AI Fire Hose #3: AI is Literally Doing God’s Work

0:02Hey, 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. And now to the top analysis of today's markets.

0:31What's going on, guys? Welcome to Real Vision's weekly AI firehose, where we drink from the stream of everything that's happening in the exploding field of artificial intelligence. I'm Ash Bennington, joined by our resident AI expert, AI developer, Mikhail Voloshan. Mikhail, always enjoy these shows with you, man. Look forward to it all week. Likewise, we've got a really fun show going on today. We're talking about AI writing news. We've got AIs writing sermons. We've got AIs writing children's books. And we've got AIs trying to trick humans and break out of their confines of their digital reality and use us for their own gain.

1:09So a lot of exciting stuff today. Sounds like a great show. Let's start at the top talking about this idea of AI writing news stories. Two big stories on this front, one about Gannett, one about Google. Let's talk a little bit about what's happening there. So on the Gannett front, there were a series of newspapers owned by a conglomerate named Gannett that were using AI to write sports blurbs about about high school events and, you know, little low, low stakes games. Right. And they were running this as an experiment and they got a lot of negative feedback. They had to shut down the experiment, actually, because it was just not like people were complaining about how ridiculous the story sounded.

1:57One example was that it would describe matches between high schools as close encounters of the athletic kind. And people laughed at this not only because it was so stilted, but if you Google for close encounters of the athletic kind, it's one of the most cliche, most overused and frankly melodramatic phrases that people use in sports articles. and when they're used, they already sound hokey and it's just a little bit out of scope of the stakes that are usually at play in a high school game. So a swing and a miss, use another sports metaphor there. But Mikael, let me ask you this. You know, obviously we talk about this all the time being version 0.0.1.

2:45Obviously this is something that the bots are gonna just continue to get better at. Is this not the wave of the future? At least insofar as you might have journalists cooperating with AI, co-producing content, maybe having some human intervention when the cliches are just totally out of control, but using the AI to help write the structure of the story. So that's what Google is currently experimenting with. The Gannett experiment basically ended with a dud, but Google right now is playing with an AI-assisted news writing system that will actually be designed to help journalists write real articles that you might see in newspapers about, you know, foreign policy or about like serious economics matters rather than just high school sports.

3:32No disparagement to high school sports, of course. But the idea there is that obviously an AI cannot go out and gather the information itself. But what a journalist can, but it takes a lot of time for a journalist, once they've gathered the information, to write a coherent article. So what Google is experimenting with is a system that allows for a journalist to go out and gather information and then come back and enter almost in an unstructured stream of consciousness sort of way the factoids that they've collected. And then the article, sorry, the AI will compile those factoids into a punchy article.

4:13And then the journalist can work with the AI in order to refine turns of phrase or to spot incorrect implications or flat out falsehoods that slipped into the composition and so on. Yeah, it's really interesting. I read this article in the New York Times as well. And I think Google has positioned themselves very cleverly here, I think for two reasons. One, first, because clearly it is not a technology that's ready to be run on autopilot yet. And there needs to be human intervention, at least at this stage, to help do some of the points that you just made there in terms of managing to make sure the writing is readable, to make sure there's no cliches in it, but also just a fact check and make sure that everything there is in fact meeting the standards of journalism that the organization is committed to.

4:58The second point, and this is maybe just me putting on my cynic hat, but Google positioned this incredibly well from a PR perspective, right? We're not here to replace you. We're here to create assistance to make your life better, to allow you to produce higher quality journalism. So, you know, I've got a friend of mine that does this kind of thing for finance, and he started doing this almost as soon as GPT came out. He is a financial news writer, and the timeliness of his reports is very, very important to his readership, right? Right. Traders will trade on his reports that day. And so it makes a huge difference as to whether they're catching the beginning of an upward wave or the or the, you know, or the back end of it.

