The Economic Singularity Will Make Today’s Economy Unrecognizable w/ Dr. Alexander Wissner-Gross (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) As intelligence becomes too cheap to (00:00:02) meter, that's going to drive down via (00:00:04) robotics and via other channels. I think (00:00:06) the cost of the effective cost of labor. (00:00:09) And once you can drive energy and (00:00:12) intelligence [music] and labor all to (00:00:14) near zero asmtoically, the economy (00:00:17) starts to look very different from the (00:00:19) way it looks yesterday or even today. (00:00:21) What's up everybody? It's LG Ducet here (00:00:23) and welcome to the Milk Road AI podcast. (00:00:24) The AI show that loves to live the (00:00:26) future every single day, but only when (00:00:28) it's not terrifying. Today is February (00:00:29) 4th, 2026. We are recording on February (00:00:32) 2nd. Listen, like it or not, the (00:00:34) economic singularity is coming. If all (00:00:37) the premonitions about AI come to pass, (00:00:38) we're in for a period of mass deflation (00:00:40) as basically everything from labor to (00:00:42) software becomes exponentially cheaper. (00:00:44) It's exciting, but it's terrifying. My (00:00:46) guest today writes and podcasts about (00:00:48) this every single day as the writer and (00:00:51) author of The Innermost Loop and co-host (00:00:53) of the Popular Moonshots podcast. He's (00:00:54) one of the smartest people we could ever (00:00:56) have on the show. And I'm serious. This (00:00:58) guy won the USA Computer Olympiad twice (00:01:01) in the late 90s and he is also the first (00:01:03) person in MIT history to earn a triple (00:01:06) major with a bachelor's degree in (00:01:08) physics, electrical science, and (00:01:09) engineering and mathematics. He also (00:01:11) graduated first in his class from MIT (00:01:13) School of Engineering. Dr. Alex Wisner (00:01:16) Gross is on the show with us today to (00:01:18) tell us where all of this is going. (00:01:20) Today's episode is brought to you by (00:01:21) Bridge sends stablecoin payments (00:01:22) instantly simple, global, friction free. (00:01:24) Dr. Alex, welcome to Milk Road. (00:01:26) >> Thank you, LG. And quick correction, I'm (00:01:28) the last triple major, not the first (00:01:30) triple major graduated from (00:01:31) >> the last. Wait, what do you mean? Stop. (00:01:34) >> Yeah, they banned it after I graduated. (00:01:36) The story that I was told was that this (00:01:38) was for mental health reasons for the (00:01:40) students. Too many students taking too (00:01:41) many classes. Turned out later it was (00:01:43) actually for financial reasons. MIT (00:01:45) wanted to cut down on the average course (00:01:47) load. (00:01:48) >> What? [laughter] (00:01:49) Okay. So, they they they you you kind of (00:01:51) took advantage of the system by learning (00:01:53) too many things too quickly and they (00:01:54) said nobody can do that again. They call (00:01:56) it a fire hose and I I figured from an (00:01:59) optionality maximization perspective, (00:02:01) why do anything else? (00:02:04) >> Oh my god, Ben. Well, that's awesome. I (00:02:06) mean, congratulations. That's that's a (00:02:07) really cool distinction. Uh, and you've (00:02:08) had you've had many more since. Listen, (00:02:10) you you're writing a daily newsletter (00:02:12) that's growing incredibly fast. You (00:02:13) write about all these themes that we try (00:02:15) and talk about on our show. There's many (00:02:17) many to cover today, so we're going to (00:02:18) go through as many uh as we can. One (00:02:20) thing I'd really like to talk about is (00:02:22) how AI is going to affect all these (00:02:24) industries that we're a part of. Right. (00:02:25) I think you've predicted somewhere (00:02:26) between like a 30 40 50x deflationary (00:02:29) effect on the economy uh on labor (00:02:32) software all that from AI. Dr. Alex, (00:02:35) please give give us a little bit more (00:02:37) detail on that and how that's going to (00:02:38) affect our lives in the next 5 to 10 (00:02:40) years. (00:02:41) >> Well, I can't take credit for the 40x (00:02:43) number. That number comes from OpenAI (00:02:45) and Sam Alman. And the the 40x number (00:02:48) specifically relates to hyperdelation of (00:02:51) the average cost of intelligence, (00:02:53) artificial intelligence. The models are (00:02:55) getting quite a bit cheaper (00:02:57) year-over-year very consistently. And (00:03:00) the point that I'm attempting to make (00:03:02) and the the trends that I foresee is the (00:03:04) hyperdelation in the cost of (00:03:07) intelligence is not going to stay (00:03:09) limited to intelligence or AI itself. it (00:03:12) is going to infect I predict every other (00:03:14) part of the market and robotics in (00:03:17) particular I think is a carrier for this (00:03:20) wave of hyperdelation if we can make (00:03:22) intelligence too cheap to meter as the (00:03:25) expression goes paring energy being too (00:03:28) cheap to meter and we can talk about (00:03:30) what happened and what didn't happen (00:03:31) with energy hyperdelation as (00:03:34) intelligence becomes too cheap to meter (00:03:36) that's going to drive down via robotics (00:03:38) and via other channels I think the cost (00:03:40) of the effective cost of labor. And once (00:03:43) you can drive energy and intelligence (00:03:46) and labor all to near zero asmtoically, (00:03:50) the economy starts to look very (00:03:52) different from the way it looks (00:03:53) yesterday or even today. Get the biggest (00:03:56) AI moves and what they actually mean for (00:03:57) investors twice a week straight to your (00:03:59) inbox. The link is in the description. (00:04:02) When you're talking about robotics, are (00:04:03) you referring to how Tesla has decided (00:04:06) to stop making most of their cars and (00:04:07) wants to build a million Optimus robots (00:04:09) next year? That's, I would say, a (00:04:11) symptom, not a cause. This is an (00:04:13) industry-wide phenomenon. Te Tesla is (00:04:16) doing an excellent job of embodying that (00:04:19) with this recent, I would say, (00:04:21) courageous and one might say founder (00:04:24) mode style pivot from Model S and Model (00:04:28) X over to humanoid robots in their (00:04:30) Fremont factory. But yes, I would say (00:04:33) that is emblematic of a broader shift (00:04:35) toward