AI Fire Hose #2 - AI's Role in Our Future: Tesla, DoorDash, GPT & Linguistics

0:00If you've been considering futures trading, now might be the time to take a closer look. The futures market has seen increased activity recently, and Plus 500 Futures offers a straightforward entry point. The platform provides access to major instruments, including the S &P 500, NASDAQ, Bitcoin, natural gas, and other key markets across equity indices, energy, metals, forex, and crypto. Their interface is designed for accessibility. You can monitor and execute trades from your phone with a$100 minimum deposit. Once your account is open, potential trades can be executed in two clicks. For those who prefer to practice first, Plus500 offers an unlimited demo account with full charting and analytical tools.

0:45No risk involves while you familiarize yourself with the platform. The company has been operating in the trading space for over 20 years. Download the Plus500 app. Trading in futures involves the risk of loss. It is not suitable for everyone. Not all applicants will qualify. Hey, everyone. If you like this podcast, go behind the paywall to get privileged access to the smartest minds in finance. Visit realvision.com slash RVpod and use the promo code podcast10. That's podcast10 to get 10 % off our essential membership for the first year. Join the Real Vision community and learn how to become a better investor.

1:24And now to the top analysis of today's markets. What's going on, guys? Welcome to Real Vision's weekly AI firehose, where we drink from the pounding stream of everything happening this week in the expanding field of AI. I'm Ash Bennington, joined as always by our resident AI expert, AI developer, Mikhail Voloshin. Mikhail, welcome. Great to be here as always, Ash. We've got a really exciting episode with us today. We've got Tesla being sued. We've got DoorDash using AI for taking voice orders. And I'm going to play with some stuffed animals. Mikhail fantastic let's jump right in I know there's some big stories out this week in this field um yeah the by the way that's not an AI cat right it is not an AI cat as far as I know I've never looked inside but I I don't think any of us would enjoy that that looks like a real tale is however uh very eager to examine my new media rig and put her scent all over it which will mess up my camera angles so she's just gonna stay right here for a second all right yeah let's jump in and start out with tesla because this is the story that caught my attention so um tesla is slated to appear in court this uh this month and the beginning of next month uh for a couple of driver fatalities uh that occurred in 2019 uh one of the incidents involved uh tesla's self-driving autopilot driven car driving straight off of a highway leading to a crash that killed the driver and seriously injured two passengers.

3:01I know what a great note to start an episode on. But this mirrors what a lot of people feel about this space, which is excitement, enthusiasm, but also tremendous concern. And when you hear a story like that, it's very visceral. It hits home. You can see, obviously, some of the downsides, some of the risks of this technology. Absolutely. And you know, Tesla is denying liability, which is not surprising. Their claim is, quote, there are no self-driving cars on the road today. That's the official company line. And what they mean by that is that this is still prototype technology. The cars are not marketed as really being self-driving.

3:43Self-driving is just a term that we use to describe the cars. But Tesla insists that you always keep your hands on the wheel and that the driver still needs to intervene in case the machine is about to make a mistake. That's their sort of. I mean, basically what Tesla is saying, their defense here, and this is just my interpretation of it, is that this is kind of just a beta test, right? That they're saying that this isn't really full self-driving car technology. You know, it's, I'm sure, very material for this case in terms of their legal defense. defense but kind of in another view it also sidesteps the core question about where we're going to be uh in whatever the number of years is whether it's one whether it's three whether it's five whether it's ten where we do have these full self-driving cars you're not going to be able to make that defense and then this question about who is liable how it works what the rates of fatality are i mean i've i've gone into this story quite a bit myself because i think it's just so uh well terrible, obviously, on the one hand, in terms of the human tragedy, but also just such a harbinger of all the things to come.