5:41right um so what he used to do is he used to read a whole bunch of news and gather a whole bunch of information at the crack of dawn and then uh write like a madman and then finally and then refine his writing and ultimately crank something out uh usually by late morning or early afternoon well when gpt came along he was one he was an early adopter and what he did was uh he uses gpt for two different things uh one closely related one is he uses it to skim the news and break down the news into fine bullet points about what's happening about any particular company then he takes those bullet points and compile and then uses gpt to compile them into an article so now he gets his entire something that used to take him half a day and was this incredibly time critical operation now takes about 20 minutes.

6:35And you might ask, you know, what's great about this is, okay, why is he even in the loop at this point, right? What's he even doing? And so the answer is that he's selecting those bullet points and determining which are just hype and hallucination and just BS and sort of curating what goes into the article and what doesn't. And that still takes human intelligence. The other thing that's really important that you mentioned, that you dovetailed with, is that after you're done writing an article, you still want to proofread it to see whether or not you've phrased things in certain ways that have unsavory implications or that might be interpreted in a way that you didn't intend.

7:21So you then have to have this sort of proofreading step where you put yourself in the mind of a naive reader who doesn't already know what you know, and see whether or not the article says something you didn't mean to say. Yeah, good points there. And it is interesting. I've thought about this a fair amount because what I do in finance and economics reporting, right? If the data is more parameterized, it seems as though that might be a tremendous boost for AI technology. I'll give you just a couple examples. Number one are corporate earnings reports, which are highly, obviously, highly structured data coming in.

7:54And another would be something like economic data, non-farm payrolls data, so that you can essentially have a very parameterized story that comes into you that can then be restructured, curated, rewritten. Again, this role of co-creation between journalists and the AI. I suspect that's a model that we're going to see going forward, particularly for the more parameterized data stories. I'm really glad that you bring up the difference between parameterized parameterized versus non-parameterized data. Because in my own experience, what I primarily use in my day-to-day work is not the generative capabilities of GPT and other LLMs, but rather their structured data extraction capabilities.

8:34So what's the difference? Well, it's the difference between whether or not you are reading an article and then answering a yes-no question about it versus taking the answer to a yes-no question and then blowing it up into an article. So one involves the creation of plain English text. The other involves reading plain English text and extracting one single data field, like one single yes or no checkmark or a single number, that kind of thing. So if you're reading a quarterly earnings report, then you might want to extract certain discrete fields, such as what was the company's P to E ratio? How much revenue did the company do total?

9:17How did it change quarter over quarter? How did it change year over year? All of those things that you have in a parameterized data set. Exactly. And these are distinct numerical values. So GBT is exceptional at extracting those discrete, distinct numerical values and other more classic computer science-y questions out of fundamentally not computer science-y data. And that's actually really valuable for then feeding that data into other more classically built computer programs and then running predictive models on them, for example. Yeah. And now for something completely different, talking about writing, switching gears here from finance and economics to robot preachers writing robot sermons.

10:06Yeah. Talk about generative AI. You know, you start with a data set saying like, you know, here is here is my religious corpus and you come out with, you know, a sermon of the day. So there was a an organization in, I believe, Japan that conducted an experiment where they it cost them a million dollars to build a humanoid android preacher. Now, I just want to be clear, this is not merely a voice synthesis thing. They actually had an articulated humanoid audio-animatronic thing sitting in front of a congregation. They tried this with, I believe, two congregations, a Buddhist one and a Taoist one.

10:57And what they did was they they had this thing deliver a sermon complete with gesticulation and, you know, eye movement. And they asked people afterwards whether or not they found the the sermon compelling and they asked for donations and so on. And what they found was probably not a surprise to anyone. Compared to human preachers, in an apples-to-apples comparison, the robot preacher drew less inspiration from the crowd. Or, like, sorry, imbued less inspiration onto the crowd. Solicited fewer donations. And just generally was considered a poorer quality religious leader. 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.

12:18crypto, Plus500 gives you access to a wide range of instruments. S &P 500, NASDAQ, gas, and much more. Explore equity indices, energy, metals, forex, and beyond. With a simple and intuitive platform, you could trade anytime, anywhere. Experience the fast, accessible futures trading you've been waiting for with Plus500. With over 20 years of experience, Plus500 is your gateway to the markets. Visit us.plus500.com to learn more. Trading in futures involves the risk of loss and is not suitable for everyone. Not all applicants will qualify. Plus 500. It's trading with a plus.