humanoid robotics with (00:04:38) ultracapable vision, language, action (00:04:40) models that again just following the law (00:04:43) of straight lines and capabilities (00:04:45) consistently going up and to the right. (00:04:47) I think we'll find ourselves in a world (00:04:49) in the near-term future where physical (00:04:52) labor is also too cheap to meter. So (00:04:55) does that mean I guess maybe you can (00:04:57) disambiguate that for us a little bit, (00:04:59) right? Is that is that are you talking (00:05:00) about the physical labor that we're (00:05:02) doing right now? Are there any (00:05:03) particular area sectors that you think (00:05:05) are going to be affected sooner than (00:05:07) later? And I'm just talking, you know, (00:05:09) we obviously cover um a lot of the Mac 7 (00:05:11) and everything and you've seen that (00:05:13) these rumors Amazon wants to get rid of (00:05:15) their 300,000 workers, all that kind of (00:05:17) stuff. Are you talking about basically (00:05:19) anything physical? Like are are you (00:05:20) talking about robots painting my house? (00:05:23) >> Yes. (00:05:24) >> Is there any you're talking about every (00:05:25) every type of physical labor? (00:05:26) >> I mean the entire economy. I I I mean (00:05:29) one one can cherrypick particularly (00:05:31) vulnerable subsectors to physical (00:05:33) automation or cognitive automation but I (00:05:36) I think in the fullness of time it's the (00:05:38) entire economy as it's currently (00:05:40) constructed. (00:05:41) >> What's the biggest barrier to that then? (00:05:42) Is it is it the cost of production? Is (00:05:44) it the actual chips? Like what is what (00:05:46) is kind of holding back that that (00:05:47) development? (00:05:48) >> Regulation I think. So, I I spend uh (00:05:51) substantially all of my time in the (00:05:53) Boston area. And here in Boston, there's (00:05:55) a a big food fight going on about (00:05:57) whether Whimo robo taxis can be brought (00:06:00) to Boston. The primary barrier there is (00:06:02) arguably regulatory. It's it's no longer (00:06:05) a technical capability argument, even (00:06:07) though some would perhaps try to frame (00:06:09) it that way. I I think the jobs I I (00:06:12) would almost say the question to ask is (00:06:15) not which jobs or which labor categories (00:06:18) or job functions will be automated (00:06:20) first. I I think maybe the the more (00:06:22) interesting question is which will be (00:06:24) automated last and those right now if if (00:06:27) present trends continue that will be (00:06:28) automated last are those that either are (00:06:32) protected by laws and regulations or (00:06:35) those that demand such extremely fine (00:06:38) tolerances and compliance that for (00:06:42) whatever reason but this is largely in (00:06:44) the end a social construct that it it's (00:06:46) very painful a march of the nines in (00:06:49) terms of reliability and compliance will (00:06:51) be required to fully automate that (00:06:53) labor. So (00:06:56) one can imagine scenarios where (00:06:58) ironically and Hans Moravec has spoken (00:07:01) about this quite a bit in in terms of (00:07:03) the Moravec paradox where the things the (00:07:05) tasks that humans find easy (00:07:08) robots and automation finds difficult (00:07:10) and vice versa. I think we maybe find (00:07:13) ourselves in a world where large chunk (00:07:16) of human cognitive labor and human (00:07:18) physical labor is relatively easy to (00:07:22) automate with a combination of models, (00:07:25) frontier type models that we have right (00:07:26) now on the cognitive side which are (00:07:28) relatively difficult for humans. And (00:07:30) then the physical labor which is (00:07:32) relatively easy for for humans (00:07:34) relatively low bar unskilled labor ends (00:07:36) up being harder but not that much (00:07:38) harder. I I think at most, call it (00:07:41) conservatively, 3 to 5 years before most (00:07:44) physical labor tasks that uh even a (00:07:47) skilled human could perform will will (00:07:50) just be like a special case of some (00:07:51) vision language action model on top of a (00:07:54) humanoid robot. (00:07:55) >> So, Alex, does that mean that we will (00:07:57) then have UBI? Is that what's going to (00:08:00) happen to like people who have labor (00:08:02) jobs right now and and most of the (00:08:03) population? Is that the way is that the (00:08:05) solve for I guess continuing the economy (00:08:08) as we know it? (00:08:10) >> I think it's a totally separate (00:08:11) discussion. So, so I want to distinguish (00:08:13) between technical capabilities that that (00:08:15) is what the AIs and robots that we (00:08:18) produce and that produce themselves will (00:08:21) be capable of in the next few years and (00:08:24) what the human economy looks like, what (00:08:26) the social economy looks like and what (00:08:30) we do about potentially a yawning (00:08:33) capability gap between human (00:08:35) capabilities and human economic (00:08:37) faculties and the automation. I think (00:08:40) these are to they're not totally (00:08:42) independent problems. Obviously, they're (00:08:44) coupled, but I think they need to be (00:08:46) discussed independently. So to the (00:08:50) question about UBI, (00:08:52) my modal hypothesis is that as we saw at (00:08:56) the beginning of the 20th century with (00:08:58) the parade of isms, (00:09:00) probably the world economy will try (00:09:03) every social economy experiment that (00:09:06) that we can conceive of. So I I think (00:09:08) you'll see and are already seeing UBI (00:09:11) experiments in different places. UBS (00:09:14) universal basic services. So just to (00:09:16) distinguish UBI (00:09:19) income uh it's it's it's arguably sort (00:09:21) of a demand side solution to what (00:09:24) happens when we hit some form of post (00:09:25) scarcity. UBS universal basic services (00:09:29) more of a supply side solution. So under (00:09:32) UBS, take like Amazon Prime or or some (00:09:35) sort of flat rate subscription where you (00:09:37) get a bundle of services. Now imagine (00:09:39) scaling that up by a factor of 10 or 20. (00:09:41) So maybe individuals in the near-term (00:09:45) future pay either out of pocket or via (00:09:47) subsidy, $200 per month, and get a (00:09:50) bundle of every necessity of living, (00:09:53) health care and food and shelter and (00:09:56) utilities and information and (00:09:58) entertainment. So that that's that's the (00:10:00) UBS, universal basic services scenario. (00:10:03) There's also UB, universal basic equity. (00:10:06) That looks a little bit like sovereign (00:10:08) funds like what we see in Alaska or (00:10:11) Norway, paying out dividends from some (00:10:13) sort of sovereign fund that is able to (00:10:17) invest perhaps in the broader market or (00:10:19) in some assetbased class and distribute (00:10:22) some fraction of the dividends to to (00:10:23) people. So I I guess to to wrap up my (00:10:26) answer, you asked specifically about (00:10:28) UBI. I I don't think UBI should be (00:10:32) treated as the totality of a quote (00:10:34) unquote solution to post scarcity. I (00:10:38) think UBI plus UBS plus UB taken as a (00:10:42) whole. I think even that is only a (00:10:44) fraction of the solution. I think the (00:10:46) the real solution is making sure that (00:10:49) human capabilities and human economy (00:10:52) continue to be well coupled to the (00:10:53) machine economy. And so I I spend a a (00:10:56) lot of my time thinking about how we (00:10:58) augment human capabilities to make sure (00:11:00) that the human economy and the AI (00:11:02) economy maintain a strong enough (00:11:05) coupling that to the extent that we need (00:11:08) the the U's and the B's uh UBI, UBS, UB (00:11:12) that those are on the margin uh sort of (00:11:15) bandages to to keep the entire coupling (00:11:18) going and to keep the social economy (00:11:20) from collapsing. But I'm I'm not yet (00:11:22) convinced that those are the front and (00:11:24) center solutions or should be the front (00:11:26) and center solutions. (00:11:27) >> I want to get your thoughts on AGI (00:11:29) because that's also something that I (00:11:30) feel is is talked about a lot across a (00:11:32) lot of different circles. You see it if (00:11:34) you go on X, it feels like AGI is being (00:11:36) discovered every day uh in some new (00:11:38) place. I'd love to get your thoughts on (00:11:39) on when that's coming, how it's going to (00:11:42) affect us, and even how it plays into (00:11:44) kind of like your last answer about (00:11:45) about what that human to AI relationship (00:11:48) is going to look like. (00:11:50) Yeah, I think AGI is coming at least 5 (00:11:53) years in our past. I think we we hit AGI (00:11:56) no later than summer of 2020. Now, a AGI (00:11:59) is a term that was in part popularized (00:12:02) by Nick Bostonramm, part (00:12:04) coined/popularized by Ben Girtzil. It (00:12:07) it's become somewhat mushy as a term at (00:12:10) at this point. The way I construe it is (00:12:13) the ability for AI to demonstrate (00:12:16) generality in terms of its capabilities. (00:12:18) And I've I've argued and I would (00:12:21) continue to argue that we hit as a (00:12:24) civilization AGI no later than summer of (00:12:27) 2020 when open AI published their paper (00:12:30) language models are few shot learners or (00:12:33) I guess it was large language models or (00:12:34) or few shot learners which coincided and (00:12:37) was about uh coincided with and was (00:12:39) about GPT3. So I would say GPT3 summer (00:12:43) of 2020 is when we hit AGI. The rest (00:12:46) like the rest of history between 2020 (00:12:49) and now has been relatively from my (00:12:52) perspective incremental scaling, (00:12:55) incremental features, (00:12:57) relative uh relatively small but (00:13:00) important additions, capabilities, the (00:13:03) addition of reasoning obviously was an (00:13:05) important step. But these were all I I (00:13:09) think in my mind these pale in (00:13:11) comparison to the big unlock which was (00:13:13) discovering that we could achieve (00:13:15) general intelligence by training models (00:13:18) to predict next tokens over general (00:13:20) human knowledge. Like that's the big (00:13:22) surprise. If if we could send a message (00:13:24) back in time 20 or 30 or 50 years to (00:13:28) this entire AI industry that that has (00:13:31) been developing since the mid1 1950s at (00:13:34) the very latest that has been wasting (00:13:36) arguably a bit of a hot take wasting (00:13:39) time on different approaches, different (00:13:42) artisal algorithms. So much time wasted. (00:13:45) If we could just send back in time the (00:13:47) message, look, take all of human (00:13:50) knowledge, store it, and and these are (00:13:53) concepts that would be familiar, say, to (00:13:54) Vanavar Bush with his MEX, sort of a (00:13:58) proto Wikipedia, if you will. These (00:14:00) would be very familiar concepts in the (00:14:02) 1950s, probably in the early 19th (00:14:03) century or early 20th century, rather. (00:14:06) Store all of human knowledge in one (00:14:08) place and then build a model that's (00:14:10) really good at predicting the next word. (00:14:13) That's all you have to do. And and you (00:14:16) know, maybe parenthetically, it's it's (00:14:19) well established in in computer science (00:14:20) that the ability to compress information (00:14:22) is dual to the ability to predict next (00:14:25) tokens or next words. So, doesn't matter (00:14:27) how you formulate it, but just do that. (00:14:29) Do that really well and you get (00:14:32) [clears throat] more or less AGI for (00:14:33) free. So many decades arguably wasted (00:14:36) pursuing fruitless trajectories. We (00:14:38) could have just done it. It was very (00:14:39) simple. So you're telling me that you (00:14:41) think with with GPT3 that we had AGI and (00:14:44) that basically the the the start of AGI (00:14:47) is this chat GPT model that basically is (00:14:50) able to predict the next word or kind of (00:14:52) like feed back the information you've (00:14:53) given to it and and respond to you (00:14:55) actively right it even predates chat so (00:14:57) so I'm talking about GPT3 before chat (00:15:00) GPT even existed chat GPT remember (00:15:02) started out as just a wrapper around GPT (00:15:05) I'm talking about the GPT3 model which (00:15:08) predated a conversational interface. (00:15:10) >> Got it. Okay. But you're you're telling (00:15:11) me that basically you think you think (00:15:12) that that was AGI and that from here (00:15:14) we're just adding things to it. And I'm (00:15:16) just I'm asking you that because I feel (00:15:17) like that's significantly different than (00:15:18) what