A large number of departures from xAI from the founding team.
It wasn't clear to me whether they were fired or whether they left, you know, because they all leave on good terms.
I don't know the answer to that question. I will say that it's...
Recursive self-improvement, RSI, is the real trigger for the singularity and it happened a while ago. We're exiting the industrial age permanently. As we're talking, we're obviously going into a new world like with the industrial revolution. And I think it's scary at times to think about.
I think we have 150,000 people per day dying on Earth and I think AI is probably the best chance we have at stopping that.
Whoever is building the AI has a lot of control about how society is going to work. So I do think there's real danger along these lines of attempting to posit.
When are we going to have discovery by an AI of something as significant as relativity on its own? I don't think it's the next 12 months. I think it's...
Now that's a moonshot, ladies and gentlemen.
So everybody, welcome to Moonshots. Another episode of WTF. Here with my Moonshot mates, DB2, AWG, Mr. Exo, and a friend of the pod, someone who's been with us before, the amazing Ben Horowitz of Andreessen Horowitz. Ben, where are you today?
I am in Las Vegas today actually. It is my wife's 60th birthday weekend super celebration.
Oh, happy... Oh, you get to merge that with Valentine's Day.
Yeah. Yeah. Yeah. Yeah. Yeah.
You get to say... which is good. Good news. Bad news.
Well, last time I was in Vegas, we can cut this from the pod if we need to, but last time I was in Vegas, I told my mom, 'Hey, I'm in Vegas.' She's in her mid-80s and she said, 'Oh, you're in Vegas? Did you shoot your wad?' And I was like, 'Is that a phrase?' And she's like, 'What do you mean? I mean, did you lose all your money?'
Yeah, I'm glad she clarified that.
Thank you for clarifying, Mom.
Yeah, that would have been a very aggressive question from your mom otherwise.
Oh, yeah. That didn't stand the test of time, that particular phrase.
Well, as I like to say every week, welcome to the number one podcast in AI and exponential tech. Our job here is getting you future ready. And it is an insane week. We've actually recorded two podcasts this week just because the speed is over the top. And we're going to be recording again another four days. I mean, Ben, it's like...
It's good. My goodness. Right. Thank god my AI avatar is getting really good.
Let's open up with top stories in Voices, Video, xAI, Multis. All right. First one. Here we go. We're starting to see a little bit of doomer conversations coming. AI disruption will soon hit sooner than most expect. This is something that's been making the rounds from Matt Schumer, CEO of Other Side AI. Ben, have you seen this article?
Thoughts? So, I mean, is this what we already know? I mean, we're about to hit recursive self-improvement. Once that hits, all of our curves go out the window. Everything accelerates. Everything we've been preparing for is being redefined on new time scales. What do you think of it?
Look, I think the AI timeline is somewhat unpredictable but kind of certainly more predictable than what he's talking about which is societal change. I think that, you know, it tends... I would be very surprised just in seeing how even companies in Silicon Valley have changed so far if companies outside of that sphere, you know, just completely change everything they did in one to five years. I think that's a little aggressive for societal change. Look, we're obviously going into a new world. And, you know, like with the industrial revolution, I think it's scary at times to think about, but, you know, I feel like he kind of highlighted all the negative changes and not the positive ones. And I feel like there's way more positive change coming than negative change at a much more rapid rate. So I don't know, I thought it was a little aggressive.
I mean I'm just trying to understand why it made the rounds since this is not a brand new conversation but I got it sent to me by everybody. Dave or Alex. I think it really has to do with Open Claw and, you know, kind of the new coding models. So people in Silicon Valley are talking about AI differently since then because it is kind of different. And I think that's a trigger for this one being so viral.
Yeah, I completely agree. I think a lot of things that we've been saying for, you know, up to a year now on the pod are suddenly resonating a lot more because of, you know, Open Claw, but also a bunch of other eye-opening, nano-banana type things where, you know, denial a year ago was very easy. Denial today is much, much harder in the face of what's right in front of you. But also the flavor of this rollout has changed a lot in my mind recently because, you know, this was board meeting week for me. You know, three back-to-back mega board meetings, 1,100 people affected. And what they were thinking of before was, well, will AI be able to do what I do and replace me? No way. Now they're like, oh wait, AI is easily going to make me three times more productive. Okay, well that's the same thing, right? In terms of the headcount you need to get a job done, that's effectively the same thing. You're like, oh, okay, I didn't think of it that way. Now you're exactly right. So the hope is that these companies will grow into it and can keep current headcount and expand 3x. But if you don't expand 3x, you're still looking at, you know, a two-thirds reduction in headcount to get the same job done. So it effectively is a huge amount of displacement because, you know, big banks, insurance companies are not going to triple their size in the time frame where it has.