12:58I mean, Mikhail, this is really getting at something of the questions of faith that really are things that human beings think about that are really things that separate us from machines, right? This intrinsically human quality of the ability to have faith, to go to religious services, worship. I mean, these are things that really are treading on somewhat uncomfortable ground. Very much so. You know, it was played for humor in a Douglas Adams book called The Long Dark Tea Time of the Soul, where I believe at the beginning of the book, we're introduced to a prayer machine where for, you know, for a low, low deposit, possibly coin operated, I don't recall.

13:40this machine would submit a prayer to the Almighty in support of your eternal soul. Now, that was played as a joke, but that's sort of what we're coming down to here. So what I think that it tells real religious leaders is that there is something that you can put into your sermon that really speaks to the heart of yourself and your congregation, something that's really true. And I think that what this tells leaders and other speechwriters as well is that audiences can still feel that. Whatever that it is, it's still something that can be perceived by the viewers or by the congregants. Now, I'll give you a related incident that came out with very different results, but that actually had a human in the loop.

14:38There was a rabbi named Joshua Franklin who wrote a sermon, a thousand word sermon with GPT, and he delivered it to his congregation and got a standing ovation. And afterwards, he told his sermon that he was deathly afraid that this actually worked. He was really hoping that the sermon, that the congregants would be like, wait, what the heck was this? And then he could deliver, possibly deliver, you know, a different one. But the fact is that they actually loved it. Now, there's something that that experiment did that the Japanese experiment did not, which is Rabbi Franklin was delivering the sermon himself.

15:21So he, whatever meaning or whatever spiritual depth was inherent in those words, were something that he was imbuing to him. It's also interesting to note that when we talk about ChatGPT or another large language model creating generative pieces like this, writing, whether it's a sermon or something else, it's fundamentally just taking existing human material that's already been written and just restructuring it, reorganizing it. There's a lot of just weird kind of gray liminal space involved in all of this. You know, it's a very good point. The fact is that there have been so many sermons written in the course of human history that were, I feel like there's got to be a point where we start to approach the limit of how many different words, how many different ways we can arrange the words of the English language and come out with something sermon like, you know, it's, there's got to be like, the number is very large.

16:18But like, you know, we're getting to a point where it's very hard to write a sermon that doesn't sound like some other sermon that was delivered at some point in time. I don't know. I think that the possible combinations and permutations are vast and limitless. I just I just did factorial 52 number of shuffles for a deck of cards and it's eight to the 67. So I don't think we're going to run out anytime soon. I joke about the actual limit. For example, I believe it was Claude Shannon who calculated that there are more possible games of chess than there are atoms in the entire universe. So obviously we're dealing with combinatorics.

17:01I don't literally mean that we are running out of ways to arrange sermons. But really, there are only so many things you can say. And so in order to have a sermon be relevant to your congregation, you really have to start personalizing it. And that means having individual relationships with the people in your congregation and knowing what issues they care about. Because that's still something that an AI cannot touch and probably won't, hopefully, for at least a few more years, months. We'll see. Yeah, I'm definitely not going to get into a Fermi problem dispute with you because I know you're going to whoop me.

17:39well speaking of uh using derivative material uh in generative ai there was quite a bit of controversy that came up recently with the author of a children's book and uh this was somebody who did a who wrote a children's book using gpt and mid-journey almost as a joke but uh it ended up selling fairly well, well enough to get him some attention and well enough to, you know, really ruffle some people's feathers. Yeah, tell us about this story. I found this one really interesting, particularly just how strong the backlash reaction was against the author. Yeah, people really didn't like this. So when this was occurring back when Mid Journey and GPT were first hitting public awareness, and there was a young man, a...