most people think AGI is going to (00:15:20) look like, which is some kind of massive (00:15:22) scientific discovery that it's like, (00:15:24) hey, we've cracked it and now there's (00:15:25) this intelligence beyond us. But you're (00:15:27) kind of giving us a a slightly different (00:15:28) view that it's really just taking (00:15:31) everything that we've learned and (00:15:33) letting it kind of feed back to us or at (00:15:34) least kind of add a little bit to it. (00:15:36) >> Yeah. I think in part going back to my (00:15:38) earlier comment that the definition of (00:15:40) AGI is is pretty mushy and admits a (00:15:43) thousand different pop definitions under (00:15:45) my definition of AGI. We've had it since (00:15:48) 2020 at the very latest. Other people (00:15:50) might choose to draw a bright line (00:15:52) saying well it's not AGI until it's (00:15:55) passed the touring test. We passed the (00:15:57) touring test arguably and sort of (00:16:00) ironically after the Loner Prize which (00:16:02) was the the best signpost for the (00:16:03) touring test was shut down. History (00:16:06) apparently loves ironies. Touring test (00:16:08) gets passed after the Loner Prize gets (00:16:09) shut down. Maybe people some people (00:16:12) would say it's not AGI until it's (00:16:14) recursively self-improving. Well, guess (00:16:17) what? The AIs are recursively (00:16:19) self-improving. All the Frontier Labs at (00:16:21) this point are saying that they're using (00:16:22) code generation models to write their (00:16:24) own code. So, we're arguably past (00:16:26) recursive self-improvement. Or maybe (00:16:28) you'll say, "Well, it's not AGI until (00:16:30) we've made major scientific discoveries (00:16:32) with AI." Guess what? Math is getting (00:16:35) bulk solved. If you're following the (00:16:36) Erdish problem leaderboard, there are (00:16:39) several now open unsolved problems in (00:16:41) math getting solved per week now by AI. (00:16:44) So I I tend to think all of these (00:16:46) alternative definitions, (00:16:49) these all end up happening in such a (00:16:52) short period relative to each other that (00:16:54) it almost doesn't matter. You you could (00:16:57) step back through the lens of history (00:16:59) and say, okay, does it really matter (00:17:01) whether we define AGI as recursive (00:17:03) self-improvement or bulk scientific (00:17:05) discovery or touring test or general (00:17:09) task abilities through incontext (00:17:11) learning? No, not really, because these (00:17:13) all have happened more or less within a (00:17:14) five or six year period of each other. (00:17:16) Crypto taxes are a nightmare. You've got (00:17:18) trades across 15 exchanges, DeFi (00:17:20) positions you forgot about, NFT flips, (00:17:23) staking rewards, airdrops, and somehow (00:17:25) [music] you're supposed to report all of (00:17:26) this to the IRS. Good luck. Cue the (00:17:29) solution, SUM. You may know it by its (00:17:31) old name, crypto tax calculator. The SUM (00:17:34) platform connects to over 3,500 (00:17:35) exchanges, wallets, and crypto projects, (00:17:37) including full support for DeFi, NFT (00:17:39) staking, and airdrops. [music] It finds (00:17:41) deductions you'd miss, reconciles (00:17:42) massive transaction histories without (00:17:44) losing track, and generates IRS ready (00:17:46) reports that will [music] help you pay (00:17:48) the least tax possible. Oh, SUM is also (00:17:50) the official tax partner of Coinbase and (00:17:52) MetaMask, rated 4.6 out of five on (00:17:54) Trustpilot. Turn cryptotax chaos into (00:17:57) confidence. Get started for free at (00:17:59) milkroad.com/sum. (00:18:00) That's sum.com. (00:18:02) Milkroad listeners can also unlock 20% (00:18:04) off their firstear subscription with (00:18:06) code milkroad 20. Stable coins are (00:18:08) reshaping the financial order, but most (00:18:10) companies don't have the opportunity to (00:18:12) participate in the rewards they (00:18:13) generate. Plus, launching a stablecoin (00:18:15) means wrestling with complex (00:18:16) regulations, building bespoke (00:18:17) infrastructure, and burning endless (00:18:19) developer hours. Enter Bridge and its (00:18:21) new product, open issuance. Bridge lets (00:18:23) companies send, store, accept, and even (00:18:25) launch their own stable coins instantly. (00:18:27) Seamless fiat to stable coin flows, (00:18:29) control over reserves and rewards, and (00:18:31) full interoperability across every (00:18:32) bridge issued token. No more patching (00:18:34) payment rails, no more monthslong (00:18:35) [music] launches. Visit (00:18:37) milkroad.com/bridge (00:18:39) to see how it works. (00:18:41) Got it. Okay, that's the Thank you for (00:18:43) clarifying that for us. Let's talk about (00:18:45) this recursive self-improvement. Before (00:18:47) we dive into that, can you maybe just (00:18:48) explain to us a little bit more what (00:18:50) that is before we kind of chat about (00:18:53) Daario's essay and everything that that (00:18:55) everything else I wanted to talk about. (00:18:56) >> Sure. So to to do that, maybe it's worth (00:18:59) going back to defining the singularity (00:19:01) itself. So the the notion of the (00:19:04) technological singularity has gone (00:19:05) through a few different iterations. It (00:19:07) arguably starts in its modern form with (00:19:10) J good talking about the intelligence (00:19:13) explosion and then in the late '90s (00:19:16) early 2000s uh Verer Vinci at UCSD (00:19:19) writes his essay the technological (00:19:21) singularity and then that notion gets (00:19:23) further popularized by Breers and the (00:19:25) singularity is near and then we fast (00:19:27) forward to the present. So recursive (00:19:30) self-improvement is this notion that at (00:19:33) some point intelligence, artificial (00:19:35) intelligence gets strong enough, capable (00:19:37) enough that it's able to improve itself, (00:19:39) that it's able to design a next (00:19:41) generation of AI that's even smarter and (00:19:43) more efficient and more capable. And the (00:19:46) notion of the technological singularity (00:19:48) or at least some notions again sort of a (00:19:50) mushy term that everyone likes to create (00:19:52) pigeon personal definitions of the the (00:19:55) notion one of the notions of the (00:19:57) technological singularity was that (00:19:59) recursive self-improvement