I also think they're not going to go to total efficiency very fast. Like, I mean, I could be wrong but like I've dealt with these guys. They've had plenty of opportunities to be more efficient and we'll see. But we'll see. We'll see. We'll see.
I have a couple of thoughts about that article.
One, I thought it was like a summary of the podcast for the last eight months. That's all we've been talking about is stuff's going to change. It seemed a little dramatic as Ben put it. And I don't agree with the timelines. But definitely there's something coming. I think it's just the reason it's making the rounds. It's just got the zeitgeist of what's exactly happening that we need to track right now, which is we're in a singularity of multiple types. So that's what I got.
By the way, it's also like super well written, like a really compelling...
Written AI. Yeah. Yeah. Yeah. I'll chime in here and just say while I enjoy writing about the singularity in general, obviously I'm doing it almost every day at this point, I found it completely unremarkable. Maybe I'm just too deep in the weeds of writing about more pressing advances every day. But the sort of style where you talk about advances that by the way are moving even more quickly than I think described in the essay and comparing it back to the COVID pandemic which I think is a relatable touch point for a lot of people like something big is about to happen. Let's be really millennialist. And you know if you read my essay it's the ultimate viral hook. Read my essay to know exactly what's about to happen and how to survive the next 5 minutes of your life. It's a natural viral moment, but I don't think the information value was especially high compared to other sources.
Exactly what I said, except Alex said it more eloquently.
Yeah. Hey everybody, you may not know this, but I have an incredible research team and every week myself and my research team study the meta trends that are impacting the world. Topics like computation, sensors, networks, AI, robotics, 3D printing, synthetic biology. And these Metatrend reports I put out once a week enable you to see the future 10 years ahead of anybody else. If you'd like to get access to the Metatrends newsletter every week, go to diamandis.com/metatrends. That's diamandis.com/metatrends.
We have a couple of articles here on Seedance 2.0 out of ByteDance. That thing good?
Oh my god, it's amazing. It's, you know, it's going to change everything. Let me play this particular video first. The response here is, you know, and you've said this before, Alex, Hollywood is cooked. All right. So, this is a video clip with a one-line prompt. Let's take a look at it. Oh my god. Just check it out.
So, you know, Tom and Brad fighting karate on a rooftop and it generates what? 10-second clips right now. Alex, what are your thoughts?
I think at the risk of sounding again ho-hum, unremarkable, we saw copyright infringement at scale already with or alleged I should say with earlier video models and we saw the industry response and we saw settlements that eventually deals were struck to handle it. I think people are at this point again maybe I'm just sounding overly jaded with some of these advances but I've seen remarkable advances in video models. I tend to think that people are so easily awed by video models that are able to show celebrity faces and scenes that they recognize that maybe they overindex on the underlying quality of the models. I think world models that are interactive in real time are profoundly more interesting than video models. I think this is just 10 different copyright infringement lawsuits waiting to happen. But I still was wowed. I'm still one of those people that said, 'Yeah.'
So I would say on this one, the two videos that I watched that were like where the entertainment quality was so high were one that Kanye doing his song in Chinese was so good that video like I watched it three times. It was that entertaining. And then the other one was the Waffle House one. And they're both kind of representative almost of a new medium. It's not like okay, this is film generated by AI. It's like no, this is a whole another thing that we've never seen before. So I think this model is at least for me personally as a consumer was an impressive step up. I'll just...
I think YouTube wins in this model, right? Because everybody's going to be producing so much content and it's going to become resonant on YouTube. It's not going to a lot of it may not go to the theaters or television and so forth.
Yeah. Dave, you want to say?
Yeah, I was going to say the same thing. Like TikTok also is the big winner. When you personalize the content, when it's almost movie quality and it's personalized to very narrow topic areas and narrow interests and many languages around the world and everything else, it just takes over. And so, you know, when the movie people say, 'Well, look, we're still a little bit better.'
Yeah, but you're missing the bigger picture, which is you don't need to be as better than a movie. If you can push the production cost down to an individual producer, then the volume goes through the roof. But the narrow casting is just so much more compelling. You know, something that you and only your group really care about a lot in full movie format is so exciting.