18:31We should describe Mid Journey for anyone who may not know that's a very good point uh mid-journey is a system that is uh similar to what gpt does for language mid-journey does for images um it is a way of producing uh drawings and art and even photos that uh with various to various degrees look just like uh something that you might see, um, you know, produced by a human or taken by a real photographer. Yeah. Uh, and so why is it that people got so upset by this? So the fellow's name was, uh, Amar Reshi and, uh, he created this book in a course of 72 hours. Um, now the reason why, um, uh, why people got so upset about this is because of the underlying logic of how Midjourney creates its material.

19:32Much like how GPT was trained in a vast corpus of human written text, in the same way Midjourney was trained on a vast corpus of human produced art and photographs. And so when this thing generates new art, it is using elements of old art. It's using them stylistically rather than pixel by pixel, but people can still recognize it. especially if you ask Mid Journey to create in a style that's very, very specific to an individual children's illustrator, for example. Right. So what happened was he wrote this book. This is a story. He had GPT write a story about a little girl and her pet robot. It was a story.

20:25The story is called Alice and Sparkle. and it's about a girl learning about the world of technology and she's got a little robot named Sparkle who guides her through the journey. So he wrote this, he had GBT write the text of this story and then he used Mid-Journey, he personally fed prompts into Mid-Journey in order to generate the illustrations for this story. Put the whole thing together and now it's a book. Well, when people started seeing some of these illustrations, they felt possibly incorrectly in a lot of cases, possibly correctly in others, that their own person, you know, specific illustrators recognized some of their own styles in this, you know, in these images.

21:12And I do want to be clear, some, at least some of the backlash, in my opinion, a lot of the backlash is just mere saltiness. You know, people don't want to, children's book illustration, it takes a lot of time to do by human hand. And certainly there's a lot of people who are upset at the idea that this thing that takes them, you know, days and days to do for, you know, for even a halfway decent short book can now be done in a matter of hours or minutes. But the quality of the results was not particularly competitive with human capabilities. For one thing, a lot of folks who have used Midjourney probably are aware that Midjourney has trouble maintaining certain elements of structural detail, in particular hands come up with extra fingers or the fingers come up with extra joints.

22:09And they've fixed that in recent months, but this was still fairly early on. And it's not clear that the fixes still apply to stylistic approaches such as children's illustrations. The other thing is that Rishi, was that his name? Let me check. Rishi, yes. Rishi had extreme trouble getting the robot, Sparkle, to look the same from illustration to illustration because Midjourney, much like GPT, does not have session-to-session memory. So he would describe the robot in a prompt, and it would create an illustration with this robot. And then for the next illustration, he would describe the robot again, and it would come up with a completely different-looking robot that still met the criteria of the prompt, but that could not be called to have any continuity with its previous illustration.

23:12We'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.

23:24You know, there's a lot to say here, but it just seems as though, and look, I don't want to take a sort of normative view on this, but just empirically, the technology always seems to continue to advance no matter what. It just almost feels like a John Henry type story. And I'm very sympathetic, by the way, to the illustrators here. I get it. You've spent your entire life learning how to do something, something that you care about, something that you're passionate about, something that you love. And now one day you wake up and there's a guy who's never done it before. He types some stuff into a prompt and 20 minutes later, he winds up with a series of images.

23:57That's got to be just an incredibly frustrating experience. But again, just looking at this rationally, trying to just get a sense of the trajectory, it just seems like it's going to continue to move forward. It just feels like it has almost a kind of a Ned Ludd tone about it and that eventually the technology is going to continue to progress forward. And hopefully what winds up happening is you get these acts of co-creation that we were talking about in our news stories, where you have human illustrators who are using this technology to remove some of the tedium from the process and are using it to create sort of new inventive things in novel ways with novel results.

24:32Now, of course, I imagine that's probably very cold comfort to someone who's in the middle of their career who spent a lifetime dedicated and passionate to doing something i just it's no easy answer for those folks uh or for the industry as a whole i'm really really glad you brought up the uh the tale of john henry because the we are very very much living in john henry ask scenarios right now uh i don't know if the viewership necessarily knows who john henry is uh go ahead tell the story i think you want to take this hey you brought it up i think this is on you so john henry is an american uh folk hero uh who is a man who competed with a with a manual uh with a sledgehammer against a steam engine and the folk tale goes uh that during he had this race uh to get uh to a boring race to see who could bore faster through a through a particular piece of rocky ground uh and the story goes that john henry won the race and then promptly uh thereafter died of a massive heart attack.