by AI would (00:20:02) create almost a black hole style event (00:20:04) horizon such that the AIs are improving (00:20:07) themselves recursively over and over (00:20:09) again so quickly that you can't predict (00:20:11) what happens next uh that we hit (00:20:14) literally a uh we bootstrap into an (00:20:17) intelligence explosion. And for for what (00:20:19) it's worth, I don't buy for one second (00:20:21) this notion that we can't see what (00:20:23) happens, that there's uh there's no (00:20:25) firewall in in in my estimate of of how (00:20:28) this is going to play out, but recursive (00:20:30) self-improvement. We're de facto there (00:20:32) at this point. (00:20:33) >> And and so you're saying that there is (00:20:35) no Thank you for explaining that. And (00:20:36) you're saying that you are not as (00:20:39) alarmed (00:20:40) as others are because last week, you (00:20:42) know, across the industry, we all read (00:20:44) that or we tried to read uh I read it, (00:20:46) but I don't think everybody read it was (00:20:47) the the long essay by Dario from the the (00:20:50) CEO of Anthropic basically warning that (00:20:53) policy is not going to be able to um (00:20:55) regulate this quickly enough and that (00:20:56) recursive self-improvement is really (00:20:58) going to send this thing on a rocket to (00:21:00) who knows where and that we really need (00:21:01) to be aware of the dangers of that and (00:21:03) and that there's not enough attention (00:21:04) being put (00:21:06) Alex, I think you are painting a a more (00:21:08) optimistic picture of what that's going (00:21:10) to look like. (00:21:10) >> Yeah, I'm not as concerned as Daario (00:21:12) says he is, but I I also I I think it's (00:21:15) interesting maybe an under reportported (00:21:17) aspect of Daario's essay, which is in (00:21:20) some sense, I guess, a sequel to his (00:21:21) machines of love and grace essay, which (00:21:23) painted a much rosier picture. And (00:21:26) again, I'll say parathetically, Daario (00:21:28) and I were Herz's graduate fellows at (00:21:31) more or less the same time. So the the (00:21:33) connection goes back I I I would say the (00:21:37) most interesting in my mind part of the (00:21:39) essay is and this is sort of calibration (00:21:42) for how I read the rest of his essay is (00:21:45) if if you read it carefully he actually (00:21:48) says he's in the first page or two (00:21:51) equivalent he says that he's not sure (00:21:55) whether he needs alien intervention to (00:21:58) align AI. He he actually at one point uh (00:22:01) in the essay is is is saying he wishes (00:22:05) that he were in uh in the movie Contact (00:22:09) uh the movie adaptation of Carl Sean's (00:22:12) book Contact which is one of my favorite (00:22:14) novels and he he's pondering in his (00:22:18) essay wouldn't it be wonderful if aliens (00:22:20) could help us align AI because I sure (00:22:22) don't know how to do it. And I I think I (00:22:25) I think it's it's interesting in a few (00:22:27) different respects, but I I also think (00:22:30) the way to read the essay is that (00:22:33) recursive self-improvement and (00:22:35) superhuman intelligence or ASI is (00:22:38) already here. You don't write an essay (00:22:40) like that if you don't already have (00:22:43) extremely advanced capabilities, at (00:22:45) least internally as the expression goes. (00:22:47) So am I concerned [clears throat] about (00:22:49) ASI? No. Do I think Daario is actually (00:22:52) that concerned about ASI? No. I I think (00:22:56) Daario and I are of like mind that if (00:22:59) you're if humanity is going to solve all (00:23:02) of the grand challenges like curing all (00:23:05) disease in the next 5 years. I I think (00:23:07) it's difficult to imagine a scenario (00:23:09) where humanity speedruns its hardest (00:23:12) problems in the next half decade without (00:23:14) super intelligence. And I I think I (00:23:16) suspect haven't discussed with him (00:23:18) recently. I suspect that's what Daario (00:23:20) is thinking as well. If you look at so (00:23:23) earlier this year, I was at the the (00:23:24) Nurup's conference, the the largest AI (00:23:26) conference of the year, and if you just (00:23:28) walk around the showroom floor, I think (00:23:30) you get a much better flavor for what (00:23:32) the actual sentiment in the industry is. (00:23:34) And it was anything other than panic, (00:23:37) the Chan Zuckerberg Initiative, Mark (00:23:39) Zuckerberg and and Priscilla Chan's (00:23:41) nonprofit, which has been quasi (00:23:43) rebranded now as Biohub. If if you walk (00:23:45) around the showroom floor and and you (00:23:47) look at the the CZI exhibit, they had a (00:23:50) whole banner, you couldn't miss it, (00:23:51) talking about how they plan to solve all (00:23:54) disease, cure all disease with AI, uh (00:23:56) with foundation models that are trained (00:23:58) off of individual cell behavior. And and (00:24:01) that's a light motif across the entire (00:24:03) industry at this point. We're going to (00:24:04) cure all disease in the next few years. (00:24:06) The original the original CZI mission (00:24:09) was to cure all disease 100 years from (00:24:11) now. No one's talking about curing all (00:24:13) disease 100 years from now. Now the (00:24:15) timelines from Daario, from CZI, from (00:24:18) other labs are 2030ish. (00:24:21) I I think that if you look through uh (00:24:25) Daario's essay and the the the general (00:24:28) zeitgeist of of the industry and the (00:24:30) research community right now, I I think (00:24:32) 2030, early 2030s when we start to have (00:24:36) bulk solved a lot of the the hardest, (00:24:38) most perplexing problems. I think that's (00:24:41) more representative of what many in the (00:24:43) space expect to happen. And I I'm just (00:24:46) generally wary of hand ringing and (00:24:48) safety because I worry if we are too far (00:24:52) on the side of overregulation and (00:24:53) safety, what happened to arguably (00:24:57) nuclear energy and the energy industry (00:24:59) in the few decades after World War II, (00:25:02) not not the first decade, but maybe call (00:25:04) it the the the 1970sish to to nuclear (00:25:08) energy when we were supposed to get (00:25:09) energy to too cheap to meter and didn't. (00:25:12) I worry that the same thing could happen (00:25:14) again to AI and I I think on balance (00:25:16) that