Latest / Last Humans in the Loop / Are you marketing to AI yet?, Humans talking like AI now, Live AI made TV shows, China exports surge
Transcript
- Christopher: Hey everybody, welcome to this, the next installment of Last Humans in the Loop. I'm Chris Slow. I'm here with Zach Wienersmith and Marty Wiener as ever. And today I'm actually really happy to announce we have our first guest. we will be ending ending this episode with an interview with professor of mathematics Jonathan Wise to figure out if the if the solution of do do a proof is the future of mathematics. so stay tuned for that one. Let's get right into it though today, gentlemen. so birthdays and anniversaries. so on on this day, on our August 29th of 1997, Netflix was founded by Reed Hastings and Mark Randolph. the actual webs website landed landed a couple months later in April of ninety-eight, but of course, major change in the in the industry from that one. Skype launched in 2003, also on August 29th. and going back into the well, John John Mockley was born August 30th, 1907. for those of you who don't know, John and Jay Presper Eckert built the ENIAC, which is one of the seminal first computers back in 1943. they subsequently went and worked on EDVAC and Univac. Univac, of course, being highlighted in such films as 2001 Space Odyssey as being the future of computing, as you recall. let's jump Marty Weiner: I still use one. Christopher: I mean you have to have your vacuum tubes. How else do you keep your office warm? That's my that's my solution. Zach: That's right. Christopher: all right, let's kick into the to the news. so apparently Sam Altman says that OpenAI will achieve AGI by the end of the year. the c I don't think you mentioned which year. maybe Marty Weiner: Well i i i and it's newsworthy 'cause he hasn't said this since last year, so that's exciting. Zach: That's right. Christopher: It's been a whole year since we've discovered that we're gonna have AGI by the end of the year. So we're super excited to hear about that. you know, I I for one look forward to being able to hand this off to to an AGI to run the podcast. Marty Weiner: Do a podcast. Christopher: do a podcast. in other news, speaking of open AI, open AI pulls out of cursor, specifically citing Elon. Zach: Yeah, so this was amazing. you know, SpaceX just took its option to buy Cursor for 60 billion, I believe. And OpenAI just put out a note which was like surprisingly non-legalistic. I'm gonna read a section of it. This decision was incredibly tough as we care deeply about our models being broadly available for developers. We are making this choice because we cannot be confident that SpaceX will use our technology with our terms of service based on our experience with Elon Musk's companies violating contracts. Marty Weiner: Wow. Zach: Yeah, so, like, I opened this up and I was expected to be like, you know, lawyer stuff. And it was like, Elon Musk is a liar. So goodbye. And as I actually want to ask you guys who have more like business experience, like, is this just personal? I mean, I know you don't know, but like, do companies do stuff just because the CEOs hate each other? Or do they actually believe this? Okay. Marty Weiner: yeah, yeah, yeah. I was I w well, I kn they don't usually come out and say it this vocally, but but I think we've seen the Sam Altman Elon fight for a little while. I was at a company where supposedly our CEO went into a discussion with the CEO of Sun and they each brought their own new chip and our CEO slapped it down and was like, I'm coming after you, and the other ones had his chip sitting in his pocket and he slapped it Zach: You Marty Weiner: down, you know. And the rest the Zach: It's, it's, Marty Weiner: rest of us are actually trying to build shit and these guys are just, you know, hissy fitting. No, no, no, there's a lot of irrational behavior. Whether or not they do it in Zach: Yeah. Rational capitalism at work. Yeah. Marty Weiner: public is another thing. Now Musk has definitely been showing that he is not trustworthy in in a thousand directions, but I've never seen anything quite so beautifully in the public. More. I want more. Zach: Yeah. Christopher: I I do think what's interesting about this what was interesting though has been the fact that unlike I like traditionally the way these are handled in press releases is with some sort of like legalistic or an actual press release that involves some sort of like the the probably the comms team is involved in those but we felt it but we felt it was time to move on. Or like, you know, ultimately Marty Weiner: The comms team, yeah. The Zach: Yeah, we felt it was time to move on and go in different directions. Marty Weiner: We wish them luck. Christopher: It's an it's an easy argument here too, because if you think about it, right, OpenAI and Anthropic and SpaceX now are all competitors in the same space trying to trying to basically pedal models. And so the fact that they would start to not use each other's technology, that's kind of the done thing. I look the same thing happened with the clouds. Here it's just literally like, Yeah, we just don't like that guy. Like we're just gonna like we're pulling out and say we decided we're not gonna do it. So I'll go with I'll go with I'll go with re refreshingly honest maybe? this Zach: You Christopher: is kind of a nice nice Marty Weiner: Yeah. Yeah. Christopher: well speaking of news in news in open AI, we should we should just rename those this one as the open AI news section this time around. Marty Weiner: Apparently. Yeah. Christopher: word in the street we have Astra. The new model from OpenAI is gonna be coming out supposedly within the next week, which is timed, I believe for Fable Five One is coming out on a similar timeline. So maybe by the time this actually this podcast lands, those will already be in the in the in the rearview mirror, given how quickly these things go. OpenAI also announced that their their jalapeno results are in. Jalapeno, for those of you who f have haven't been tracking, is their own brand of of of chips that they're going to try to use to replace their d dependency on NVIDIA and other other peddlers. I I don't know what the joke is here. I assume it's because they you they may you'd say you would think they actually wouldn't want to run as hot and say jalapeno. Maybe I guess it's best is not it's to be Marty Weiner: It's it's spicy. Yeah. Christopher: to be followed with ghost pepper is the future one. Marty Weiner: Mm-hmm. Zach: you Christopher: they apparently the benchmarks are there better than Blackwell. now, of course, the benchmarks are coming from OpenAI, so take it with the grain salt. But I would say in this case, this the OpenAI was able to bootstrap an entire chip arm in about eighteen months with apparently LLM assist. So it's interesting from that standpoint alone. and b and be competitive. Marty Weiner: Yeah, and be competitive. Yeah. Zach: Yeah. Although we're worth noting like we didn't see like Nvidia stock price drop. So it doesn't seem like everybody believes that this has been a complete broadside. Christopher: Is NVIDIA's stock price allowed to drop? I thought that wasn't allowed anymore. That they were only only one direction. It only goes because, you know, the demand the demand curve hasn't been met yet. So Zach: That's right, I think that was bad. The line goes up. Marty Weiner: Up and to the right. Christopher: and and last but not least, Apple Apple announced its new M6 chips early. I think they've actually had Apple's had a a rare a rare