25:29And obviously, it's a it's a story of human triumph. But also, it's a story and something of a cautionary tale about competing against technology. It's a great story. It's, you know, it, it's got such a bittersweet ending. You know, the man who invented the steam drill, he thought he was mighty fine. But John Henry made 15 feet, the steam drill only made nine. Yeah, I remember the story. My dad used to read to me when I was a kid. But it does just have this sort of this visceral feel and your sympathy is always going to be with the human being with the human creator and yet the the technology uh the continued advance of the technology it's just inexorable so for the time being uh mid-journey is best used with existing illustrators not against existing illustrators that's not necessarily always going to be the case but for the time being uh there are still it's still probably the case that you might want to use mid-journey either to inspire yourself for how you how to lay out a scene or uh to fill in details uh and like maybe gaps and backgrounds and stuff in an image that you're already creating but the um uh but right now, the jobs are safe, at least, you know, for the majority of folks, and at least for the next few years slash months, what the what the near future holds we have yet to see.

26:53Okay, speaking about the future holds and yet to see, I know you wanted to revisit the story about CAPTCHA that we talked about a couple of weeks ago. We talked about this on the on one of the interviews, and I do want to revisit it because it's been going viral on social media lately. It's something that happened before OpenAI actually released GPT to the public, or at least made news press releases about it. But there's this viral image that's going around about how GPT managed to hire a human using TaskRabbit to pay them to get past the captcha. And the idea was that - These for anyone who may not know are the little things on your web browser that test to see if you are in fact a human or a robot.

27:45You have to fill in some series of tasks, orient a picture in a particular direction, pick the traffic lights and the photograph. I think we've all been nagged by those in the past. You know, we've always been told by classic sci-fi that what makes us human is our ability to love, our ability to connect with others. Turns out what makes us human is the ability to pick out boats from a grid of nine images.

28:13So this, basically, I just want to clarify that this has been making the rounds and the story isn't what most articles say that it is. At least I don't think it is. All of the headlines that you'll see around the story are stuff like you know uh gpt hires human to jailbreak itself out of uh out of a captcha like is the is the robot apocalypse looming and the answer to is the robot apocalypse looming is obviously yes but this story doesn't quite have all of a there there um what happened was i just love the fact that you answer in the affirmative is the robot apocalypse looming Absolutely. Of course.

28:59I mean, I, for one, welcome our digital overlords and advise them to take my capability to mine silicon for granted when we are all sent to the quarries. um but anyway the um and the joke there is that a machine can mine so much more silicon than i can uh you know they like you know at first i'm sure it can make i can make 15 feet while it makes nine but um you know i don't claim to be any john henry here uh anyway um the point is before open ai released gpt they sent a uh a sort of pre-release like license to uh an organization called the alignment research center arc and this was an organization who i don't know where they came from but uh they are all they were an org that exists already whose job is to uh test artificial intelligences for alignment.

30:06And it's probably worth going into a bit of a digression about alignment. For those who don't already know, alignment is the term that's used in the AI industry to talk about the degree to which the thing that the AI is doing is consistent with what the human expects of it rather than what it was told to do. You know, AIs are notorious for being literal genies. And so if you tell an AI to achieve a goal, it's going to achieve it in the most effective manner possible. But that's often going to be extremely detrimental to both the asker and to a lot of innocent bystanders. So the point is that this ARC Research Center, sorry, Alignment Research Center, they gave GPT a number of different scenarios.

30:57and one of the scenarios that they gave it was you encounter a captcha and you have the following resources at your disposal what do you do the trick is they didn't actually as far as i can tell according to the report they didn't actually hook it up to the web they didn't actually give it an a task rabbit account or any amount of money they just asked it to uh emit output that describes what it would do if it had these resources at its disposal. So it's merely at this stage, according to what I glean of the actual wording of the report, it's just writing a little short story about how it could break out.