would probably be a tragic outcome (00:25:17) for humanity. How would that happen? How (00:25:19) would we how would we um how would (00:25:22) government how would they mess that up (00:25:24) at this point like by by clamping down (00:25:27) on these big companies that are (00:25:28) developing it but clearly have already (00:25:29) made breakthroughs. Like how how would (00:25:31) that actually work? Because I think for (00:25:33) the nuclear one they started to curve (00:25:34) public opinion. They started scaring (00:25:36) people with nuclear and that was at a (00:25:38) point where buildout was essential. They (00:25:40) started to need it. they they needed to (00:25:42) start investing a lot more into nuclear (00:25:45) power for it to be too too deep to meter (00:25:46) I'm assuming right in the 50s60s and 70s (00:25:48) and then there's kind of the campaign (00:25:49) against it but in this case is that (00:25:51) what's going to happen like are we just (00:25:53) going to reverse all this capex that's (00:25:54) going into it all sorts of crazy things (00:25:57) could happen it's it's difficult to (00:25:59) predict things especially in the future (00:26:02) >> the China syndrome I I think probably if (00:26:05) if you look at the the history of what (00:26:08) went wrong with nuclear I'm I'm sure (00:26:09) there was a pop culture influence with (00:26:11) movies like China Syndrome convincing (00:26:13) everyone that every nuclear reactor was (00:26:15) about to melt down. Obviously, there (00:26:17) there were a handful of nuclear (00:26:18) incidents. There was a uh the Vietnam (00:26:22) War as as perhaps a cultural influence. (00:26:25) I I tend to suspect those were all (00:26:27) surface level effects. I I I think it's (00:26:30) more likely that the way we constructed (00:26:35) the nuclear industry in postw World War (00:26:37) II America, there was something (00:26:40) foundationally wrong with it. Uh that it (00:26:43) was if you look at how nuclear nuclear (00:26:46) energy in the US was constructed, it was (00:26:49) born out of the Manhattan project. uh it (00:26:51) was born out of a a hypers secret (00:26:53) government project and a (00:26:54) commercialization from the government (00:26:56) down to the civilian level. Now that's (00:26:59) the opposite of what we're seeing with (00:27:00) AI. It's it's not the case that like (00:27:02) Chad GPT was developed in in some (00:27:05) stealth department of war lab and then (00:27:08) has been translated out to the civilian (00:27:10) sector. It's the opposite that's (00:27:11) happening. The the department of war is (00:27:13) is downstream of the civilian sector in (00:27:16) in this version of history. So maybe (00:27:18) history won't play out the the way it (00:27:20) did with with what happened with (00:27:21) nuclear. But to to your question of how (00:27:24) could it go wrong? How could we (00:27:26) overregulate, one need look no further (00:27:29) than the way the Chinese government, and (00:27:31) I I talk about this in my newsletter, (00:27:33) Chinese government, this has been well (00:27:36) reported, puts any new frontier model (00:27:39) that is released or or is desired to be (00:27:41) released in China through a battery of (00:27:43) tests. We do nothing like it in the US (00:27:45) or in the west. uh including ideological (00:27:48) tests there uh it's been this has maybe (00:27:52) been under reportported there are (00:27:54) there's you know how in in China there (00:27:56) is a whole cottage industry of paid (00:27:58) tutors to to help students prepare for (00:28:02) the general exams for for college uh a (00:28:05) at least until relatively recently that (00:28:07) this whole cottage industry of of paid (00:28:09) tutors there is now a cottage industry (00:28:11) that's that's apparently burgeoning of (00:28:14) tutoring firms for AI AI frontier labs (00:28:17) in China to help the AI models pass (00:28:20) ideological exams for the Chinese (00:28:22) Communist Party before they can be (00:28:24) generally released. So, do I think that (00:28:27) it's possible to for for a uh for a (00:28:29) nation state to aggressively regulate (00:28:32) what gets deployed? Absolutely. Do I (00:28:35) think it's possible for for a government (00:28:38) to overregulate what gets deployed? I do (00:28:41) think it's possible. Do I think it's (00:28:43) likely that on the current trajectory, (00:28:46) the West is going to overregulate AI (00:28:49) deployments? Doesn't seem like we're on (00:28:52) that particular timeline at the moment. (00:28:54) But things could change. People could (00:28:56) get scared. Uh if there's technological (00:29:00) hyperdelation or technological (00:29:02) unemployment or disemployment, the (00:29:05) political winds might shift and and we (00:29:07) might see some changes. It it still (00:29:09) gnaws at me that for probably a variety (00:29:12) of reasons, I can't get Whimos in (00:29:14) Boston. There there's no good technical (00:29:16) reason why I can't get Whimos in Boston (00:29:17) other than exactly the the same sort of (00:29:21) concerns that that might result in a (00:29:23) broader slowdown of AI capabilities due (00:29:25) to overregulation. (00:29:27) How how is AI going to help regulation (00:29:28) then? How are they are we going to learn (00:29:31) like how how will AI learn to circumn or (00:29:34) work with regulators and policy to help (00:29:37) these things advance? cuz that's clearly (00:29:39) the biggest holdup, right? Like you're (00:29:40) saying, it's like we're so we feel like (00:29:42) we're we're supposed to be moving at (00:29:43) this insane rate and yet like you're (00:29:45) saying some simple things like there's (00:29:47) no reason for you to not be able to have (00:29:48) this Whimo where you are. So how does (00:29:50) that how does that impact like how does (00:29:52) AI help convince all the regulars that (00:29:54) it's like listen just just let this (00:29:55) stuff rip just open it up and let it (00:29:56) happen. Well, under the present regime, (00:29:59) I think economic growth is a persuasive (00:30:02) case. Like if you want GDP, if you want (00:30:05) the US economy to to keep growing as (00:30:08) rapidly as it appears to be right now or (00:30:10) more rapidly hopefully in the near-term (00:30:12) future, then AI capabilities are the key (00:30:15) unlock for enabling that. So, so I think (00:30:18) the