Yeah. I mean, the elephant in the room here though is, you know, this 2K quality multi-scene video. It doesn't just threaten Hollywood. It threatens the whole concept of video as evidence, right? Court testimonies, journalism, political campaigns.
I mean, well, and also that's going to be real time, right? Like, so then yes, just any kind of security mechanic that you have where you recognize the person via video or voice is shot to hell.
Yeah. The echo chamber effect is crazy, too.
I'll maybe just add again. I think this is several months behind the bleeding edge that I think video models have been approximately this good. Yeah, sure. You can upscale it. You can increase the fidelity of the faces. You can certainly use faces that you probably shouldn't be using. We've been able to do this for months where I think the frontier actually lies in being able to do this in real time and being able to do this on a single modern Nvidia GPU and being able to do this at a cost-effective speed. And I think again I want to avoid overindexing on just video models that produce two Hollywood celebrities fighting each other on a rooftop. We were able to do that many months ago. This has already been thoroughly litigated. We saw OpenAI strike deals with relevant movie studios for Sora 2. We saw the Disney deal. We're in some sense I think past this where we are now.
But Alex, I don't think that's the point. I think the point is this is making it out into the ecosystem of common users. I think it's the notion that yes it's the cutting edge and yes it was possible but now it's something that is going to grow in its utility and its consumer base. And then all of a sudden we've had basically democratization of film production for some time but it's now going 10x, 100x more. I mean I think that's the issue.
You're bringing up a really interesting point too which is that our starting point for this journey was probably autocomplete and code, you know, and like wow that's incredible. But then Alex in his newsletter has been tracking every event along the way so nothing surprises him.
Yeah. I'm just spoiled my problem at this point. I'm so spoiled.
So 6 months ago but if you think about it...
On something though that Peter said like the thing that I do think is different and it's not any aspect of it but it's the combination of that you could give it a one-line prompt and produce something entertaining. Isn't something that we were at least I hadn't seen at this level of entertainment. So I think from a consumer product standpoint as opposed to a technological standpoint, which I agree on, it's kind of exciting.
That too is exactly what you're saying there, Ben. Like if the later your first exposure to AI, the more of a holy crap moment it is because, you know, it's something truly mind-boggling to the unexposed. And you're still seeing that when you survey around at random in a city, you know, outside of San Francisco or Boston, the exposure rate to AI is still very, very low. Shockingly low.
The first thing you see is so mind-blowing.
Yeah. For me, this was ho-hum because, you know, if you trace the trajectory of where we've been going, you should expect to see this about now or even earlier. So, there's nothing radically magical about this. I may hit a new group of users. Yeah. People go and another batch of people goes, oh my god, fine, there's another segment falling over into the new world, great. And the more voices out there, this is the plus side of also the Schumer essay, yeah, we know, but the more voices out there talking about this the better because it'll accelerate the whole thing.
All right let's get to this second ByteDance Seedance 2.0 article. Here we have Seedance 2.0 was paused by ByteDance after it was found to recreate real voices from just facial photos. I find that almost impossible just from a facial photo. Alex, what did you learn about this?
I think there's something interesting here, which is as we start to scale data sets, it is possible that we could start to see positive transfer between modalities. That's unexpected. We've spoken in episodes past about people claiming that they're uploading their whole genome into Claude and being able to generate facsimiles of their face that resemble the real thing. I do think it's possible that if all of YouTube and all of the world's video were uploaded into a single joint embedding model, which is the foundational technology behind Seedance 2.0, I do think it is conceivable that if we just aligned all of the world's audio and spoken audio with all of the world's faces, we would find some positive transfer between the two and be able to reconstruct to reasonably high fidelity your voice from your face or your face from your DNA or some attribute from some other attribute. I do think it is possible with enough scaling. Whether Seedance 2.0 actually achieved it or whether it was just a happy coincidence. Hard to tell at this stage.
What's interesting though is that they voluntarily stopped it, right? They voluntarily shut it down. Which is a great move by a hyperscaler. But the reality is once it's out of the bag, once you know it can be done, it can't be uninvented. So it's out there if in fact it works.
Yeah. Totally. So, so far, you know, ByteDance is like Google, they have to be cautious and conscientious. They can't just... But then every time this happens, a small startup then does it again right after, you know, they don't care because they're a startup.
