expensive metals or anything like that. (01:46:06) So yeah, point it into deep space like (01:46:08) Alex has been saying forever and it it's (01:46:11) for whatever reason just flat out (01:46:12) working. (01:46:13) >> So most of (01:46:15) >> I took all of the comments from our last (01:46:17) two pods and ran them through one of the (01:46:20) LLMs and said, "Okay, pull out the the (01:46:22) most interesting AMA questions. Here we (01:46:25) see a list of 10 of them, gentlemen. Um (01:46:28) uh let's pick out a few to answer. I'll (01:46:31) start with one which is how do you make (01:46:34) these space-based AI data centers fault (01:46:36) tolerant right there's sunspots there is (01:46:39) the potential for you know disruption (01:46:42) from a uh even from an EMP at some point (01:46:45) uh god forbid uh any ideas on making (01:46:48) them fault tolerant (01:46:51) >> those are two very different faults (01:46:53) >> yeah yeah both there are lots (01:46:55) >> disruptive (01:46:56) >> there are lots of different failure (01:46:57) modes so I I do think this is another (01:47:00) multi-billion dollar company that (01:47:02) someone should start. There are many (01:47:04) techniques right now ranging from uh (01:47:07) switching from silicon based electronics (01:47:09) to to maybe uh other semiconductors. (01:47:12) Yeah. like gallium arsenide, uh, 26 or (01:47:14) or 37 semiconductors that are more fault (01:47:17) tolerant, have different band gaps to (01:47:20) designing just electronics that are (01:47:22) intrinsically at at the at the design (01:47:24) level better able to tolerate faults to (01:47:29) uh just doing what what right now is a (01:47:31) standard protocol, which is if if (01:47:33) there's uh if there's a solar storm or (01:47:35) bad space weather, you shut down or you (01:47:37) switch them to to safety mode. So that (01:47:39) there are lots of partial solutions (01:47:42) here. To my knowledge, there isn't like (01:47:44) the definitive industry standard (01:47:46) solution of what happens if you're in (01:47:48) the middle of a training run. (01:47:49) >> I just hate to think about the idea of (01:47:51) your all the data centers in orbit (01:47:53) shutting down because there's a solar (01:47:54) storm for the next 12 hours. We're (01:47:57) getting hit by uh by alpha particles. (01:47:59) >> But how do we solve that in general? (01:48:01) Like if there's bad weather or a (01:48:03) blackout on Earth, you have (01:48:04) diversification. So, so if anything (01:48:06) again like let's put space-based AI data (01:48:09) centers throughout the solar system. So (01:48:11) if there's bad space weather in one (01:48:13) part, there isn't in another. (01:48:15) >> That's a great point actually. I bet (01:48:16) earthquakes and tsunamis and hurricanes (01:48:18) are much bigger problem than solar (01:48:21) storms. (01:48:22) >> All right, let's pick another one of (01:48:24) these. (01:48:24) >> Hey, just just to make a point though, (01:48:26) there's a there's a kind of a flaw in (01:48:27) the question, too, because when you have (01:48:29) Skylab up there, you want it to be up (01:48:30) there for 20 30 years and you don't want (01:48:32) it to get hit and destroyed or anything. (01:48:34) But this space-based data centers need (01:48:36) to be replaced every three years with (01:48:37) new chips. (01:48:38) >> And so they're not it's a constant (01:48:40) launch, recycle, launch, recycle, (01:48:41) launch, recycle thing. (01:48:43) >> Somebody EMPs the entire thing and (01:48:45) destroys it, then there's a war, of (01:48:47) course. But it was going to get replaced (01:48:48) in a three-ear cycle. Anyway, it's not (01:48:50) it's not like Skyab. (01:48:53) >> Interesting. One of the things we did in (01:48:55) the uh uh for planetary resources when (01:48:58) we're looking at asteroid mining, we we (01:49:00) set up the the software so we would (01:49:03) expect constant disruption. Um and the (01:49:07) system we focused on rapid restart of (01:49:09) the system so it would boot up (01:49:11) extraordinarily fast. Um all right. (01:49:15) >> Can I tell a quick story here? (01:49:16) >> You can, but I want you to choose one of (01:49:18) these uh one of these AMA questions (01:49:20) also. (01:49:21) >> Sure. Um, you and I were sitting in a (01:49:23) hotel in Dubai and Richard Branson walks (01:49:26) by and he said, "Hello." And we grabbed (01:49:28) a quick drink and he said, "Peter, how's (01:49:31) my investment in, you know, planetary (01:49:33) resources going and you