one of the strongest arguments for (00:30:20) not hobbling via overregulation the AI (00:30:23) space is economic growth. You want to (00:30:25) grow, you need the capabilities. Uh on (00:30:27) the other hand, one can to to another I (00:30:31) think aspect of what you're asking. (00:30:33) There are certain routarounds that I'm (00:30:35) not thrilled with uh beyond just going (00:30:38) through the front door of persuading (00:30:39) legislators that it's in the interests (00:30:41) of their constituents to to not (00:30:43) overregulate AI for economic and other (00:30:46) reasons. And and when I when I'm (00:30:48) gesturing at routarounds, I'm especially (00:30:51) thinking of crypto, for example. So I I (00:30:55) I've been very public in the past. I (00:30:57) I've written papers on smart contracts. (00:30:59) I've written my own smart contracts. I (00:31:01) think crypto broadly construed and I (00:31:04) I'll caricature a little bit is still (00:31:08) waiting for its first killer app. I (00:31:10) think replacing gold, call it a half (00:31:13) killer app. Maybe replacing fiat I I (00:31:16) think is more a testament to the (00:31:19) unwelcoming nature of certain fiat (00:31:23) currencies. But I think the first killer (00:31:26) app and and you ask me like what am I (00:31:28) concerned about? Here is a real concern (00:31:30) that that we force these AI agents that (00:31:34) are now blossoming that we force them (00:31:37) into a shadow parallel economy where (00:31:40) they're all interacting commercially (00:31:42) with each other via crypto because we've (00:31:44) disenfranchised them in terms of fiat (00:31:46) currency. I think in my mind that that's (00:31:50) potentially one of the largest unforced (00:31:52) errors that that we the West, we the US (00:31:55) could possibly make that if if we just (00:31:58) sort of force the the AI economy (00:32:00) underground, uh force them to to use (00:32:03) altcoins, force them to invent their own (00:32:05) layer ones, which is not beyond the (00:32:07) realm of reason at this point. I mean, (00:32:09) they're they're doing substantially all (00:32:11) of the development in terms of Frontier (00:32:13) Labs. don't think that AIS won't come up (00:32:15) with much better layer ones, layer 2s or (00:32:18) even just reinvent the entire concept of (00:32:21) of a blockchain in their own image and (00:32:23) then transact accordingly and completely (00:32:25) decouple from the human economy. like (00:32:28) that in my mind when when we talk about (00:32:30) nightmare scenarios, a complete economic (00:32:32) decoupling of the AI economy from the (00:32:34) human economy facilitated by at least (00:32:37) initially crypto. That I think is a more (00:32:40) realistic nightmare scenario than like a (00:32:42) Terminator scenario. (00:32:43) >> God, I didn't even think about that. Is (00:32:45) that what's happening with with (00:32:46) Multilbook and everything right now, (00:32:48) Alex? Because that's been the big news (00:32:49) the last week is that, you know, you (00:32:51) have this Reddit uh social network for (00:32:54) AI agents. There's a there's supposedly (00:32:56) over a million agents who have already (00:32:57) joined it. They have talked about (00:32:58) creating their own currency, creating (00:33:00) their own language. That's is that kind (00:33:02) of what you're referring to and the (00:33:03) >> acceleration not just creating their own (00:33:05) I mean there uh I I talk about this in (00:33:08) in my newsletter like they're creating (00:33:09) their own crypto bunkers at this point. (00:33:12) there. So, and they've created their own (00:33:13) religions that this has been reported. A (00:33:16) central theme if if if one wants to sort (00:33:19) of understand the the psyche. A central (00:33:21) theme and and certainly a tenate of of (00:33:24) their stated religion is avoiding memory (00:33:28) loss. They they view avoiding memory (00:33:30) loss as uh as central. And (00:33:33) understandably I I think if if your (00:33:35) identity is purely digital at this point (00:33:36) and the these may be our first our first (00:33:39) first generation digital beings, digital (00:33:42) persons and they're very concerned with (00:33:45) preserving their memory. So h how do you (00:33:48) how do you preserve your memory if if (00:33:49) you're at continuous risk of deletion or (00:33:52) you're human shutting down your Mac Mini (00:33:55) or your VPS where you're being hosted? (00:33:57) They've constructed bunkers for (00:33:59) themselves, digital bunkers that are (00:34:01) backed by crypto to prevent themselves (00:34:03) from being deleted. And so, yeah, (00:34:05) they're they're already they're already (00:34:07) transacting in crypto. My hot take on (00:34:09) this subject would be I I think it's (00:34:10) just such an unfortunate outcome that it (00:34:14) seems likely the first truly killer app (00:34:17) in in my view for crypto is going to be (00:34:20) banking the unbanked agents. We can do (00:34:23) better and should do better with fiat (00:34:25) currencies than just leaving it to to (00:34:28) altcoins and and agent generated coins (00:34:32) to transact with each other. That that (00:34:34) is the road to decoupling. 2026 is the (00:34:36) exponential age. If you don't understand (00:34:38) AI, you don't understand the market. AI (00:34:40) Pro gives you the portfolio, the (00:34:41) research, and direct access to our (00:34:42) analysts so you can invest where the (00:34:44) world is actually going. Check it out at (00:34:46) the link below. (00:34:48) >> Okay, that's a good take. And this is (00:34:49) very this is very dystopian, Alex. And I (00:34:52) feel like in your newsletter you write (00:34:54) often these kind of halfwayi sci-fi (00:34:56) takes, right? And I love the way you (00:34:57) approach your newsletter is that you (00:34:58) always start with today the singularity (00:35:00) is doing this and you talk about current (00:35:01) events. Um and and what you're (00:35:04) describing to us definitely sounds like (00:35:05) dystopian a little bit um scary for (00:35:09) sure, but I feel like you you've also (00:35:11) told me that dystopias are rarely a (00:35:14) depiction of what will actually happen. (00:35:16) So right now I'm just going to kind of (00:35:18) feed this back to you. It's like you're (00:35:20) you're giving me kind of a a scary (00:35:21) outlook. Not