All right, I want to play this video clip from ElevenLabs. You know, I think all of us have ElevenLabs voices that we use for different projects and so forth, but I was just blown away by this. And it's the humanlike quality, the cadence, the hums and the hms that come out of this. So, let's take a listen and discuss it because this is a game changer for me. I know what you're going to say, Alex. This has been around for a while. Not saying it in advance. But let's take a dive in.
So, I don't know. I'm moved by that. Both excited and frightened in a way that is, you know, again, we're cooked in terms of, you know, at my home, our family picked a secret code word. And again, everybody listening, if you've not done this yet tonight at dinner with your parents or your kids, pick a secret code word. If someone's asking you to do something that is kind of unusual or crazy that you don't expect, you may be talking to an AI. So Ben, you may not know this, but we have a fool your spouse challenge for this calendar year. The first podcast listener who can fool your spouse for three minutes on a Zoom call and has to be a totally fake you may or may not have to record it and send it in. Peter is always asking the multis, the Open Claw agents to dox him by calling him at home.
Yes, that's my AGI has arrived when an AI is calling me at home.
So, you know what happens? They end up emailing me. I get several emails probably per day at this point from Open Claw agents asking me what Peter's number is.
That's hilarious. And when they call me and ask them where they got it from, Alex...
That's right. Maybe just a comment on the ElevenLabs. So, if you've been using the version 3, v3 alpha model from ElevenLabs, this should hardly be surprising. V3 enables you if you use text to speech in the ElevenLabs platform to specify with brackets emotional expressions and this has been around for many months. What's somewhat interesting here to me at least is we've known how to do audio to audio transfer for probably a year plus at this point. If you've used advanced voice mode from OpenAI, you've used audio to audio transformers. But what's somewhat interesting here is ElevenLabs is better known for text to speech than speech to speech. And as far as I can tell, this new expressive mode that we're talking about here seems to still leverage the text modality, which historically has been very difficult. You'd have to go from speech to text back to speech, which was high latency. It was slow. It didn't feel very conversational. And somehow it seems like without having to do direct audio to audio, ElevenLabs has found a way to do speech to text, back to speech in a way that feels natural and turn-taking and real time. So I think it's in some sense an incremental advance, but another sense if they really are still keeping text, which is much easier to mine and analyze, in the loop, it probably is a material advance.
Yeah, Alex, I would say we've crossed the uncanny valley on voice at this point with this demonstration. And then voice becomes the new interface in the AI era, right? I mean, I can't tell you the amount of time that I'm just speaking to AI versus this cumbersome typing at it. So, I think those two things are really an important takeaway from this for me.
I think that's a great way of putting it, Peter.
Yeah. But the problem is, do you really want to use audio as your primary modality? I mean, it works well in isolation.
Okay. Yeah, you want BCI. I want a BCI too and I want wearables and I want gestural interfaces. I want it all. For most people. I mean, is it New York, this is sort of famous anecdote, New York Times in the 1980s did a famous study where they just put all of their reporters on then state-of-the-art speech-to-text systems and asked them to use voice. And what happened was the writing quality went down. Why did it go down? Because it's difficult to think ahead as well as you can when if you're just typing. If you're speaking, you're leveraging, right, similar portions of the brain. So, I'm not 100% sold yet that speech is the modality of the future. I do like BCIs. I do like wearables and gestural interfaces. I do like typing, but speech I think jury is still out for me for high bandwidth operation.
I was just going to say for regular people who don't necessarily write for the New York Times I think speech is often or at least in my experience is the mode of choice. For sure. The other thing I'd say about ElevenLabs and kind of quick disclaimer that we're a big investor in ElevenLabs and my daughter Sophia works there so I'm biased towards them. But the, you know, one of the really amazingly just kind of landscape shocking things to me about ElevenLabs was, you know, when we invested originally, the big question was, well, like aren't the state-of-the-art models going to be able to talk? I mean, like, of course they're going to be able to talk. But speaking correctly with the right nuance and building the right products for developers and so forth has proven to be very sustainable for them. Which I think is interesting as you look at the entire landscape the difference between the capability and the product is significant.
Yeah. You know what's amazing to me about ElevenLabs? We have two companies here in the lab that do voice, Run and Vocara. And what's amazing is these are self-organizing systems that are trained off raw data and what they do well just blows your mind. But you know within voice it turned out the turn management was very, very hard.
And you're like what I didn't ever thought like it seems so trivial compared to actually doing these incredible synthetic voices that can say intelligent things but they don't know when to stop talking. You know, like on this podcast, when I stop, you start. When you stop, it's like very natural to us. But because that wasn't in the training data, they're just, you know, up until now terrible at it.