bout of early press releases on new technology. usually they are pretty tight lipped about this about this stuff, but even they are in the hype cycle. six six chips with an obvious target towards doing AI compute. and Marty Weiner: Yeah. Christopher: at some point maybe that rebranding of AI being Apple intelligence will actually stick. Marty Weiner: Yeah, a and what's crazy is th they're even short of this chip, their new lines of Mac minis and and pro computers are flying off the shelves simply because they're pretty well targeted toward running local LLMs, including lots of RAM. So they've actually been caught kind of surprised by people wanting their their computers. Christopher: also been a rare situation where the the new hardware is priced above the price of last generation's hardware just by simple supply chain constraints and it's still flying off the shelves. Zach: It's good. Do you guys think like I was reading a guy who saying we're not going to have cheap RAM for 10 years at the soonest. Like is that just the world Marty Weiner: my god. Yeah. Zach: we're in now? Marty Weiner: No, I I do think maybe that it's different from some other areas because I think assuming that there's a green field of everybody wants more RAM, I think it it's different from like the toilet paper situation where the we're we're running remember w we all ran out of toilet paper during COVID. There people were like, Well just make more and the problem is we were already utterly efficient at it. and there wasn't motivation to make more because by the time new factories are turned up we're gonna be caught up, right? I mean there's only so much toilet paper that could be actually consumed by families at a at a certain rate. But with Ram, I think that I I wouldn't be surprised if a whole bunch of companies suddenly jump up to meet the demand and we actually get back to a to a s kind of a more of a steady state that makes sense. Christopher: With the issue being that the time to bring up a fab from scratch is somewhere in the vicinity of two to three years. And so I think the hard part here has been there's Marty Weiner: Yeah. Christopher: cons there's been concerns that if they bring fab up and then the and the market decides to tank, then they're left holding the bag with this giant capex that they just spent a bunch of money on. so it's a it's a tricky it's a tricky situation. And there has the industry is not is not fast to build new facilities. You're talking about multi billion dollar capex investments. Marty Weiner: Maybe it is now with AI. Zach: Yeah. Christopher: You never know. well let's jump into let's jump into some deep dives. so qu first first interesting bit is that there seems to be some indication that marketing as a field is changing simply because AI is now a consumer. Marty Weiner: Yeah, so so I started diving into this a little bit with asking a marketing friend about how AI has affected their industry, and it started to become clear to me that you know, marketing has been built on the assumption that the humans are the final consumers, right? You know, if you're selling Coca-Cola, a human's gonna go buy that and you're trying to get them to buy that over Pepsi. and in addition, you also have a bunch of intermediers, right? So if you're a restaurant you're gonna try to target these restaurants to choose your product so that your ultimate consumer has a limited selection, and one of the limited selections is is your product. And what we're seeing now is the AI is now being freed up to purchase. So the AI is now one of the purchasers. we actually saw an example of this with Anthropic's famous vending machine set up where it could actually go out and buy things for the vending machine and have it installed. so that means that AI is now actually the one consuming, the one buying things. And and obviously it's not buying Coca-Cola, but it can be buying other kinds of products, licenses and and compute and all of that. And then on the other side of it, we're also seeing AI as a clear filter in between the consumer and the producer. So AI is you know, the people when they're making their buying decisions, if they need to buy you know a new product for for their company, a new Xerox machine they're going to go to AI first, and AI is going to say, Here's my recommendations. And the companies making the the copy machines are going to want to be at the top of that list. And so I think we're we're seeing two new fields that are about to come out or starting to come out. One is targeting your your marketing so that it hits the LLMs that are going to buy your product. So that's instead of humans buying your product, the LLMs are now buying it. And then the other side is targeting the LLMs in such a way, almost like SEO. where you want the LLMs to to find your content and surface it to an individual. And I was curious if these AI startups already exist. I talked to my co-founder David Graham at at Verify You, and he actually dug up a few. there's one called Merch Kit and that that's addressing this. And there's this whole new area called ACO, a Gentec Commerce Organization, which I I guess Zach: God. Marty Weiner: is like this whole new field where you are building entire companies now to sell to what amounts to a whole new creature. so anyway, yeah. Zach: Yeah, but do you know like What's interesting to me about this is like, you know, I haven't I haven't used AI to buy products often occasionally, I like just really don't know about something. I'm like, can you make a recommendation? It hasn't always been great. But like one of things that I like about AI is it'll do the stuff you wouldn't do. So like, if you go to Amazon, you're like, I want paint, you're gonna get like 40 drop ship companies, you don't know who's any good. Christopher: Mm. Zach: In principle, AI can go to like Consumer Reports or whatever. And so I guess it's not clear to me, like, you when you're marketing to humans, what you do is you kind of like half lie to humans about how when you buy this, you'll be sexy and popular, presumably an LLF. doesn't have that bias but presumably it does have other weird alien biases and so I don't know if that yeah but but but what are they what do you think god Marty Weiner: That's what I'm saying. yeah. Christopher: It has a it has a percovity to read white on white text, is the is the trick there. It's white on white text is the is what's hot right now. Zach: Well, but that's what I mean. Like, is this just evil stuff or is it like like one way maybe you Marty Weiner: No. Zach: could get an alum to buy the product is by having good information that's relevant to the query. And one way is evil. And I don't know what. Marty Weiner: I I don't know the answer to that except that it's definitely different than humans. Zach: Yeah. Marty Weiner: and and I wonder if Christopher: But that's Marty Weiner: it's almost Vulcan, you know, almost like logical. Christopher: But that but that's been that's been the you know, if you if you if you make the analogy with SEO, SEO Zach: Yeah. Christopher: always felt like it was that same kind of grasp bag of like there's a set of things which obviously lead to better search results when you do the SEO playbook. Zach: Yes. Christopher: And there's a set of things which are clearly just taking advantage of the algorithms that are used by the crawlers to try to optimize for the result. And so I'd be willing to wager that you're gonna find the same thing here. You're gonna find you're gonna find white hats and black hats, and the white