4:42Like, here's an example of a question that I came up with thinking about this, Mikhail. If you have a fatality rate that declines by 90 % for self-driving cars, right, but the 10 % of errors are errors that a human being would not have made, for example, a car that drives itself off a cliff for no apparent reason, how do you assess liability in that? I Is it the entity that owned the code? I mean, it just gets really weird, really murky, and really just interesting from the perspective of trying to assign liability and advance these technologies. So you're absolutely right that the rate actually has been going down.

5:24And my understanding, and this understanding is backed up by some studies, in particular, some driverless car statistics. there's on a mile by mile basis Tesla cars are actually or self-driving cars in general are actually much safer than humans and that's a counterintuitive statistic one of them one of those one of the contributing factors to that statistic is that self-driving mode is usually engaged on the highway where you can score the highest number of miles with the least amount of stuff happening um highway driving is relatively simple compared to urban driving uh the other consideration uh is that you know people often forget uh two things one is that when you think about whether or not you every human is going to assess themselves to be a safer driver than than an automated you know than a machine right but that's only when you're thinking about yourself driving.

6:22You're not thinking about the moments of when you're behind the wheel when you're not driving, which is, you know, when you space out to think about something at work, when like, you know, you get that little blip on your phone. And like, even if you don't check the message, like that grabs your attention for a second, that kind of stuff, right? You're fiddling with the radio. I mean, it could be anything. The other thing that a lot of people don't consider is that the car literally has senses that you don't um you know you have sight and you know and hearing that may be superior to the car in many ways but you don't have lidar um i mean not yet um uh and if you do please leave a comment below um you don't uh you know the human body has uh you know it has uh the ability to detect acceleration but we don't have the ability to detect absolute road speed which the you know the speedometer does.

7:11So we can feed that information back to the driver. But right now it's being integrated into the self-driving systems of these cars. And so the car is literally operating on a different set of inputs that you are. And that's the reason for the enhanced safety, what you're saying is this idea that you get a broader sort of spectrum of data than human beings are able to gather through their five senses. But here's the paradox and the kind of that human beings are not very well suited to cope with, which is this idea. If you have a 90 % improvement in safety, you know, there is this joke in medicine that things, the odds of something happening are either 100 % or 0 % either happens to you or it doesn't happen to you.

7:51And, you know, someone says, well, listen, the fatality rate has been reduced by 95%. That's really cold comfort to you. If it's your loved one, your, you know, mother, father, child, brother, sister who gets killed in the accident involving the self-driving car that malfunctions. And these are just really difficult problems for fallible human beings like ourselves to get our heads around. You know, the thing that really gets me about the ethics of self-driving cars is that it is a literal implementation of the trolley problem. Tell people what the trolley problem is for those you may not know so it's an ethics problem um i want to say composed by john rawls in like the i want to say around the uh early 60s or so but the idea is that there's a trolley that is well it's a runaway train that's heading towards a that's heading on a track and there's a person that's tied to the railroad track now there's a switch on the track that allows the trolley to go off to a different direction however on that second track there is also another person tied to that track because apparently the mustachioed mastermind villain has really been getting around now the the ethics problem is that you are standing at the switch of the track and you can choose to either make the trolley go one way or the other way um there's a lot of variations about like maybe you know what if there's five people on the on the track what if there's like two only two people but like one is a baby and the other is like really really old like you know what if uh one is some you know evil villain and the other is like some average joe that kind of stuff right so the um uh the ethics problem comes into the fact it comes into play when determining like uh whether or not the act of enter of intervention in and of itself uh like it warrants some kind of moral culpability so um there's another permutation of the problem that's particularly sticky which is uh you have a track that splits left and right there's one person tied uh to each side and then if you do nothing if the train keeps going straight it kills 100 people so basically it forces you away from this idea of uh kind of uh benign neglect or not uh making a decision so you have to choose and these are exactly as you pointed out i think this is real life trolley problems that people have to code today not in some ethics class uh not in a philosophical discussion but they have to make these decisions about how you know when you have a car that's lost control the brakes are failing uh do you slam into another car do you slam into an area where there might be a pedestrian so the um the ai that drives uh teslas is a hybrid neural network slash heuristic evaluation system, which basically means that it uses neural networks for vision identification of objects.