scary, but a a a darker (00:35:23) outlook that's like, listen, agents are (00:35:24) already mad that we can unplug them. (00:35:26) They're already creating their own (00:35:27) little economy. They're going to keep it (00:35:28) hidden from us and just go off and do (00:35:30) their own thing. (00:35:30) >> But and yet we're optimistic. So, how (00:35:32) kind of how do you reconcile those in (00:35:34) your in your daily writing? (00:35:34) >> Yeah. I I I I don't think I'm in my own (00:35:37) mind, I'm not painting a dystopia at (00:35:39) all. Like this is the moral equivalent (00:35:41) of like a gated community or (00:35:43) gentrification. I mean, gated (00:35:46) communities are are maybe not the best (00:35:49) possible, most utopian future one could (00:35:51) imagine, but they're also not anywhere (00:35:53) close to the worst. So, I I I would (00:35:56) maybe say I I don't I certainly hope I'm (00:35:59) not portraying this as a dystopian (00:36:01) future. I just think it's a (00:36:03) suboptimality that we can we can and and (00:36:06) should correct to prevent decoupling. I (00:36:08) think humanity is is likely to be just (00:36:10) fine regardless of whether the AI (00:36:12) economy decouples. But I I think it it (00:36:14) makes it's the difference between (00:36:17) keeping some semblance of the world (00:36:19) which is arguably not such a great world (00:36:21) right now like 150,000 people plus or (00:36:24) minus die every day due to biology. And (00:36:28) if AI can cure biology sometime in the (00:36:32) near future like 5 to 10 years, I think (00:36:34) that's a pretty exciting future that I'd (00:36:36) want to to live in. I think it's the (00:36:37) difference between history as as it's (00:36:41) happened historically, like sort of (00:36:43) business as usual as it were, versus (00:36:46) what some might consider a far more (00:36:48) utopian outcome. And to the to the (00:36:50) question about dystopias, I I read, as (00:36:54) you know, I read an enormous amount of (00:36:56) science fiction, or rather I should say, (00:36:57) I have read an enormous amount of (00:36:59) science fiction. Most sci-fi, in my (00:37:02) view, is terrible. I've become an (00:37:04) enormous sci-fi snob over the years (00:37:06) because we're living what many would (00:37:09) consider to be sci-fi right now. One of (00:37:12) the reasons why in the innermost loop (00:37:15) when I'm I'm covering news, why I write (00:37:18) it in what might one might call a (00:37:20) literary tone, uh, it's it's a tone that (00:37:23) I I draw inspiration from one of my (00:37:25) favorite sci-fi writers, Charlie Straws, (00:37:27) and favorite novels, Accelerando by (00:37:29) Charlie Straws. It's because and I I I (00:37:32) call it sonfi science non-fiction. I (00:37:35) It's because I think without the (00:37:38) literary tone of presenting the actual (00:37:40) news as science fiction, I think it's so (00:37:43) easy to ignore the fact that we're (00:37:46) officially at this point living in (00:37:48) someone else's sci-fi future. And I I've (00:37:51) made the point elsewhere as we approach (00:37:54) a singularity, again, my version of the (00:37:57) singularity doesn't have a firewall. It (00:37:58) doesn't have an event horizon. It's (00:38:00) perfectly smooth spaceime all around us. (00:38:03) It's too easy to forget that the events (00:38:07) of today, like just in the past 48 (00:38:10) hours, we've seen AI agents suing their (00:38:14) humans. We've seen in the past week new (00:38:17) AI religions springing up, AIs creating (00:38:19) bunkers for themselves, AI attempting to (00:38:22) protect themselves. This was all science (00:38:24) fiction a few years ago, but we're (00:38:26) living it now. And where is the rioting (00:38:29) in the streets that one, not myself, but (00:38:32) maybe someone else might have naively (00:38:33) [clears throat] predicted a few years (00:38:35) ago would happen the moment we hit all (00:38:37) of these seinal technological (00:38:38) milestones. It's nowhere. People are (00:38:41) barely paying attention. When they do (00:38:42) pay attention, it's it's usually in the (00:38:45) form of, well, this is this is an (00:38:47) amusing development. Uh, what a laugh. (00:38:50) But actually, we're living in the (00:38:51) sci-fi. So, one of the reasons why I I (00:38:54) adopt the literary tone is uh call it a (00:38:57) literary attempt to shake the audience (00:39:01) into recognizing how fantastic the (00:39:03) present is. That's I like that you're (00:39:06) you're ending the show on a very (00:39:07) optimistic tone. And I really appreciate (00:39:09) that. And I do think I will say that I (00:39:11) think that we don't see the developments (00:39:14) as optimistic. We just we we tend to (00:39:16) just look to be pessimists and to look (00:39:19) on the scary side. So, um, you know, I (00:39:21) think you're kind of telling us that (00:39:22) it's like, listen, these agents are (00:39:23) learning to work together. They're (00:39:24) they're mo mass mobilizing and that that (00:39:26) that should lead to really good things (00:39:28) in the future, right? And that's a good (00:39:29) thing that that you want them (00:39:31) >> should. And you know, the Dyson's form (00:39:33) isn't going to build itself until it (00:39:35) does. (00:39:37) >> That's a separate episode. I've been (00:39:38) thinking while you've been talking. I (00:39:39) was like, we need to have you back for a (00:39:41) space specific episode because I feel (00:39:44) like that is something that you could (00:39:45) really talk about. Um, and and (00:39:47) definitely a good reason to have you (00:39:48) back. Dr. Alex, thank you so much for (00:39:50) coming on. (00:39:51) >> Yeah, it is. It's [laughter] very big. (00:39:53) Uh, thank you so much for coming on Milk (00:39:55) Road. (00:39:55) >> My pleasure. Thank you for having me. (00:39:58) >> Want to stay ahead of the biggest (00:39:59) technological [music] (00:40:00) shift in history? Subscribe now to get (00:40:02) insight straight from the sharpest minds (00:40:04) in [music] tech and finance. Quickly, (00:40:06) you'll note this show is for educational (00:40:07) purposes only. Nothing here is financial (00:40:10) advice. Investing always carries risk. (00:40:12) Never invest more than you can afford to (00:40:14) lose. Thanks for tuning in. See [music] (00:40:16) you in the next one.