Huh. Wait. According to our listeners, we're talking over each other all the time in a very unsynchronized way. So the exact same thing as like we have so much to learn from these speech to text to speech models, not turn taking.
Yeah. Oh my god. Yeah. All right. Yeah. All right. Next topic here. If you guys are okay with it, I'll move us along. So, just following the merger of SpaceX and xAI and in the spirit of full disclosure, I'm an investor in both. Probably you are as well, Ben. A large number of departures...
Yes. A large number of departures from xAI from the founding team mostly ethnic Chinese co-founders and it's likely due to the ITAR regulations of SpaceX, right? And it was interesting on a few video presentations that Elon's done that's had the team from xAI at least half of the team there was ethnic Chinese and, you know, the culture there breeds incredible mathematicians and programmers. Any comments?
Ben, love to get your thoughts on this, but I looked at the timeline and the exodus of senior AI talent predates even the decision on who to merge with what. So the theory that this is related to ITAR and it wasn't clear to me whether they were fired or whether they left, you know, because they all leave on good terms. I don't know if you have any insight.
Yeah, I don't know the answer to that question. I will say that it's so in a related matter we recently heard from a few Chinese nationals who are PhD students that the Chinese government is cracking down on the Americas or US academia's use of Chinese open source models like they particularly post the Meta Manus acquisition, they're very, very worried about actually secrets going this way. So kind of the opposite of what we've been worried about which our whole open source work is based on the Chinese models since there haven't been as many US open source models. So this whole thing I think is about to get more complicated.
It is. I mean, if you read these tweets, if you read these tweets, Mr. Woo and Mr. Bob both are super enthusiastic about xAI. They're not leaving with any kind of angst. You know, they're saying that they love the xAI family. We're heading towards an age of 100x productivity. So, I don't think they would have left on their own accord. That's my personal opinion. So, what drove it? You know, enter Alex's comment from last time. Well, when did their stock vest?
I mean, there is another explanation which is just SpaceX is a very large company relative to xAI's headcount and maybe there was a natural reorganization that happened as a result of that. I'll also point out that xAI was selling foundation model services to the Department of War prior to this merger. So I would again query whether some sort of nationality concern was really the trigger. It seems more likely to me this is just the result of a natural reorg of xAI starting to merge into SpaceX.
Yeah. Well, if it were a result of ITAR and such, you know, America's AI dominance is really built significantly on immigrant talent. So if we're going to start having this kind of a reaction to PhD immigrants, I mean I personally think and I'm curious what you think, Ben, that everybody going through a PhD program in the US should have a green card stapled to their PhD when they graduate. I think the idea of kicking people out is the wrong approach.
Yeah, I think in general that's correct. I do think that there's... I don't think we've quite thought through the case of China totally in that look we have amazing Chinese nationals who work for us and in every company we have, I think...
So, you know, the talent that comes here is extraordinary. I would say there is no risk to that idea. I guess I would just say that but...
Right. I think being open, I think you're right, General. I think we should be open. We should accept the phenomenal talent that comes over here and not fight it because we'll lose anyway. I mean like we're not keeping anything secret in America. Certainly not in American companies. None of them have skiffs. None of them have good security protocols around personnel.
Don't get me started on the idea that privacy is long since dead. But so here's correct.
Here's a tweet from Jimmy Ba, a co-founder at xAI. Recursive self-improvement loops likely to go live in the next 12 months. 2026 is going to be insane and likely the busiest and most consequential year for the future of our species. So you're going to hear this a few times...
Go off. This is I mean you're going to hear this a few times that you know these next few years are a massive inflection point that everything changes and we're going to look back thousands of years in the future back to this inflection point or is it just a smooth singularity? Alex...
I think it can be both. I think we've already hit the era of recursive self-improvement. I'm banging the table rhetorically every episode and every day in my newsletter talking about recursive self-improvement. We're there. All of the Frontier Labs are using their own models at this point to develop their models. That's practically the definition of recursive self-improvement at this point in practice. I don't think it's the next 12 months. I think it's now. And is 2026 going to be insane relative to years past if you just sort of skip over all of the interim time? Absolutely. Is it already insane in some sense? Absolutely. Even if we just look at the events of the past 24 hours, some of which I think we'll get to like self-replicating AI and courts where AIs can mediate their own disputes in front of an AI jury. That would have been pure science fiction several years ago. That's not 12 months from now. It's not quote unquote within the next 12 months. That's the past 24 hours. So I think...
