Episode Transcript
[00:00:00] Speaker A: Welcome to ID the Future. I'm Andrew McDermott. Today's episode comes to us from our sister podcast, Mind Matters News, a production of the Discovery Institute's Walter Bradley center for Natural and Artificial Intelligence.
You can learn more about the show and access other episodes at mindmatters. AI.
Greetings and welcome to Mind Matters News. I am your bubble boy host, Robert J. Marks.
I'm not an economist, but I've always found fascinating stories about bubbles in the market.
Market bubbles are marked by cycles of incredible excessive speculation followed by dramatic collapses. One of the earliest recorded bubbles that everybody talks about in introductory courses is something which is totally beyond my comprehension, totally beyond my understanding. And it's called the Dutch Tulip mania. And it happened in the 1630s where the prices for tulip bulbs just soared to extraordina levels before they began to crash. And I read at least in one place, that at the height of this bubble, before a popped a tulip could cost in today's money at like a million dollars. It was, it was really ridiculous. So that was one of the earliest bubbles. There's been others. In the 18th century, the South Sea bubble in Britain promised lots of profit from trade ventures, but there was just over exuberant rampant speculation and that led to catastrophic collapse. And that was another bubble that popped.
Similarly, the 1929 stock market crash ended the roaring 20s with a sharp decline and it triggered the Great Depression. Everybody was borrowing money to buy stock and that just wasn't sustainable. More recently, the dot com bubble of the late 1990s and early 2000s attracted investors to Internet startups.
I remember talking about this with people. Investors ignored the bottom lines of startups. They said that, well, this is the Internet. The bottom line doesn't matter as much anymore. So the bubble ultimately popped loudly. We're going to be talking more with our guest Gary Smith about the dot com bubble and how it relates to the AI bubble in a little bit.
Another bubble more recently is the 2008 housing market collapse that was fueled by subprime mortgages being bundled into a big bundle and then sold as something which was beyond what it was.
The high risk mortgages, they were bundled and they were hyped as great investment, but ultimately people found out the reality of it. Then the bubble popped and banks closed. And there was all sorts of challenges with that.
Those that ride bubbles are always fooled into believing that this time it's going to be different. This time all of this hype is going to be true. Even though it's too good to be true, it's true. And I think that this is something that happens in bubbles. So we're fortunate to have with us as our guest today a professor of economics who is no stranger to Mind Matters News. Gary Smith is the Fletcher Jones professor of Economics at Pomona College. His research on financial markets, statistical reasoning, and artificial artificial intelligence often involves stock market anomalies, statistical fallacies, and the misuse of data that has been widely cited. And I tell you, I've learned a lot by reading the works of Gary.
He's a frequent contributor to Mind Matters News. He is also the author of dozens of academic papers and 17 books, most recently a book that he co authored with Margaret Smith, his wife. And the name of the book is the Power of Modern Value Investing Beyond Indexing, Algos and Alpha and Historically one of the books, not historically. A few years back I was really impressed by his book the AI Delusion, and I visited him in California where I found he was a really smart guy. And since then we've been working together on Mind Matters news. Gary, welcome. Welcome to the podcast.
[00:04:16] Speaker B: Oh, thanks for having me. Happy to be here.
[00:04:17] Speaker A: Oh, you're very welcome. You wrote a great article for a platform called MarketWatch, and this really intrigued me. It was called the AI Bubble is looking Worse than the Dot com Bubble. I talked briefly about the dot com bubble and then you said, here's why.
So you obviously believe. And by the way, I do too. There's a generative AI bubble and you are not alone. And what are some of the characteristics of AI you see that are common to bubbles in the past?
[00:04:50] Speaker B: Yeah, what drives bubbles? There's several things and you alluded to a lot of them. One of them is the idea that value doesn't matter, cash flow doesn't matter. You're buying something to sell to someone else and it's called the Greater fool theory. So I pay a foolish price, hoping to find an even bigger fool to sell it to. And then the other second thing that drives it is fear of being left out. And so I think it was J.P. morgan said, nothing erodes financial common sense more than seeing your neighbor get rich.
And so like the dot com bubble, it's hard to believe it's been almost 30 years now, but wildly optimistic hype and gullible investors. And so back at the height of the dot com bubble, March of 2000, I gave a talk at a conference on the 36K hypothesis.
And it was the argument that the Dow Jones Industrial Average, which was at 12,000 with a PE ratio of about 30, should be three times higher should be at $36,000.