described that (01:49:37) how it was going? It had NASA contracts, (01:49:39) etc." And Richard turns to me and goes, (01:49:40) "This is why Peter's interesting because (01:49:42) in a random hotel lobby, I'm suddenly (01:49:44) having a conversation about asteroid (01:49:47) mining off planet just like this. This (01:49:50) conversation happens nowhere else in the (01:49:52) world except with Peter. We love you so (01:49:54) much. (01:49:54) >> It was fun. All right, Sel, pick a (01:49:56) question here. Is this question bingo? (01:49:59) >> Should we expect G20 level initiatives (01:50:02) for UBI within the decade? I would hope (01:50:04) it would be within a year. Uh it needs (01:50:07) to happen very very fast. I think it'll (01:50:09) force the conversation. But um uh (01:50:13) >> universal basic right universal replaced (01:50:16) soon by UBS, universal basic services. (01:50:19) Uh but I think you shouldn't expect much (01:50:21) from the G20 period. I think that's the (01:50:23) flaw in the question. But in general, (01:50:25) we're going to expect to see this uh (01:50:27) rolling out in a pretty rapid way. Lots (01:50:29) and lots of experiments being done all (01:50:31) over the world on this because they have (01:50:32) to do we have to move to something like (01:50:34) that. The social contract is completely (01:50:36) being wiped out in the current model. (01:50:39) >> Dave, why don't you pick a question (01:50:40) next? (01:50:42) >> Uh okay, I'll take number one. How can (01:50:44) AI lift up those who aren't (01:50:45) international entrepreneurs? I I I think (01:50:48) one listen to the podcast, get (01:50:50) subscriptions, play with the tools, and (01:50:52) then brand yourself as an AI expert (01:50:54) within your company, you know, or if (01:50:56) you're not going to be an entrepreneur, (01:50:57) that's fine. You know, the demand for (01:50:59) this knowledge inside regular corporate (01:51:01) world is going to go through the roof in (01:51:03) 2026. And if everybody around you knows (01:51:06) you're the AI person, and also don't be (01:51:09) intimidated. The the historically, if (01:51:12) you wanted to be a software god, you (01:51:13) needed to be very, very softwary. That's (01:51:16) not true with AI. It's it's much more (01:51:18) intuition based. You can build virtually (01:51:20) anything with voice prompts. Uh it's (01:51:22) just knowing how it applies in your (01:51:24) industry will separate you. So just jump (01:51:26) in the game. (01:51:28) >> Yep. Amazing. Uh (01:51:31) Alex, do you have one? (01:51:33) >> I I'll take question number four for 10 (01:51:35) trillion. (01:51:38) >> Uh is pure scaling enough or what comes (01:51:40) after? Uh so so I think the answer I (01:51:44) think it's a trick question. I I think (01:51:46) pure scaling probably is by pure scaling (01:51:49) I I I I'll construe the question to mean (01:51:52) we freeze all algorithms. No new (01:51:54) algorithms are allowed to be developed (01:51:56) in AI but we're allowed to shovel more (01:51:58) and more compute especially inference (01:52:00) time compute into the existing (01:52:01) algorithms. I I do strongly suspect that (01:52:04) if we froze all the algorithms we have (01:52:06) today, no new architectures, but we get (01:52:09) lots more compute coming online. The (01:52:12) existing architectures combined with (01:52:13) scaled compute will be enough to give us (01:52:16) AI smart enough to tell us what a (01:52:19) perfect algorithm would be to the point (01:52:22) where we get our uh highly coveted AI (01:52:25) researcher recursive self-improvement (01:52:28) the final algorithm and we can just ask (01:52:30) our scaled algorithms what comes after. (01:52:33) So in in in summary my answer to (01:52:35) question number four is yes. I I think (01:52:37) probably pure scaling is sufficient. Is (01:52:40) it is it all that we need? No. Of (01:52:43) course, algorithm in the real world, (01:52:44) algorithmic development is continuing (01:52:46) and we're going to get both. But could (01:52:48) we live with pure scaling at this point? (01:52:50) My guess is probably yes. All right, (01:52:52) let's answer one more here. Number (01:52:54) three, how do the Moonshot Mates prepare (01:52:56) daytoday for each podcast episode? Uh (01:53:00) yeah, I think we we can share that. So, (01:53:02) uh uh let's see. Alex, you're constantly (01:53:05) providing uh the team with a incredible (01:53:09) list of all the breakthrough stories (01:53:11) you're searching. You're probably (01:53:12) generating