hats are gonna try and make things better, and the black hats are gonna be like, Look, there's this like Point one percent increase. Marty Weiner: Yeah. Christopher: Right? It's like the vacuum up nickels problem. Zach: haaa... Marty Weiner: I mean that's exactly what marketers do, right? You're looking for that point one percent. If I use yellow instead of red, people are more likely to click on the button. and I think that's exactly Zach: If I say delve. Marty Weiner: what we're gonna see here. I mean, this is a very niche, super nerdy thing, but you'll see my Apple II behind me. The power supply broke on it. I went to Zach: Nice. Marty Weiner: ChatGPT, and ChatGPT gave me three or so sites, and I chose the one that would make me look the most sexy. So Zach: Yeah Marty Weiner: Yeah, but no it is it's funny. I I I use it for my my well, Christopher: As sexy as one can be holding an apple two. Zach: I will too, yeah. Marty Weiner: with a soldering iron too, just to add that. And Zach: Yeah, yeah. Marty Weiner: but no, I I I'm making my purchasing decisions now with with the AI and I don't know how it shows it. I mean it's it's a very limited set world that I'm in. But I did also make that decision recently when buying a new watch. So you know, I I wanted just kind of bring that up because I think it's a whole new field that if it's not exploding now, it's about to explode. Christopher: Makes total sense. Alright, move it moving on. Google publishes a paper on a new LLM architecture that could cut token usage by ninety four percent. Marty Weiner: Yeah, so you know, it's funny, I as I read this more, I'm starting to think it's not as exciting as I originally read it, the Zach: Ha ha ha! Marty Weiner: first time I read it. But so yeah, the idea is what Google proposes in this paper is the way that LLMs currently w work or any transformer is you take all the text that somebody just wrote and including original instructions from the LLM, and then you predict the next word. Then you add that next word To all the other stuff that you already have, and then you predict the next word. And you can imagine how this would explode very rapidly into just a huge number of tokens that are being consumed. the proposal here is: well, why don't we occasionally compact all the stuff that we built beforehand? And and I thought that was kind of neat, kind of interesting. It does destroy your the all the caching that's used with within the LLMs, but that's okay if you've turned one million tokens into 60k. Maybe you know that you're still winning. I kind of also realized while this is interesting, this is kind of what we're already doing with vibe coding. It's very common to have a session where you you update the code and then you tell Claude or whatever you're using to go make you an MD file so you can remember what the hell you were working on and communicate it with others. So I think the idea is kind of interesting that that they did see a result that dropped the token usage by as much as ninety-four percent. Zach: Yeah. Well, but so like, could that be exciting if you're serving inference, you know, if you're anthropic and you can drop the cost to serve people by like 90%, that's obviously a. Marty Weiner: Yeah, yeah, that could be a huge change. The the problem you have to be careful with is that when you compact the entire conversation, you compact it in a way that you know, aligns with what the person was trying to accomplish. 'Cause you've just Zach: Yeah. Cool. Marty Weiner: you've just changed the previous conversation. Christopher: No, Marty, dumb question, is this is this actually a new chip architecture or is this new a s a new software architecture? So this is is this deployable? Marty Weiner: This would be just No, this is just this is just software. So, you know, you you can kind of imagine if you took your entire Chat GPT conversation you've been having in in one thread for a week or so and you said, you know what, I want you to smash this into a more compact state, and you know, drop the irrelevant stuff, all that, then then this is what I would end up with. And I I am I am butchering the paper a little bit, but that's that's the way I I took away what they're trying to get Christopher: That's what we do I mean, I I think Zach: What? Christopher: I think butchering the paper a little bit would have been a good alternate title for this podcast. We could have actually gone with that one. what's we're I think it's it we Marty Weiner: w we're we're up to twenty great titles here, man. Zach: you Christopher: we find them as we we know, we might we we st it's still not too early to rebrand. We can figure that out. do a but Zach: Dude, do a butcher. Marty Weiner: Yeah. Do a butcher. Christopher: okay, so another news. NVIDIA buys Hugging Face for what I think what we would call stupid money of twelve point nine billion dollars. for those of you who don't know, Hugging Face is a it's a community site where all the all the highly reputable leaderboards for all the models live. It's also where researchers collaborate on the state-of-the-art models. and so Hugging Face being bought by NVIDIA, company that Builds hardware for LMs, buys company that peddles LMs. Seems like a good synergy for me. I'm not sure about you guys. Zach: I guess I don't Marty Weiner: I Zach: understand like, like a why this much money and why they want it assuming it's not like just to advertise their own product or is that it. Christopher: I mean I I do think that there's a bit of it, which is that Hugging Face has an opportunity to be the community product that is NVIDIA's community. I mean Nvidia it's is also pushing its own its own models in this process, right? And so it's it's good to be able to show that it's it it it has a it has a community organized around what it's building. I think it's also just good it's it it I think there's only s so many places that NVIDIA can put their money right now too. to some extent, like if you think about the major players that are running around, Hugging Face is a is a big surface. we've talked about it in previous weeks over the the hacks from the LLM vendors who who broke into Hugging Face in l in previous weeks. in this case I think it's just a good case of like they're in the same space, the price was right and off they go. Marty Weiner: Now I'm not I'm not entirely certain that this deal has been closed yet. I believe that it was proposed. and I'm curious if the DOJ will start to raise their eyebrows on this, although I don't think the D DOJ understands anything of what's going on. Christopher: You want you want to hug a face? Marty Weiner: Yeah, what what does that mean? Zach: Yeah. Christopher: I la last up on the frontier side we've got a as I mentioned before, Apple's been not doing as good a job on its anti leak prevention, and so the current leak is that Apple may be making an AI server. which I think by itself is interesting because I don't think Apple has made a server in a good twenty years at this point. Marty Weiner: They literally made like a server when I think macOS X came out and they promptly Christopher: Yeah. Marty Weiner: we never heard anything about it again. in fact I think we used to have jokes about it at at Pinterest, the theory being that Twitter had bought a bunch of Apple servers and and why the hell would they do it? But I think it was a malicious rumor because anybody who would spend an extra forty percent premium on servers from Apple just be a i idiotic. I d Zach: They have really cool casing Christopher: You don't didn't