10:48Like basically it uses a neural network to be like, that's a stop sign. That's a child. That's another car. That kind of stuff, right? But then having established like these are these objects at these locations, it uses hard-coded heuristics with a scoring system to determine what action to take next. And this gets into game players and basically tries to maximize its good score, or minimize its cost, or however it's formulated. But the point is that it literally has a score for this action. I am currently in this situation. If I take this action, it would lead to this outcome, which has this badness rating.

11:33If I take this other action, it'll lead to this other outcome, which has this other badness rating. which means that it literally has to decide, like, if the car's brakes have failed, and it can choose to either run over a pedestrian in the intersection versus slam into a nearby tree and kill the driver. There's actually hard-coded heuristics that Tesla's programmers had to sit down and evaluate in order to put numerical scores on the quality of one outcome over the other. We're going to take a quick break and be right back with more of the day's top analysis on the Real Vision daily briefing.

12:11And the key there that you just said is that the developers have to make those decisions for the waiting. This isn't something that machines are deciding for themselves. And that's why it's an intractable ethical problem. I think something that you mentioned earlier was you said, do you really want to be the last guy on the road who's driving his own vehicle? That's something that I think about a lot, which is where as time goes on and more and more vehicles are self-driving, are you going to be at an advantage or a disadvantage if you're the last human standing? On the one hand, you'd think you can weave and swerve around all of the other vehicles on the road because you're willing to break rules that they aren't.

12:53But on the other hand, does that increase your safety rating or not? Can you really say that you're a better driver if that's the way that you behave on the road. We may reach a point where we're going to have to explain to our grandchildren what a DWI was and why it was something that society looked down upon so much. You literally may reach a point where people can't understand that human beings used to operate these vehicles. So I do want to interject with one other point from abstract philosophy, which was, I believe, it was a Socratic dialogue. I'm not sure if it was original to Socrates or if it came from Plato, but it presented a thought problem that goes like this.

13:38Imagine you're in some kind of personal danger for a little while, like there's some stalker that's chasing after you, and you borrow a weapon from a friend. um the danger you have that weapon around for a while uh the danger passes the uh and it's time to return this thing that you borrowed uh the in the time that you had been borrowing this weapon your friend whom you borrowed it from has gone stark raving mad uh and is now liable to hurt someone else with this weapon so the question becomes is the uh in more modern uh philosophical terms, is which Kantian categorical imperative wins out, the obligation to return things that you borrowed or the obligation to not give weapons to crazy people.

14:30And I love this analogy for or this thought experiment for self-driving cars, because it's not entirely clear to me whether the stark raving mad entity is the car itself, the human driver or the programmer. Well, it's always humbling whenever you have a conversation about ethics to realize that Plato and Socrates got there first and maybe Kant did the elegant framing in a more modern sense. Listen, we've got a lot of stories that we want to talk about here. Another one that I'm really interested in getting to is DoorDash, the idea of DoorDash using automated AI to help process some of inbound ordering.

15:08Talk a little bit about this. It's an interesting one. so this is a great one from the point of view of talking about ai taking our jobs and ultimately whether or not that's necessarily a bad thing uh the uh the system that doordash has just rolled out and doordash of course is a food delivery system uh for those who don't know actually ash have you ever used it um it's uh i've only used github i understand it's similar grubhub sorry sorry github um see this is what happens when you have a programmer on the show right it is not a code repository it is a food repository mcow i mean in the to pair to misquote uh erasmus when i have a little bit of money i use it on code and then if any is left over i spend it on food it's a perfect freudian slip um but anyway doordash uh is a food delivery system where you order online from nearby restaurants and they put a little brown paper bag together for you and then a driver comes by picks up the bag and drives it to your house.