how many how many AI stories (01:53:14) per day do you think you generate for us (01:53:16) to look at? (01:53:17) >> Oh gosh. Um or order of magnitude 20 (01:53:20) important stories per day. I'm also at (01:53:22) this point like I spend so much time (01:53:25) just reading reading uh primary sources, (01:53:28) archive papers, etc. living in the (01:53:31) zeitgeist of of the moment because after (01:53:33) all drink singularity comes around only (01:53:36) approximately one time per planet. So so (01:53:39) it's a special time. Uh I I I do also at (01:53:43) this point um you know probably should (01:53:45) say uh I'm I'm also turning all of these (01:53:48) stories in addition obviously to (01:53:50) research for this show into quasi daily (01:53:53) newsletter. Uh just trying to (01:53:57) >> follow Alex on X. Uh he puts out some (01:53:59) incredible uh daily uh uh sort of (01:54:03) interesting AI rants I would say or AI. (01:54:06) >> Follow me on X follow me on LinkedIn. (01:54:08) It's it's a genre I'm trying to (01:54:10) popularize. I'm calling it sigh nonfi. (01:54:12) It's it's written in the style inspired (01:54:14) by Charlie Strawk Salando others written (01:54:17) in the style of science fiction except (01:54:19) it's all grounded in what's actually (01:54:20) happening. So, Alex generates uh you (01:54:22) know on the order of (01:54:25) 150 stories a week. I'll generate (01:54:27) probably 20 or 30 stories a week. We get (01:54:29) some from Salem, some from Dave. All (01:54:32) this gets sort of put into different (01:54:34) categories. We then sort of cut it down (01:54:38) to the top uh 30 stories. I typically (01:54:41) spend uh about 10 hours sort of playing (01:54:46) slide shuffle working with Jan Luca and (01:54:49) and Dana. uh who are incredible members (01:54:52) of our team and then we do research on (01:54:55) those stories uh to get the details and (01:54:57) and think about them and uh I'm probably (01:54:59) spending a good 15 hours of my week (01:55:03) focused on this. How about you Dave and (01:55:05) Seline? (01:55:06) >> Well, everything you just said, you (01:55:08) know, I lean entirely on Alex's internal (01:55:10) feed, which now you can get on X. You (01:55:12) know, it's a digest of the same thing (01:55:13) that's brand new as of the last week or (01:55:15) so, so take advantage of it. Um, but (01:55:18) I've been reading that internally for (01:55:20) what, a year now, I guess, or more. Uh, (01:55:22) which is very timeconuming, but I need (01:55:24) to know it all. The only other thing I (01:55:26) do is I route all the really big stuff (01:55:28) over to the venture capital team and (01:55:30) say, what are the business implications (01:55:32) of this, which we need to know anyway to (01:55:34) run our venture fund? And then I try and (01:55:36) bring those stories back into the (01:55:37) moonshots feed so that we can talk about (01:55:39) not just the technology, but what it (01:55:42) means to investors, to business people, (01:55:44) to people with career planning and all (01:55:46) that. (01:55:49) um I spend um I source a few stories but (01:55:52) nowhere near as much as the rest of you (01:55:54) but I think the I spend a chunk of time (01:55:57) the minute you guys release the the deck (01:55:59) I look through it and then find it's (01:56:00) changed again and so I have to restart (01:56:02) again uh so I'm always playing catchup (01:56:05) with the slides that you and then Peter (01:56:08) on the last night you go God knows what (01:56:09) you do but you change it all again and I (01:56:11) have to re I do re research it um I (01:56:14) spend half a few hours a week looking up (01:56:16) the term terms in the papers that Alex (01:56:18) surfaces because half of it is Greek. (01:56:20) Um, and then I'll ask also my community (01:56:23) me, my opening exo community. So there (01:56:25) there's a hive mind reaction to some of (01:56:27) this which I think is very powerful (01:56:29) similar to Dave asking his team. (01:56:32) >> Uh, just again to let our subscribers (01:56:34) know (01:56:35) >> though it's just sucking up more and (01:56:36) more time per week, but it's such a (01:56:39) important thing. (01:56:40) >> It's the most thing we do. Come on. (01:56:43) Super fun. But what no one ever warned (01:56:45) you of Seem is like the the singularity (01:56:48) of covering the singularity. It's a (01:56:50) singularity of time suck. (01:56:51) >> It's it's just it's a black hole. It's a (01:56:54) black hole. Dyson swarm forming around (01:56:56) my own head. (01:56:57) >> Singularity wants your attention. (01:56:58) >> So we we hope for all our subscribers (01:57:01) and listeners that you guys appreciate. (01:57:03) We put a huge amount of work because we (01:57:05) care about this deeply. Uh (01:57:07) >> I need to give a quick plug quick plug. (01:57:10) Um I'm doing my meaning of life session (01:57:11) next week. We've already we're almost (01:57:13) sold out. Uh it's going to be pretty (01:57:15) amazing. It's going to go for several (01:57:16) hours starting 11:00 Wednesday. Come (01:57:19) armed with any question you have about (01:57:21) life and judge me by how well this (01:57:23) framework answers that question. Boom. (01:57:25) >> All right, let's get to our outro music (01:57:27) here uh from David Drinkall. I think (01:57:29) it's the perfect name for a drinking (01:57:32) game. (01:57:33) >> That can't be real. Oh my god, it's a (01:57:34) bingo card every (01:57:36) >> And so this is a bingo card. Uh, and you (01:57:39) can see tile the earth. Uh, uh, (01:57:42) >> have our glasses of water ready. (01:57:44) >> Yeah, I do. Cybernetics. Okay, let's (01:57:47) listen. (01:57:48) >> Where's the humanoid robot entry? (01:57:49) >> Uh, six arm humanoid robots. Robots down (01:57:52) at the bottom and cloud computing on the (01:57:54) bottom left. Okay, let's take a listen (01:57:56) to uh to David's uh outro music. Thank (01:58:00) you, David, for producing this for us. (01:58:01) And again, if you're listening and you (01:58:04) are creating music videos and you want (01:58:07) to create an outro song for us, send it (01:58:09) over. We'd love to we'd love to listen (01:58:12) to it and perhaps select it. All right, (01:58:14) let's take a listen. (01:58:26) Take a sip when Peter says (01:58:30) go try and gentlemen (01:58:32) two if he name drops just got back from (01:58:36) again (01:58:37) drink when Alex says better benchmarks (01:58:40) abandoning bench and finishing glass if (01:58:44) he whispers dyson swarm at last (01:58:47) moonshine lingo Sorry. (01:58:53) The earth with comput (01:59:05) bag chug (01:59:09) sip when someone says we'll cure every (01:59:13) disease. (01:59:15) When they mention startups or (01:59:17) singularity (01:59:20) drinking drops insert my usual objection (01:59:25) and the phrase red (01:59:28) successively (01:59:30) moonshot bleed (01:59:38) training (01:59:42) go Shopping (01:59:49) quick. (02:00:13) One sip for every code red. Two for (02:00:16) humanity's last exam. Three when Alex (02:00:19) says solving math. Yes, that old plan. (02:00:23) Big up when anyone says universal basic (02:00:27) services. (02:00:29) Pass out when Peter yells. That's a (02:00:31) moonshot. Ladies and gentlemen, drop. (02:00:37) >> All right. (02:00:39) >> Amazing. (02:00:39) >> That is awesome. (02:00:40) >> Yeah, it's a moonshot, ladies and (02:00:42) gentlemen. (02:00:43) >> You know, this is again a tribute to the (02:00:45) creative nature of all of our (02:00:46) subscribers. Thank you guys. And also (02:00:48) the tools out there to allow you to do (02:00:50) things like this, (02:00:52) >> guys. (02:00:54) >> Amazing. Have an amazing weekend. (02:00:56) >> Yeah. (02:00:57) >> Super creative. Take care, folks. Every (02:00:59) week, my team and I study the top 10 (02:01:01) technology meta trends that will (02:01:02) transform industries over the decade (02:01:04) ahead. I cover trends ranging from (02:01:06) humanoid robotics, AGI, and quantum (02:01:08) computing to transport, energy, (02:01:10) longevity, and more. There's no fluff, (02:01:12) only the most important stuff that (02:01:14) matters, that impacts our lives, our (02:01:16) companies, and our careers. If you want (02:01:18) me to share these meta trends with you, (02:01:20) I write a newsletter twice a week, (02:01:22) sending it out as a short two-minute (02:01:24) read via email. And if you want to (02:01:25) discover the most important meta trends (02:01:27) 10 years before anyone else, this (02:01:29) report's for you. Readers include (02:01:31) founders and CEOs from the world's most (02:01:33) disruptive companies and entrepreneurs (02:01:35) building the world's most disruptive (02:01:37) tech. It's not for you if you don't want (02:01:39) to be informed about what's coming, why (02:01:41) it matters, and how you can benefit from (02:01:43) it. To subscribe for free, go to (02:01:45) dmmandis.com/tatrends (02:01:48) to gain access to the trends 10 years (02:01:50) before anyone else. All right, now back (02:01:52) to this episode.