necessarily but you don't Zach: it'd be like sleek Marty Weiner: I I guess that's what it would be, but nobody would ever see them. but so I I don't know what's in these five servers yet. They're the it's just a a leaked rumor and I think Apple's trying to position themselves to be in the AI universe more. But they're definitely, you know, lagging behind a bit. Christopher: moving on to the world of robotics. Anthro Anthropic apparently has a standard for hardware now. Marty, you have something news on that? Marty Weiner: Yeah. I I'm getting kind of excited about where where these things are going. So what we're seeing more of is y the LMs can go think all day and night running software, but they have the limits that they they don't have eyes and ears on the ground. And you've got these companies wanting to say do research into biology, y some so you've got to have someone come along and, you know, squirt something on a slide and then put it under the microscope and then count something and then go to the next slide or do whatever your your experiments are gonna be. But the there is no interface for the LLMs to do that in a consistent way. So what Anthropic is proposing is a model standard for for hardware, so that all of these various kinds of equipments, cameras, incubators, everything, can be then hooked in so that Anthropic can run your experiments. So then Anthropic can go all day and night, you know, running your biology experiment. and we get that going, that that really opens up a whole new area of of rapid research. Christopher: So wait, Marty, does this mean my my current career of being the live button pusher for my LLMs is gonna be removed? Like what'll I do with my time? Marty Weiner: you know, that sounds like a startup idea. We could make a robot Christopher: Ha ha Zach: that presses just just press this tab you need one of those get a drinking bird Marty Weiner: Different Yeah. Christopher: Yeah. Chris, I'm gonna need you to go in the garage and pull the plug on the server and restart it for me. It's like no. Zach: Yeah Marty Weiner: Wait, what's the classic what's the classic line The Simpsons already did it? I think The Simpsons had an episode where the little drinking bird would hit Y for him. Christopher: It was like hit the hit the Y form, yeah, that's fair. It's totally fair. Zach: There's prior art. Alright nevermind. Marty Weiner: Prior art, yeah. Wait, wait, prior art just that doesn't mean we can't start a startup. In fact, I think Zach: right. Marty Weiner: most startups are started on prior art. Deep homage. Not not called theft, it's deep homage. Christopher: Some of the some of the best startups are are deep homages to technology that didn't quite land the first time. So it's stuff Zach: Alright. You Christopher: and other news, a new Chinese battery hit six hundred watts per watt hours per kilogram. Marty, what does that mean for what does Marty Weiner: Yeah. Christopher: it mean for drones? Let's start with drones, they're cool. Marty Weiner: Starbucks drawn. So I tried to come up with some numbers to help us understand what the is six hundred watt hours per kilogram, what does that even mean? so Christopher: How many how many horses how many horsepower per washing machines is that? Marty Weiner: I'll get you that number. I'll ask Zach: you Marty Weiner: the I'll ask AI to do it. Well let me let me give you some numbers to to give you a little bit of a s size estimate. So the Tesla c cars, some of them have what's called a Tesla four four six eighty battery, and those come in at two hundred and two hundred and forty-one watt hours per kilogram. So about two to three times smaller. I should tell you what this unit is. It's kind of a weird unit for for people who aren't nerdy in the same nerd way I am. watt hour is is energy. It's how much you can store and kilogram is of course you know the mass of the battery. So what we're talking about is more energy in a smaller amount of of mass. and and the more energy you can get, the the longer we can drive our cars, right? So if we had 600 watt hours per kilogram in my Tesla, I could go instead of I think my max limit now is like 360 miles, we're talking then being able to go like 720 to 800 miles on a charge. the challenge with this tech is it's not it's currently in the lab, but it's really interesting. and there's problems with it. Eighty percent capacity lost over just sixty cycles, so that's gotta be fixed. but if it is fixed, what it opens up is things like longer drone flights. It opens up electric motors for for aviation. this is where my understanding of the limits of aviation come in. but it has to do with gas engines have a lot of extra weight, right? and if you can start to get close to that weight limit where the the batteries are far more efficient in terms of weight usage, you suddenly open up electric motors for aviation. Gasoline, by the way, is far more dense, energy dense than than these batteries. We're talking twelve thousand to thirteen thousand watt hours per kilogram. So that's like, you know, over the Tesla, that's what is that, fifty times fi Christopher: Factor of twenty. Yeah, fift fifty over the Tesla and twenty over these new batteries. Marty Weiner: Right. but if you take into account all the extra mass you have to put into the engines with the transmission, with with the the engine, with heat dissipation, then it drops down to about twenty five hundred to five thousand, still a lot larger. but but apparently all you have to do is get to within like six hundred watt hours per kilogram and you start to open up aviation as a as a possible new space for electric motors. Christopher: Very neat. last last in robotics, and maybe we can actually get a we can get this get a video of this in post. Agbots are looking very futuristic and I would almost say Terminator-esque. Zach: Yeah. Marty Weiner: Yes. Look at these. my god, we're so screwed. Zach: What is it picking? couldn't tell. Is that cabbage? Marty Weiner: lettuce? I don't know. I've seen Zach: Lettuce? I don't know. Marty Weiner: a couple of these One is I'm seeing them now picking the actual food. but the other we've definitely seen especially with lasers, which is cool because lasers Zach: Yeah Marty Weiner: the ones where they zap the weeds. So you I d the AI can identify which plant is good and which plant is bad. and nuke the bad and pick the good. so it's it's really opening up I I think a potential for a massive increase in the density of food that could be produced from any one area. Zach: Well, and then using less pesticide and that sort of thing, ideally anyway. Marty Weiner: Using less pesticides. Exactly. Because you you know, if you could just walk around all day and zap bugs and zap the wrong plant from growing, you don't need herbicides or pesticides. And it also opens up we s I think last week we showed that there are now these AI controlled hydroponic labs, that could be multiple stories tall and can be, you know, doted upon by the AI twenty four hours a day. Zach: I'm a little skeptical of the cost benefit on the vertical lab stuff, but Marty Weiner: Yeah, but it's but it's definitely cool. Zach: it is objectively awesome. yeah, like Christopher: Ha ha ha. Zach: that is the beguiling part, but this is a big deal. there's a guy, I'm thinking on his name was proposing vertical farming for years and like had one of these engineering things where you're like, you see the cost all works out fine. And then they tried it it was like, no, it turns out having the sun for free is really nice. Marty