Yeah. That's like past 24 hours.
So I think Jimmy is underselling if anything what craziness looks like. At the same time locally, space-time is smooth and you need to look no further than my saying ho-hum to a few of these stories as being so several months ago. That would have been sci-fi years ago. But you get a little bit spoiled living inside the innermost loop.
I would say I do think there's a delineation between recursive self-improvement with a human in the loop and without one. And I think he seemed to be implying that there'd be no human in the loop, which I think is an accelerant. TBD, how much of an accelerant? But I think that could be very different.
Yeah. Permissionless self-improvement, right? Like flip the switch and go as fast as you can.
Well, also there's a distinction, this came up at the EverQuote board meeting yesterday, but the self-improvement where the inference time speed algorithms are being improved by AI, clearly well underway, way down the path. And then inference time speed can be directly translated into intelligence. Now we now have the know-how to turn more inference time loops into a higher IQ. So that is clearly underway. The question is does that notch up then allow it to work on the next part of the algorithm, the next part of the algorithm, which case we've already hit the flash point and now we're just talking about the rate at which it percolates across. But Ben, I agree that his comment is about the actual core algorithm development. The next ideas are 100% from AI and then they go into production. You know...
I also think human outside the loop, on the loop, in the loop is in some sense a pretty blurry or slippery slope. If you remember George Jetson from the Jetsons, George the cartoon...
Remember George Ben? George would go into work and he would complain about his finger. He'd have to press one button all day with his finger and then complain that his finger was sore and the finger would be swollen from pressing a single button all day occasionally. I think that's a good metaphor for the state of recursive self-improvement inside and outside the frontier labs right now where you have Claude Code instances and Claude is asking you every few minutes, do I have your permission to do the following thing? And you press the George Jetson button, yes, I approve. No, I don't approve. And we're all sitting here complaining about our swollen finger from pressing approve, approve, approve for Claude Code running Opus 4.6 with agent teams. But really it is recursive. I would argue it is recursive self-improvement even if we're pretending we're in the loop by pressing the George Jetson button.
Exactly what I'm doing on Telegram right now with Skippy. Yes, go on to the next stage. I don't want to slow things down. And as we've said so many times, this is the slowest and most expensive it's ever going to be.
Look, a couple of points here. One is we've been talking for a while that recursive self-improvement, RSI, is the real trigger for the singularity, and it happened a while ago. So, all we're doing now is kind of accelerating that path. We're exiting the industrial age permanently as we're talking. So...
Yeah, I really think the minute-by-minute unfolding of the singularity is the most fascinating thing I've ever experienced. And you know, Alex is exactly right. There is this point in time we're in right now...
Where there's a human in the loop contributing, but it's really ambiguous what part of the progress is AI versus human. If you're in the actual coding process, was that my idea? It was kind of half my idea, but then the AI suggested this other thing and I kind of adopted it, and now it's not clear whether it was my idea or not. But we're in that mode right now where a lot of the research in these core algorithms is just 'deploy these 500 tests for me and tell me which hyperparameters worked better or which neural topology worked better.' It's not like inventing relativity, you're discovering relativity. It's just a lot of experiments with different trials and then taking the one that worked and redeploying it, and now you have a smarter AI and now it's trying more trials. It's really very likely that we're well down that path.
And I also add, I do think we're going to discover the next relativity or equivalent of relativity in physics as well with AI. I'm predicting.
Super interested in that as well.
What would you like to hear?
When are we going to have discovery by an AI of something as significant as relativity on its own?
Yeah. I would also—this is a great bellwether, Alex. The transformer algorithm in 2017 kicked off everything that we're experiencing right now. To me, the AI has already discovered things in the last six months that are harder to discover and harder to solve than the transformer was. But I'll ask you, Alex, if you feel like that's true too. But it seems pretty clear to me that the algorithm is just not up there with relativity in terms of complexity.