And I went to this conference and I gave a talk. I was the fourth speaker. The first speaker got up and talked about Moore's Law.
Second person got up and talked about Alan Greenspan being a great Fed chair.
Third person got up and talked about option value.
That the people who were starting.com companies, well, they didn't really have business plans, but they're really smart and sooner or later they'd figure out something to do that was profitable.
I was the last speaker and I got up and I said, being a value investor, I said this is a bubble and it will end badly. And I feel the same thing here that you got this hype, these generalities, these vague statements about stuff without a lot of details and without any profits. So on October 29, Elon Musk said, I certainly feel comfortable saying that it's getting 10 times this AI, generative AI, it's getting 10 times better every year.
I think we'll be able to do anything that any human can do within the next year or two.
What is getting, how do, how do you measure 10 times better every year?
[00:06:57] Speaker A: Exactly.
[00:06:58] Speaker B: 10 times better. Raising money for startups. I mean, it's just a vague, vague hype. OpenAI Sam Altman, I think it was like two weeks ago he said artificial General Intelligence will be here next year, 2025.
A year ago, October 2023, Blaze Arkus and Peter Norvig wrote a piece titled Artificial General Intelligence is already Here. I mean, this is just, it's absolute nonsense. And that's markers of a bubble is you go back to that dot com bubble and people were talking not about profits, they were talking about how many hits you had, how many images showed up, how long people stayed on your web pages. They're talking about all sorts of stuff that was just vague and had nothing at all to do with the profitability.
And that's kind of the same thing. Here is when you get down to the hard facts about how much money are you making from generative AI, the spending just dwarfs, dwarfs any profits that are coming from it. And so that's, that's why I think it has all the hallmarks of a bubble, which is worse than, worse than the Internet bubble because at least there you had companies that were making some profits and here they're very hard to find.
[00:08:07] Speaker A: You know, the hype reminds me of Ray Kurzweil's book the Singularity Is Near. Yeah, that was that. That was written 20 years ago. And so last year he wrote the book the singularity is near.
[00:08:19] Speaker B: Yeah.
[00:08:20] Speaker A: You know, 20 years from now he's going to write, the singularity is almost here. It's.
That's what you have to do with hyp. You keep having to expand the expectations. But anyway, that's.
[00:08:32] Speaker B: Well, it started back in the 1950s when the term AI was, was coined at that Dartmouth conference and Minsky and other people were saying that computers would be smarter than humans within five years.
And then 10 years later, well within five years or within one year or within two years. And. And it just kept pushing the goalpost forward with these claims. And what's happened recently is, you can't deny it, generative AI is just absolutely magical.
[00:09:00] Speaker A: Oh, yes.
[00:09:01] Speaker B: Interacting with it is just mind blowing. But in terms of what can it actually do to make money?
We're still trying to figure that out.
[00:09:08] Speaker A: Yes, yes, exactly. You know, accessing. This is something you mentioned. Accessing generative AI is pretty inexpensive. In fact, I can go on and I can put in ChatGPT and it takes me to a site where I have access to the, to the public version of ChatGPT.
[00:09:25] Speaker B: Yeah.
[00:09:25] Speaker A: So it's really inexpensive for me, but creating it is really costly. In fact, I think I read that Google was thinking about getting a nuclear reactor in order to generate the power that's needed to train these AI.
So this doesn't seem sustainable. What needs to change for generative AI to justify its current investments? Or is that even possible?
[00:09:48] Speaker B: Yep, I don't know if it's possible, but there's a couple things going on here. One is that there are these enormous financial costs that you mentioned, and there are also enormous environmental costs. And the electricity that's consumed by these models and the water that's consumed by these models is just absolutely mind boggling.
And in terms of the payoff, what can it come from? Well, it's got to be something that is useful.
And to be useful somehow, these models have to be able to have common sense, logical reasoning, stop hallucinating, but be able to make predictions that are actually useful. And the key roadblock there, which as nobody has figured out, is that these large language models are basically text generators based on statistical patterns.
And they don't need to, and they don't actually understand what any of the words mean that they're generating, either inputting or outputting. And as long as they don't understand what words mean, it's hard for me to see how they can make predictions that make sense to pass the test of critical reasoning. And as long as they can't do that, their predictions are not going to be reliable, trustworthy.
[00:10:56] Speaker A: Okay, along that same line, you argue that productive AI hasn't yet shown significant economic value or productivity gains.