Weiner: Do you now how old is that? Because I I've seen a lot of fields where the Zach: Yeah. Marty Weiner: LMs really tip over the cost analysis. Zach: It could be, I think my impression, I've only read a little in this, it kind of depends on what you do. So if you're like, well, we'll serve micro greens to yuppies inside New York City. It might be Marty Weiner: Mm-hmm. Zach: worth it because they're high value stuff and they're close. I think people tend to underestimate how cheap transportation is. know, Marty Weiner: Yeah. Zach: like, like, like so often it actually like is literally cheaper to ship like tomatoes from Argentina than to truck them from a US state for like complicated economic reasons. So I always worry about things that just look too cool via tech when there's like some giant like you're fighting the capitalist economy of scale forces. It's very hard to win. But on the other hand, as you said, like I really would like it to because it's just really cool. Marty Weiner: It's really cool. Yeah. Zach: but intuitively, it's kind of like suppose you want to grow a pumpkin like the most you can sell a nice. What does a jack-o'-lantern cost? Like eight bucks, which already feels pretty overpriced. It's like the Marty Weiner: Right. Zach: idea that you're going to grow that like in a lab bench and that the LLM is going to confer so much value that it's worth it. And, you know, you can make the pumpkin for a few dollars less seems I'm a little skeptical, I guess I would say. Christopher: well moving on, so let's talk about let's talk about culture, 'cause all of these things we've been talking about have huge implications for for how we how we operate. sounds like there's evidence for a shift in the way that LMs are influencing human communication. Zach: Yes, this was a I'm going to deep dive on this. is a really cool bit like one of these pairs where you read it. You're like, this is so obvious in retrospect. We should have done this. So what they basically do is they say, hey, look, know, technology, we think often shapes communications like books are not just a way to share information. People start talking a certain way in books, and that influences how people talk in public. Obviously, this stuff is hard to measure. But then they realized there's this period that will probably never come again from when GPT three and a half comes out to to win like other models. get big. So they say that they claim there's like a period of 18 months where basically everybody's using GBD three and a half. Like even though four comes out, most people use the default model, which is three and a half. And three and a half has quirks, right? So the way they say it is it prefers normative, socially desirable language, or they said something like that. And so we all remember this, like it would use words like Delve, Delve was seemed to be really overrepresented. There were others like showcase and intricacies, these kind of banal showman type words. And And And so what they did to actually make sure that like the human intuition that Delve was quite common, what they would do is give AI a human made essay and say, could you polish this? And then it would redo it. And then they would check the difference and they would find which words were overrepresented. And so I don't know if we could throw up some of these graphs, but indeed Delve was really overrepresented. And so they were like, okay, is there a way to check whether this is influencing how humans talk? And so first they looked at scientific papers and indeed they found scientific papers were suddenly saying like delve and these other words more Marty Weiner: Yeah. Zach: but then they're like now maybe it's just AI just wrote the papers so we're all we're finding is AI talks like AI and they did a similar thing where they looked at like YouTube videos of scientific conferences I don't know why they wanted scientific stuff but it's just nerd stuff and they found the same effect and then they're like now maybe everyone's having AI write their scripts and so they had this cool idea which was well podcasts often have a lot of spontaneous language in them and so they trained a classifier to pull apart podcasts into the spontaneous portions and the scripted portions. And then they looked at the word frequency in the spontaneous portions. And what they found is over time during this 18 month period, you could see, and there's a really cool graph of this, that the usage of what you might call like AI quirk words went up really significantly. So I think in the case of Delve, which did really turn out to be highly overrepresented, I think at peak, was something like 50 % higher users than you would have expected without people talking to LLMs. And another finding too, like podcasts that were about science, education and business did this a lot. Sports barely affected. I think the jocks were not using LLMs. They were they were not being ruined by LLM language. And then the last thing to say about that. is that one of the other really cool things is once GPT-4.0 comes out, the opposite happens and it starts going down. And there's an interesting actual question as to why, because 4.0 still overrepresented Delve, it represented by less. So one explanation is that AI used it less, so we started using it less. Another is there were other models by then. So we were just, know, diffusing into other quirks. Another thing is that the Delve thing became viral. And so it started to stink of AI, if you saw someone saying Delve. fascinating thing is that with a lot of these words that the level it got to afterwards was actually lower than the initial expected baseline. So it appears we actually say Delve less than we would have if AI had never existed now. And so the really cool thing about this to me as a broader cultural matter, was an article I wanted to talk about a few podcasts ago that we didn't get to where a guy was saying like, This is a literature guy, not a tech guy. And he said, you know, we're going to be divided into people who use AI, people who don't use AI and people who kind of blend. But the thing is, even if you don't want to use AI, you will have to react to it because you might have quirks and stuff that are AI like. And so you're going to have to move away to them to get your value added as a human. know, I think we also Hank Green got in trouble because he used the phrase, spontaneously used the phrase, thank you for the pushback, which is a very AI sounding phrase. Marty Weiner: Yeah. Zach: And so it's like, what I wonder if now we're going to have to say like not use phrases like thank you for the pushback or not say it's not just wisdom it's insight you know we can't say Marty Weiner: Yeah. Zach: that anymore because AI says it so even if you want to opt out you can't opt out because AI is such a powerful force in culture there was more but it's just just a really cool very tight paper on how AI is affecting spontaneous communication Christopher: I think the Delve example is a good is a good representation of a word that is like the it's like the physical it's like the word embodiment of the dash. Like I mean I Zach: Yeah. Christopher: think for the like speaking as somebody who liked the dash before the dash was uncool, Zach: Hahaha! Christopher: the you know the LOMs LOMs do actually have the advantage of they can actually use the entire grammar and the entire space of words and they can use them effectively because they understand all of them. And so you've just seen like much more varied punctuation