16:17Dimitri by the way as a I mean by the way as a Manhattan bachelor do you think I have not used this service extensively? I'm just useless in the kitchen so without Grubhub I'd starve to death. There's a reason why I asked you. And, you know, when I was living in Manhattan, I depended on food delivery services just a little bit. I'm just, you know, I made use just a bit. So anyway, they found that about 20 % of people can't stand ordering through the little menu system on the phone. About 20 % of DoorDash customers prefer to just call the restaurant and put in their order. And this is particularly painful during the really highly slammed hours from between about like five to eight or so.

17:08It's a classic example of what economists call a peak flow problem. And they're very difficult to staff for it because you've got the pig through Python effect. Exactly. It doesn't make sense for restaurant managers to hire extra people for necessarily in all cases. Sometimes it does, but often it does not. unfortunately what this means is that these restaurants are shorthanded for taking calls during these peak hours and they found that upwards of 50 percent of phone calls are dropped which represents if you do the math that's 10 that's a 10 percent uh total revenue stream to these restaurants that's just getting a not that's just not manifesting it's a big hit it's just falling out of the sky for no reason because they can't exactly staff this yeah so uh they're turning this over to AI.

17:57They've got an AI operator that answers the phone for the restaurant and says, you know, thank you for calling McHale's McBurgers. How can I, you know, whatever. And the customer describes what they want to buy in plain English and the AI converts that into a set of orders for the cooks and the rest of the order goes as if it would have had the person ordered by the phone app. Yeah, it's really interesting because you can start to think about some very cool implementations on this stuff. Like for example, Grubhub has all sorts of different weightings that you can sort by. I want to optimize for delivery time.

18:40I'm really hungry. I want to get the best top rated hamburger in New York City. I don't care if it takes an hour to get here. So you can just imagine having this kind of almost colloquial conversation with an agent. Listen, I'm starving. I need a hamburger. I know it's three o 'clock in the morning. Just get me the fastest possible burger. I don't care if it's gourmet or not. Bring it to my apartment as soon as you can get here. From a developer standpoint, I'm going to be keeping my eye on the on interviews or like customer satisfaction surveys that determine whether or not the people even know that they were talking to a bot.

19:14You think the technology has reached that point, Mikhail, where people would not necessarily know that they were talking to a robot? It's a loaded question because a person who is slammed and taking orders for the restaurant is not always functioning at their full cognitive capacity. So the question is - You made this point last week where you kind of made this distinction between kind of almost the highest possible rate versus the actual rate that people have when they're when they're doing a job meaning you know if a human being you told them dedicate all of your cognitive bandwidth to getting this absolutely right might they perform higher sure but in actual practice the rate is lower so the hurdle rate so to speak that the robot needs to meet declines i thought that was just really interesting the way you frame that out bingo bingo again yeah did i uh did i make the uh two hikers out running a bear analogy last week you did i don't know if you did that last week or if that was in one of our practice shows that we didn't publish but go ahead and kiss you basically two runners uh come up to a our hiking sorry two hikers uh come up to a bear in the woods and the bear starts looking at them angry and one of the hikers very quickly takes off his hiking shoes and starts putting on a pair of running sneakers and the other hiker says what the heck are you doing you're never going to outrun that bear and the hiker says i'm i don't need to outrun the bear i just need to outrun you um right so the objective is not to get a perfect score the objective is to get is to get a the objective of the ai is to get a score that is better than how the humans are doing and the humans are don't always do as good as they think they do so it's not as it's not an impossible hurdle to reach by the way one of the places you can see this phenomenon is in birth control methods highest uh actual rate uh versus effective rate differ dramatically um you know i do want to mention about uh you know more from a more abstract standpoint with regard to this uh what this spells for the future is that the application of the ai for this particular in this particular manner doesn't actually deprive any restaurants of any jobs they're not firing people uh who are just on the phone because they uh instead they're augmenting uh the the availability of these restaurants uh to take orders that are otherwise just being abandoned.