Well, I mean, I guess many of these discoveries boil down to insights that can be distilled down into equations. And you can, in principle, with special relativity, you can do a number of thought experiments. You can compare it fortunately with lots of experimental data. And one could imagine a thought experiment—one of my favorite thought experiments is a Bayesian superintelligence. So if you had a video of—this would be Newtonian gravity, not special relativity—but you could imagine taking a superintelligence, making it watch a video of an apple falling from a tree. And the argument in the thought experiment goes, within three frames of that video, it should have concluded Newtonian gravity is a pretty high likelihood posterior probability of explaining the universe. And with a few more frames, somewhere in its distribution—and there's a whole class of information theory devoted to what's called Solomonoff induction, just devoted to thinking about how you can efficiently infer theories of the universe from limited observations—somewhere in that distribution of theories should have been general relativity. So I am incredibly bullish that we'll be able to, with superintelligence, discover any new laws of physics, discover transformative inventions. For disclosure purposes, since we're playing the disclosure game, I have a portfolio company, Physical Superintelligence, that's working on these issues as well. I'm very bullish on the space.
Amazing. I want to move us to a conversation that Eric Schmidt recently had. I pulled a clip out of it and the question that we've talked about on the pod before. It's been the debate: can AI be paused? The answer quite clearly by almost everybody today is no. We had that conversation, Dave, you and I with Elon. Let's take a listen to what Eric has to say.
This technology is going to happen. It's not going to get prevented. It's not going to get stopped. There's too many countries, too many people, too many incentives. It's going to happen. So, what does this mean? It means great solutions for healthcare, new drugs, better energy solutions, better power distribution. It also means that it can be used for bad. It can be used, for example, for oppression. It can be used to limit freedom in governments, hopefully not in the West. It can be used in war. The technology itself is so addictive that it can affect our young people and we should make sure that our young people are protected from some of the worst parts of the technology. We face choices now about how we want to deal with this incredibly powerful technology. I will tell you, and it's really important to understand, that we are living through a moment that will be in history for thousands of years: nonhuman intelligence arrived and it was a competitor to us.
Amazing, right? We hold two futures in superposition, right? One future in which AI is the greatest advocate and supporter and accelerant to human spirit. Another future where it is a dystopian outcome, and it's ours to shape in these next few years. Dave, you were going to say...
Oh, Alex and I were there in the dome. You know, that was our Davos dome where he was speaking.
And it just blows my mind how he owns a room. The guy is so articulate.
Well, he's going to be opening the Abundant Summit this year and we'll be going into a lot of this conversation. Dave, you and I are going to be with him on stage, so that'll be fun.
Well, what he's describing in that clip was very similar. We did a whole interview of him, which you can find on YouTube. And yeah, he made that point in our interview of him as well that there's no force that's going to slow this down. And so all the people picketing and walking around saying 'stop it,' you're just wasting your time. There are probably ways you can help and contribute and help point it toward good, but picketing with a sign on the street is a complete waste of your time. It's not going to slow down. Sorry, Alex, I cut you off.
No worries. I would take the position—I think AI in fact can be paused, but it shouldn't be. We do know ways to pause it. Various folks have described both in science fiction and in realistic prognostications, in some case normative recommendations. But Larry and Jihad from sci-fi, or the Asilomar guidelines maybe in the case of recombinant DNA...
But those were guiding, not pausing.
Well, Asilomar was, unless you know of some other story, pretty effective for the first two decades in discouraging or pausing recombinant DNA entering the human germline, unless you have a counterexample.
It was just the five, you know, P1 through P5 labs in terms of safety and self-regulation. Yeah, we did pause human germline editing for sure.
Yes. So, we are capable of pausing a technology if there's a desire to. I just think in the case of AI there isn't, and arguably shouldn't be, a desire to pause it. I think we have 150,000 people per day dying on Earth. And I think AI is probably the best chance we have at stopping that. I think Nick Bostrom, who also in the past 24 or 48 hours put out a wonderful essay that I'd recommend everyone read called 'Optimal Timing for Superintelligence,' argues that AI can and should be paused, but only once we're on the verge of superintelligence—how he defines it, not how I define it. I think it's already here. That's sort of like station keeping. You want to get into the harbor, I think is his analogy, as quickly as possible. But when you're about to approach the dock—I'm mangling his metaphor a bit—you slow down a bit. You pause as you're about to dock. I think that sort of concept makes much more sense than some sort of tech-messianic six-month pause.
I'm sorry, I'd like to throw in a couple of things here. I find this is another ho-hum thing. Yeah, Eric is fabulously articulate, but we've been saying this for months, a year at least now. We've been talking about the fact—and there's a bunch of dimensions—that AI cannot be paused at all. One is once you have a downloaded model, people are going to do stuff with it. We have a global prisoner's dilemma model going on here. The whole thing is going to scale now no matter what we do. You know, OpenAI did two things that kind of unlocked this Pandora's box. One, it wrote code, and the second, it was released on the open internet. Once you do those things, you're done. Pandora's box is gone. We're talking about the barn door after the horse is bolted. We're using an 18th-century metaphor to try and understand what's going on.