in vocabulary. And the side effect of that is that that's where the that's where That's kind of where the taint comes from. Like if it becomes unfashionable, all of a sudden you become aware that I mean, I've I've seen examples of people complaining that you know, how how is it that like Emily Dickinson was using was using AI and they'll like cite an example of her using like an dash. It's like, well, you know, the dash Zach: Yeah, yeah. Christopher: existed before before GPT, believe it or not. And same thing with Zach: Yeah. Christopher: Delve. It's like Del like it might be new to you, but that doesn't mean the word is new. Zach: No, no, we have this in the arts now we're like now and then there'll be an artist. I mean, there are a lot of artists who are faking it. They are using AI and pretending they can do stuff and getting in trouble for it. There have been there are a few people whose style just happens to match like the open AI house style. It's a kind of sweet cartoony style who get in trouble now. And there's like there's nothing they could do. You could say like I've been doing this for 20 years, but it doesn't matter because now your style has the stink of AI. So you have to change if you don't want that. Christopher: Well, I do like the fact that one of the one of the words looks like it's going through a cycle is groundbreaking. Groundbreaking groundbreaking Zach: Hahaha Christopher: and innovative are the ones that are currently crashing. Zach: Yes, yes, innovation and groundbreaking are out. Although do you see hinges are still on the up hinges by hinges? It's going places. Marty Weiner: Hinges? What what is wha how do you use hinges in a phrase? What what do you mean? Zach: Well, you know, your brilliant insight really hinges on yeah. Christopher: That argument hinges on blah blah blah, yeah. Marty Weiner: okay, okay, okay, okay. Zach: Yeah, not like the object. I don't know that there's. Marty Weiner: yeah, okay. Yeah. I was getting really excited gate hinges is up. Christopher: I mean, actually one one interesting side effect here will be the you'll remember from a couple weeks ago, Anthropic announced that they have to start watermarking they're gonna start watermarking all the things. I was digging into this in in the last week to figure out like, well, how do you exactly watermark text? and the answer is actually an interesting one. So the way what they decided to do is they have gone through and they've I'm kind of paraphrasing the the way it works, but ultimately they've gone through and they just kind of identified words and categorized them into sort of like Red or green, right? So like and it and for words that are effectively equivalent, right? And they f they're gonna fit into one or two buckets. And so what they're able to do is even though they can't inject you know, like special characters or signatures in the in the text, what they can do is they can effectively create patterns in the text of this token is in bucket one in red red bucket, and this token's in bucket two green bucket. And then they can basically create patterns out of the tokens inside the text in a way that's opaque to the end user. which you have to imagine also creates a lot of opportunity for this kind of this kind of shift, right? Where you start seeing words that are somewhat less used appear because they're a synonym of words very common because they needed a red token instead of a green token. Zach: my God. Or Marty Weiner: Okay, but Zach: like, we'll all start doing the pattern and not realize it and then the one row won't work. Christopher: Yeah. Marty Weiner: That's what I'm I was exactly what I was about to say. That I mean this very paper is kind of saying that th those watermarks are gonna be short lived if we as a culture Zach: Bye. Marty Weiner: move toward doing what they're trying to watermark against. Zach: Fascinating. Christopher: That's fascinating. well, you know, s speaking of speaking of high brow, apparently Zach: Yeah. Christopher: you can we can now generate AI Rick and Morty episodes in faster than real time. Marty Weiner: yes. Zach: Heh. Christopher: Our long nightmare of waiting for the next season is over. Zach: you Christopher: the next season can be continuous and streaming. Zach: You Marty Weiner: Yeah, I think this is really wild stuff. And I think this is the this is almost like a a tipping point, I feel like, in in the way we consume entertainment. So the idea is that these guys are showing that you can use Minimax H3 Max to run faster than real time. The equipment they use is quite expensive. I I was hard having a hard time finding exactly what they're using, but one person was suggesting they use a couple of NVIDIA RTX A6000s, four of them maybe at forty eight hundred dollars a pop. So that's quite pricey and I'm sure the energy cost is not nothing, but it's at a point where we can literally generate have the AI start to generate movies. I'm I'm sure the scripts suck and the jokes are completely awful. but Zach: you Marty Weiner: but the video can be generated, the audio can be generated, and it can be generated live with what you're watching. So it's sort of a first step into you know, being able to have as much signfold you want, as much brick and morty. I think the part that's more interesting to me than than that is how that will change what we consider entertainment. So, you know, for us ultra, ultra nerds, you know, you can imagine wanting to just be a fly on the wall, say like, I I I saw one guy making movies before AI that was just what would it be like if you're just on, you know, the Star Trek Enterprise, right? Just just just enjoying soaking it all up. And that's, you know, obviously hard to build all the sets and have the actors do all the things, but now you could you know, generate that in a fly and just enjoy it in any way you want to enjoy the the the the entertainment. So I think it's gonna open up a whole Christopher: I mean this seems like a Marty Weiner: new thing. Christopher: this seems like a great opportunity for also for gaming if if held properly, right? Like for all the Marty Weiner: yeah. Christopher: for all the complaining about for all the complaining about you know, we can't have any of that any of that dirty AI in our games, if you could actually have if you can do if you can do this render faster than real time, you could actually have the the the viewer become one of the characters, right? Just react to what to whatever person's saying. Marty Weiner: So y yeah, there's there's a couple of variants of that coming out. One I've seen in the last few days is sort of like the game from the eighties called Dragonlayer. Dragonlayer was interestingly different from every other game. You know, every game in the eighties, like at the arcade, would look like Pac-Man, pixelated, very limited. But Dragon Lair Dragon's Lair was a cartoon that you got to control. The the thing was definitely there to take away your quarters, but it was neat because you'd actually have full animation, and then you had to push your, you know, the joystick at the right time and you would go to the next thing. people are now making that, but you have total control over it, so that y you know, your your little figure walks up to the castle. do you wanna go in the castle? Do you not want to do it? And then it just makes the AI for the next segment and you can just follow along with it. And then of course we are also seeing some really wild demos coming out of you know, Minecraft generated out of the hallucination of the AI video where you push forward and it says, Well, what would be the next frame? And you move forward. so I think I this is again goes back to my prediction that we will have Grand Theft Auto seven before Grand Theft Auto six. Do a video game. Zach: I mean, do you think, Christopher: well there Zach: know, one of the things that's culturally weird about this to me is like, I feel like one of the big discoveries that studios have made in the past 25 years is like catering to the fandom. And, and I think part of what's going on is like, you have people who just want to be part of a community. And it's almost like, like if the new Star Wars thing isn't that great, that's kind of okay, because we're all just part of this community together. And if Marty Weiner: Right. Zach: like, if our, everybody's having a different perpetual Seinfeld or whatever, you lose that thing where you get to talk about. Marty Weiner: Yeah. Yeah, that connection where you're all together and you talk about it over the water cooler. I've wondered about that too. 