21:37So this isn't putting anybody out. Instead, it's actually facilitating economic activity. So that's pretty exciting. We're going to take another quick break and be right back with more of the day's top analysis on the Real Vision Daily Briefing.

21:56Mikhail, I could put on my cynical devil's advocate hat and add a single word to the end of that which is today today it's not displacing actual workers because it is you know fine it's found this problem where you can basically optimize uh by bringing in new technology so you're not actually laying anyone off but that is today it's it's pretty clear that you can see the direction and i'm i'm sure this is a topic that we're going to cover on many future shows which is just the potential economic shock of what's going to happen when we start to see this move through the labor force i know some folks made some comments in the youtube last week about precisely that point and hopefully we'll bring in some uh some professional trained economists and have this conversation with them and get their thoughts and get an understanding from a different point of view from ours about how that's going to shake out i think it'll be very interesting to talk to economists about this uh because at the end of the day here's the deal uh i think that anybody you know all of us want to like progress towards that like glorious star trek future of a post-scarcity society right but a post-scarcity society means that it means not only that material goods are universally available through replicators, but it also means that labor is no longer scarce either, which means that these machines are capable of displacing the producers of labor just as surely as they're capable of displacing the producers of goods.

23:17And what that means is that we all have infinite amounts of labor at our disposal, as well as replicated meals coming out of the little shiny box thing. and maybe at some point in the distant or not too distant future, we end up in the matrix. Hey, Mikhail, let me ask you this in our very much non, you know, non world where we don't yet have this lack of scarcity. Let's talk about something that's really interesting, which is chat GPT, open AI producing an enterprise version of their application to help monetize it. I think it should probably surprise nobody that this was coming. Talk a little bit about what they're doing, what the use case is, and what the model is there.

23:58You know, I tried understanding how ChatGPT Enterprise really differs from ChatGPT Plus, and I'll be honest, it doesn't seem like... By the way, tell us what ChatGPT Plus is, because I think this is something that both you and I pay for. I use ChatGPT Plus pretty ubiquitously. It allows for you to get access to the latest trained models that OpenAI produces. It allows for it lets you run those models on slightly faster servers, which means that your throughput is like you you can get more words out of GPT faster. It also allows you to use their plugin system that I demonstrated previously in one of our interviews, where you can have GPT sort of outsource its processing to plugins such as Wolfram Alpha to get it to do better math.

24:55Or you can have a little code runner to get it to actually write little bits of code for itself and then run those code snippets. By the way, I should say in our demo of ChatGPT, you showed precisely some of the challenges that exist with the sort of math coprocessor units in the logical structure that is the large language model. GPT is a language model, and that means that it can't do math. The analogy that I gave in, I think, our third interview was that you can no more expect GPT to be accurate at math than you can expect Excel to work as a spell checker. so there um so uh with plug-in architectures you can get gpt to do a little bit better at that kind of thing now uh open ai just unveiled gpt enterprise chat gpt enterprise which basically lets an organization buy an entire block of access to gpt plus essentially to a to its entire staff It also lets the organization create certain chat templates that might facilitate more conversations along the lines of what exactly it is that the organization deals with.

26:08So, for example, if the organization is a marketing agency and writes a lot of marketing copy, then they can precede their organization's edition of ChatGPT with questions that might help people write better ads. they also give additional data privacy guarantees because a lot of people who would otherwise use chat gpt for business are apprehensive to do so because they know that all yeah because they don't want it to be used as training data yeah that makes perfect sense but we don't yet have true customization by enterprise or by organization with this new enterprise software is that correct Not really.