One of the points he made was everybody's economically incentivized to keep it going and race it along. The US, China, you know, all the incentives are in place that make it highly improbable and almost impossible. Ben, what are your thoughts?
Impossible. Impossible. So, I think, you know, I look at this question through the geopolitical lens, which is, you know, clearly I don't think there's any kind of leverage where we would get some global—I mean, especially when it's on people's laptops as well—where we would actually stop the technology. Like, I do agree that it's impractical. I think that there is a real danger, and we faced it probably more in the Biden administration than in this one, but it's still like a potential movement that we really slow down AI progress in the US to the point where the other thing that he mentioned, the threat to freedom, becomes completely out of our control because whoever is building the AI has a lot of control about how society is going to work. And so I do think there's real danger along these lines of attempting to pause it and maybe not actually pausing it but slowing it down enough in the US that we just become far enough behind China that it's a real problem, or whatever society Xi Jinping thinks we should be.
Yeah, completely agree. I think if you listen to Eric's words closely, he wasn't saying we couldn't pause it. He's saying that because we're in the middle of an all-out arms race with China and we're only one year into a presidential administration, so it's going to happen in this next three years. So given the current administration and the current situation with China, there is no chance of it being paused. And so react to the real reality. Don't hypothesize something that is just never going to happen. But it was all geopolitical.
Speaks to the motivation. I'm speaking of the fact that there's no mechanism to posit or stop it at all. None. You'd have to regulate every line of code. I mean, come on.
Oh, there are ways to do it. I think not being able to think of it speaks well of your character. You have to imagine a completely pathological society. Vernor Vinge wrote quite a bit about this. We have these excess transistor budgets. Imagine a society where you have literally transistors spying on other transistors on a single SOC. Imagine you have people spying on other people. Imagine there are bounties. Anyone discovers anyone else doing something that's algorithmically impermissible either at the logical level or the social level. You can construct a sufficiently pathological society where people are turned against each other in order to suppress AI. At least I can imagine it.
I can give you an actual real-world example. The last executive order from the Biden administration was that you could not sell a GPU without US government approval, not a single GPU. So I think that wouldn't stop AI, it would slow it down enough in the US that it would be extremely material.
Speaking of this subject, Alex, I put this slide up after our conversation. Do you want to...
Peter inserted this slide, Ben, as an indulgence to me. So I have to ask you this question. You and Mark towards the end of the last administration were very public making comments that you took a meeting at the White House, apropos, and if I'm relaying the comments accurately, you were dismayed to hear about plans to classify AI progress just like—and again, correct me if I'm mischaracterizing—advances in math and fundamental physics had been purportedly classified or overclassified for decades. And I'm curious at a few levels. One, if that's accurately characterizing what you heard, what do you think was classified? What do you think was the impact on the economy and the world from such classification or overclassification of math and fundamental physics? And what would you have done differently than if you had been in charge?
Yeah. So I can tell you what was said. I said, look, you know, I was trying to be pragmatic. I said, at the core, AI is math. That is what it's doing. It's math. So if you start restricting the models and you start regulating the models, you're just regulating math. You're outlawing math in some way. Either you're outlawing parts of math or you're saying you can't do enough math. And he goes, 'Yes, we can do that.' Like that was his answer. He goes, 'Yes, we can do that. We did that in the 40s around nuclear physics. And some of that stuff is still classified today.' And one, I was shocked, like my jaw hit the floor. I was like, 'Wow, that's crazy.' And then this would be even crazier. But I don't know what it was, but if you think back about—I'll just make this comment—if you look at the progress in the US and in the world in physics up until kind of the Einstein, John von Neumann era, and then since then, it's pretty startling how little progress we've made. I would just say many of the ideas that have come since then don't seem to work. And hopefully we'll get to the other side of that with AI figuring things out. But I do wonder, did we put something away that we knew that would have unlocked some of the problems we're trying to solve?
That's fascinating. Okay, that is for the record that is what I assumed you meant and/or heard or inferred. So maybe just if I may, the second part of the question: what would you do going forward now if you knew for a fact that such classification of fundamental physics had in fact happened? Like, how would you fix the world?