'cause I I even though even before AI we've we've had an explosion in the number of shows that you can Zach: Yeah. Marty Weiner: watch on streaming, it does seem like communities kind of go around certain ones. So I'm really curious to see if we have this tech, what will the community do? Will you end up with, hey, here my favorite YouTuber put out his latest Rick and Morty and my you know, my whole office watches it. So I I wonder Zach: Ha. Marty Weiner: if we will still clump around the fireplace in the same way. Christopher: well in in in gaming news, there does seem to be a a new Unreal Engine release that's come out. Marty Weiner: Yeah. So five point eight recently came out a few months ago and their big change was to add in, of course, all the things you need to AI to make games with AI. And so they added in these various MCPs, and it is pretty impressive what you can do with it. The newest version is really stepping into that. So they allow you to I I mean the the whole craft of making video games must be changing in every single studio out there because this thing just allows you to say, Hey I want an entire field of trees, they handle all of it. I tried doing this a little bit a few months ago before, you know, a lot of these nice tools were out, and it was really clunky. I did this with Roadblock Studio and it would take a screenshot and I would say I want trees and it kinda worked, but it was it was bordering on not useful. but now it's looking like they've really broken through that so that you actually have sort of like a buddy next to you building this game with you, and it's able to see the screen in the same way that you're able to see it and interact in the same way. So I think we're gonna see an explosion now in really excellent indie games. And actually I've already started seeing that. You're now seeing a lot of stuff spread onto Steam of really good quality, indie games. Christopher: There's also been there's been a shift in sentiment around DLSS I've seen recently too, where for DLSS Zach: Yeah. Christopher: was the deep learning super sampling. I had to look up the acronym. this is this was this came out a couple of about a year ago. originally it was an attempt to do AI overlay to do remastering of old games, and it was just it was just trashed in the media. the latest versions of it seem like the version five just came out, and it's now at the point where it's like, Now they're on Cany Valley because the movement's all jerky, but the faces are really good. You know, it's like the the models are the models are starting to show that they're models, but the the textures are now being generated by AI, which is a a really interesting shift in this. Marty Weiner: Yeah. Christopher: All right, moving on moving along to let's get into let's talk about politics. so apparently NVIDIA has decided to start its own pack. which I guess it's that we're at that point in the cycle. Zach: Yes. Marty Weiner: Huh huh. Christopher: I th I br the for those playing bingo, we're now gonna use the word regulatory capture, I believe, is the next word that kinda comes Zach: Hahaha Christopher: up here typically. unclear exactly what the pack is going to try to solve. Zach: Yeah. Marty Weiner: Well, I I I actually was a l I I will say I'm a little surprised I didn't see this before. it seems like a company worth about five trillion dollars and with the influence it has on the world would want to have some heavy influence into l yeah, lobbyist Christopher: Lobbyists. Lobbyists. You gotta spend money on something. Marty Weiner: and then the next level up, like we are going to raise money for the candidates we want to be in in government. Zach: Yeah, but I Christopher: and Zach: mean, it does seem like they have political desires like to I assume they'd love to sell to China. So if Marty Weiner: yeah. Yeah. Zach: there was a pro free trade with China candidate, I assume that's who they'd back. Marty Weiner: Well w where's the pro build the hell out of the data centers, you know, drill baby drill kind of candidate. Zach: Right, RIPPRO data centers, right, right, right. So, I mean, it is clear that they have political desires. There's also like, you know, I don't know what Huang's position is, but he seems pretty gung-ho in every direction. And I know like, you clearly the frontier models are getting some kind of regulation now, and there's probably gonna be more. And so I can imagine them wanting to push against that. So, clearly they do have political desires. don't think we know yet what their PAC is going to actually do though. Christopher: And and and we are seeing that we're seeing that with with I mean China's actually catching up in high tech exports, right? I mean like China China apparently exports surge forty one percent. turns out when you when you when you restrict access to technology to a large manufacturing c country, they put their effort into man manufacturing things they can't get access to. so definitely a big shift over there. potentially in another hey, maybe that's where ramp prices will go down if they're allowed to if we're allowed to import them. though yeah. Marty Weiner: Absolutely. Yeah. Christopher: in other news, the CEO of Marty Weiner: Mm-hmm. Christopher: Flock Cameras says that Americans need to compromise on privacy. Marty Weiner: You know, there's th Zach: my god. Marty Weiner: there's certain things you don't say out loud. I you know Zach: Yeah, I was gonna say. Marty Weiner: you know, but yeah Zach: my God. Wait, wait, Could Christopher: Yeah, this one Zach: you describe what Flock is for maybe people who don't know? Christopher: for those so those of you Flock has been in the news recently. So Flock is a they're a network of cameras that many, many municipalities general generally targeting sort of small towns and cities. they're kind of like the we'll call them traffic cameras on steroids. So effectively they are pull-mounted cameras that are do license break tracking, they're potentially doing face tracking, and they're being they're being used to basically cover all the way from speed traps through to looking for burglaries and fighting crime. Zach: . Christopher: but of course the flip side of it is it's a persistent always on camera that's probably streaming directly to to a number of of data mining companies. unreally c kind of unclear where the data's going. And the side effect has been has been a fairly large backlash of people cutting down flock cameras, burning them. you can imagine this