26:53You can use fine tunings, which allow you to recalibrate the weights of the neural network for your own specific needs. But the setup for fine tuning requires a bit of technical acumen, and it's also really not cheap. Now, here's the thing that I want to emphasize about this release of GPT Enterprise Edition. it's very good to me anyway this might be a controversial stance but it's pretty obvious to me that the only reason open ai is doing this is because they're trying desperately to find some way to get gpt into the black a lot of viewers might not know this or at least uh you know we i believe we discussed this last week but gpt or open ai is losing a lot of money um And they their revenue in 2022 was only 30 million dollars, which sounds like a lot.

27:48But their costs ran into the ran into half a billion. On paper, just the cash flow statement, not a great business. Obviously, the perception out there is that this technology is truly groundbreaking and therefore will advance in ways. And by the way, this is a very well-trodden Silicon Valley paradigm, which is you lose money, you get adoption, you get a huge user base. and then you figure out how to monetize it later. As long as the functionality is there, as long as the adoption rate is there, so goes the thesis, this is something that can be monetized later. But it really is something that's incredibly interesting.

28:22And at some point, you have to think that these are going to have to become cash flow positive businesses. They're going to have to generate free cash flow. Otherwise, maybe the economic model needs to be rethought a little bit. There's a specific archetype to this entire development that I'm reminded of. The inventor of movable type, Johann Gutenberg, you know, created this civilization changing invention, right? He died penniless and forgotten in some tiny little village in southern Germany without ever being able to profitably capitalize on his invention. He printed a bunch of Bibles. The Gutenberg Bible, of course, is very famous.

29:05I think he printed a total of like 300 of them or something. Um, but the, but he just couldn't move product. Uh, and he had to return the loans that he had made for, uh, building his machine and renting out his workshop and, and, uh, running his workshop and his creditors ended up, uh, acquiring his publishing house. And he never, like he, he, he was never financially successful with it. Yeah. It's an incredible story, by the way, I'm sure Silicon Valley VCs on Sand Hill road now saying you just you needed us we would have made it all better for you i'm sure um so the you know uh this does sort of once again drive home the reminder that like others will rise to the occasion if uh you know if open ai doesn't eventually uh manage to sustain itself uh somebody's gonna acquire gpt probably microsoft um and there will be other contenders uh by Facebook, by Google, by Amazon, by Microsoft, by Apple, that will provide LLMs.

30:14So this is just like emphasis of the fact that this field is just on the cusp of absolutely exploding. Okay, Mikhail, and to borrow from Monty Python, and now for something completely different, Mikhail, what the hell are wugs? uh well i wasn't gonna like i was just gonna show you one at a time this by the way this is if you're listening to this podcast we had about 10 000 people listen to the podcast uh last week so michelle is holding up a little blue bird it looks like is it a bird it's a it's a it's a stuffed plushie that uh my better half amanda spent a day making um and yeah it looks like a little like a little blue easter peep so this is a wug it does that's exactly what it looks like here comes another one now there are two of them yeah there are two wugs they're wugs exactly um now this and i'm i'm holding up a much smaller little baby version of it this is a young baby wug what would we call this like oh i don't know like a wugette a wug you know and there there is no right answer that's the point if we were speakers of spanish it would be wugito um the answer that's listed in the book i'm about to cite is wugglet or wuggling i believe wuggling okay here's one ash they're just adorable by the way if you're listening to on a podcast here um here's one ash uh what do you call a facility that that hatches and raises wugglings uh a wuggary oh my god you got it on the first try that's amazing right is that the right answer so these are a set these are a bunch of non oh oh oh sorry i forgot um this little wuggling is about to flebe would you like to see it flebe i i think so it will now flee before you uh for those that are listening i'm just rotating it in my hand like turning around and around so what would you say that this that this wuggling is doing right now i i would say it's it's about to get us a psychiatric check.

32:35It is fleeping, and yesterday it flebed. So look, these are a bunch of nonsense words. We're going to explain this complete insanity shortly. We have not totally lost our minds. These are a set of nonsense words that date back to a linguist and psychologist named Gene Berko Gleason back in 1958. OK. And they were given to the little exercises that I just gave you were given to a bunch of preschoolers and kindergartners to see whether or not they would all come up with common answers and whether or not the common answers that they came up with were the same as the answers that adults would come up with.

33:23so this was a really interesting exercise in language acquisition because prior to that prior to these experiments it wasn't entirely clear that a student that a child didn't learn language simply by rote memorization um but these are words that they could not have possibly memorized because they were words that uh that gleason had just made up and there's no such thing as a right answer because again this is just stuff that gleason made up but she did establish that the answers that adults came up with would be the same answers that children would come up with as well around age four or five or so now i want to answer two uh two obviously raging questions one why am i talking about this now and b why am i talking about this at all um the reason i'm talking about this like questions yeah like questions we go um the reason i'm talking about this now uh is because there's been a lawsuit in the last uh couple of years over the use of wub of wugs uh in uh like as a commercial product property okay um gleason has recently uh written a children's book uh in which she's trying to monetize on the you know on her invention of these little guys and unfortunately in the 60 years since her original experiments uh she's um like the the wug has become a mascot of linguists and the uh and so there's a lot of other products that are already monetizing this little image uh i feel that i feel we're probably safe to use it because we're discussing it for academic purposes and like you know educational and we're like saying like you know this comes from john burke gene burke o gleason please go buy her book um but the uh but the point is that um the uh that they've been these little obscure nerd mascots have been in the news lately the reason why is all of this relate to ai because it lends because it's shed some significant light on how large language models work and how they mimic the operation of the human brain.

35:37The original experiments were instrumental in developing an understanding of language acquisition in young humans. And what it drives home is that the total amount of information that you need to have in order to have mastery of a language, or at least operational capabilities of a language, is much smaller than the total number of words in that language. You just need some vocabulary and then transformation rules. Now here's the thing, the number of transformation rules to speak English is absolutely enormous, and linguists have tried to catalog these rules over generations and have ultimately failed.

36:20A neural network is very, very good at learning from example, and it is exactly the kind of AI that you throw at a problem when you want to say there is some pattern in this data but i'm but i can't be arsed to articulate exactly what that pattern is can you please figure it out okay let me ask you let me ask you a question about this because i find this absolutely fascinating you know while most uh young people in college may occasionally have a flirtation with noam chomsky's socialism i knew i was a died in the world capitalist even as a young man uh so i had a flirtation with his transformational generative grammars.

36:54The idea is that the human mind itself possesses these innate structures that allow us to create things like syntax, morphology, and the other transformations that we see in languages. How does that apply to computers which don't inherently possess those physiological wetware layers of creating language transformation? How does the machine do it? So I'm really glad you bring that up because Because Chomsky's innate grammar theory was very interesting and has actually proven instrumental in the development of artificial languages, specifically computer languages. Computer languages use something called context-free grammars in order to be parsed into machine code and compiled into something that the processor actually, you know, into the ones and zeros that actually write on the processor.

37:50And so Chomsky's theories of linguistics actually were really important in the development of compilers. But they have proven to have absolutely no basis in how humans process language. And they have no explanatory power in either language acquisition or in the describing of human capabilities. And the example that I always use with Chomsky Grammars is that I am perfectly capable of producing non-grammatical speech. Company, open AI, customers calls during workload unveils. Yeah. Or you could just pick up some random words. You could just pick up a contemporary sociology paper and get the same effect.

38:37Very good point. if chomsky grammar if chomsky's linguistic theories were true and there was innate circuitry in my brain that required me to emit language on a on some set of grammatical rules i would have been unable to utter that sentence uh so the um so the truth is that true we have a sort of a biological uh imperative that drives a certain type of gate in humans but i can hop on one foot if I want to be really inefficient. That's, I like that. But your knee still bends the way that your knee bends. You know what I mean? Hopefully, Mikhail, every knee shall not bend at the will of AI in the near future.