is why the CEO thought he had to kind of get in front of this by saying, Y'all just need to chill. I was like Yeah. Zach: You will have no privacy go home yeah, yeah Marty Weiner: Just be okay with that. Why yeah. D do what would that be called? a do yeah, do it there you go. Yeah, do it Yeah. Great. Christopher: Do an do an Orwell. Yeah. Zach: Do an acquiescence. Christopher: let's let's let's let's j let's jump into weird stuff. so there apparently was a there's a a new AI video, which maybe we can actually get on, where somebody Marty Weiner: yeah. Christopher: somebody overlaid so if if if you haven't seen it recently, China has been running the new humanoid the sorry, the robot games effectively. there's been some great videos of some herky jerky but extremely fast robots running down the track field. somebody went and overlaid human models on top of these on top Zach: You Christopher: of these humanoids and it's it's hilarious. I'll just say that. Zach: Yes. Marty Weiner: This is what the best part. Yeah, when they all smash into the wall and then explode. Zach: Yeah. Marty Weiner: Yeah. Anyway, fantastic. Zach: Fantastic. Christopher: Fantastic. and of course we can't we can't let a good story go to waste. so we we we started the we started the the the hour talking about how open AI has decided to not use cursor anymore or not support cursor anymore. And of course Elon is going to fire back. Zach, would you like to take Zach: Yeah, but Christopher: this one? Zach: as we got hung up on this, so what he actually said was, Marty Weiner: Yeah. Zach: scam Altman, and Marty Weiner: Mm. Zach: Greg, Greg Stockman. So his name is Greg Brockman. And Elon said stock man are utterly untrustworthy a holes, etc, etc. But what happened is what is is insulting about Greg Stockman? Because his name is Greg Brockman. Marty Weiner: I I don't know, i i is he portly, like he's saying he's stocky? He's not. I don't know. wha what's the insult? I Zach: stocky? That's a reach. don't know. Like, he is like Christopher: He only he only cares about the stock? Zach: he owns stock. Yeah. Is it right? Marty Weiner: What? Zach: So I was trying to make up better ones. So you Marty Weiner: Yeah. Zach: know, so I came up with Greg, Crockman. Greg Schlockman, Marty Weiner: Yeah, like it Zach: like as a crock I don't know. Greg Snotman was I thought pretty clever. The Marty Weiner: that that's good. Yeah. Zach: best I actually think what he should have done and I'm you know, Maybe he should listen. I probably listen to this podcast. He should have gone with Greg Brockscam. Because Marty Weiner: Yeah, that's that's way better. Zach: then the problem is you have Sam Scamaltman and Greg Brockscam, which feels a little lazy. Marty Weiner: okay. Zach: So you change Sam. I think we decided I was talking to Kelly about this last day. think it should have been Sam Faultman, like a man who has faults and Marty Weiner: Yeah. Zach: then Greg Brockscam. I'm just saying the guy has like eight billion employees and his own personally trained AI. He could do better. Marty Weiner: Yeah, i i i he's really he's he's really not up to what I would consider presidential level name calling. Name pod yeah. Yeah, yeah. He's got a lot more practice in petty Christopher: Do an do an insult. Groc, do an insult. Zach: it's no literally that's right should have called should have called the president yeah Christopher: Name punts, name puns. Zach: yeah that's saying you can't make it too clever you it has to be only Marty Weiner: No. Zach: as clever as scam altman Christopher: Ugh Well, in other news, so we're moving on we're gonna we're gonna end this time on on math and science. apparently the first AI assisted surgery is in the works. this is a and in fact I guess they're starting ambitious where it's an AI assisted brain surgery to have a tumor removed. very interesting set of results here. I mean I guess the the the hard part with typically with with with brain surgery is finding the tumor where the tumor ends and where the brain begins is a very complicated problem and they're deploying AI techniques to make that process a lot easier for surgeons, which is Marty Weiner: Yeah, apparently there's a second screen live video feed that is then handed over to essentially a s what they're calling quote unquote surgical chat GPT. I don't know what the actual underlying model is, and I think they then use that as another tool to watch for things as the surgeon's going along. So s like a second set of eyes. Zach: Did you guys see there was a video of like AI to draw blood and I just can't. just I can't do it. Marty Weiner: God. Yeah, like it's coming straight at you. Zach: I don't know. amount of evidence will get me to put Marty Weiner: Ha ha. Zach: my arm in that thing. Forget about it. Marty Weiner: Yeah, that was terrifying. But apparently it's really good. Which is way but but but you know, that needle coming Zach: no, no. Marty Weiner: straight at you and the one chance in ten thousand that you're the one they're like that they're Zach: Yeah, no, I want a disgruntled nursing student. I'm sticking with that. Marty Weiner: Yeah. yeah. Christopher: In in in non in non-AI news, there's been a I would I would call this a breakthrough, a breakthrough in in location tracking without GPS. and so the interesting the interesting model here is that it's actually it's a it's a quantum navigation system. So it's a super super advanced system for detecting very precise measurements of magnetic and gravitational fields. and so the trick here is If you you're they've actually over the years been able to map the earth and basically cr basically create a map of at every point on the earth what is the local magnetic field and the local gravitational field. Because of course there's fluctuations in both over the course the globe. There's pockets of ores, there's going to be different densities of the crust. And they're able to use that to f basically find your spot because of knowing exactly what the local what the local fields are. so very Marty Weiner: Right. Christopher: very I guess differently hard to jam. It's a side effect here. Zach: Ha! Marty Weiner: Yeah, yeah. And you know, I where whereas GPS is accurate to within like what, a foot or so, this is one nautical mile range. So it's not tiny, but they the with things happening which with GPS like getting jammed, apparently hundreds of thousands of times let me actually got a number here. More than nine hundred and seventy eight thousand GPS jamming and spoofing incidents recorded globally in the first quarter of twenty twenty six. so you know, it's becoming a very real problem. And a mile might be plenty. You know, as if you're out in the open sea, you just need to know where you are within a mile, you're good to go. And this is almost surely unjammable, at least with any technology we currently have. Christopher: Okay, and last but not least, we would we would not be doing our jobs on this podcast if we didn't mention another AirDosh problem that has been affected by LLMs. Marty Weiner: Yeah. Christopher: apparently ChatGPT improves the improves a a a bound on the gap between primes. this is a another one of those we're we are collectively not smart enough to basically speak to this one well. but fortunately, Our guest coming up, we're we're gonna end with a special guest today where we have math professor Jonathan Wise. gonna do a quick we're gonna do a quick cut through the magic of Marty having to do a bunch of work after this, and we're gonna come back with with Jonathan.