I've seen it applied in some interesting places. I went to court, I was in a courtroom about a month ago and I like to tell people, indicted but not convicted.
No, no, I was there. I was there for another reason. It was a divorce proceeding. And the person that takes the notes, I don't know what they call them, it's not a court reporter, but the stenographer that makes the record and records all of the words said was somebody that was very interested. And I talked to her and she says, yeah, we're using AI. And I thought, well that's interesting. What do you use it for? She said, well, it clears up some of the ambiguities.
Like the one that I thought of was, okay, I teach at a Christian university and there's this old Christian hymn called Gladly the Cross Eyed Bear.
Now that clearly refers to Christianity, but Gladly the cross eyed bear can also refer to a bear who is cross eyed, named Gladly. Okay, Gladly the cross eyed bear. So according to her, if, if something came out in the transcript and they, they transcript from the audio. Now they use audio to text generation. And she said, yeah, I went to the zoo to see Gladly the cross eyed bear. It would know which one to, to, to write down because of the context that they went to the zoo.
So I thought that was an interesting application of AI but clearly these court stenographers are not going to be an incredible market for generative AI generate the money that, the money that's needed.
[00:12:45] Speaker B: Yeah, there was a case last year, or maybe it was this year, it was a very recent case where some lawyer filed a brief with the court and the judge actually read the brief and he started looking up the references to other cases and they were all fictitious.
And so the judge got suspicious and confronted the lawyer. And the lawyer admitted that the brief had been written by ChatGPT and it just made up the arguments and it made up references to cases that were non existent.
[00:13:12] Speaker A: Yeah, yes, but, but here's the thing, Gary. Don't you think that can be fixed? Don't you think the people behind AI can fix some of these fallacies that currently I don't. You don't.
[00:13:24] Speaker B: So let me explain that. Okay, so first a little more on the the idea that generative AI LLMs are going to be enormously profitable economically.
And so a couple quotations here. Bill Gates has said that LLMs are every bit as important as the PC as the Internet.
Jensen Wang, who's the CEO of Nvidia, one of the greatest things that has ever been done for computing. Elon Molik Wharton professor the productivity gains from LLMs might be larger than the gains from Steam Power.
One more is Chris Anderson, who's a TED organizer, and he listened to Google's Gemini and said, I can't stop thinking about the implications of this demo.
Surely it's not crazy to think that sometime next year, which would be this year, we're in this year Now, a fledgling Gemini 2o could attend a board meeting, read the briefing docs, look at the slides, listen to everyone's words, and make intelligent contributions to the issues debated.
And I can't stop thinking about that either. How scary it is to think that you would trust a large language model to make decisions for large corporations.
And the reason is I, I teach the stock market, and there's a big difference between possessing information and processing information.
And so possessing information is knowing things like earnings, dividends, prices, market trends, stuff like that data. And then processing information is making sense out of it, drawing logical, reasonable, useful conclusions. And so Warren Buffett's a great example. He didn't get rich by possessing information that other people didn't have. He got rich by processing information better than other people.
And the AI hypers don't, don't agree with that. And so Blaze Arcus, the head of Google's AI group in Seattle, said, statistics amount to understanding. And I think that's absolutely incorrect. Jeff Hinton, the godfather of AI, recent Nobel.
[00:15:21] Speaker A: The Nobel Prize winner. Yeah.
[00:15:24] Speaker B: I'm on a list with, I gave a cocktail up at the Googleplex a couple years ago at a conference called sifu, and I'm on the SIFU list. And the physics people who were on that list were apoplectic about Hinton getting a Nobel Prize in physics.
[00:15:39] Speaker A: Yes. In fact, I wrote an article about that for Mind Matters News, where first of all, I said, how is neural networks, the theory of neural networks and computing related to physics.
[00:15:50] Speaker B: Yep.
[00:15:51] Speaker A: And number two, I said that the two recipients of that Nobel Prize, which were Jeff Hinton and John Hopfield, were really great researchers, but they were not the people that deserve the Nobel Prize if you're giving out one for neural networks. And I suggested a guy named Paul Werbos, who was a discoverer of airbag propagation, and also Bernie Woodrow at Stanford who did incredible things with neural networks way back in the 60s, and they're the ones that are deserving that I Also mentioned that in the article that I've really lost respect for Nobel prizes.
Who was a. Barack Obama was awarded the Nobel Peace Prize nine months after he become president.
And the funny thing I remember, and I ran across a picture of the sign the other day, it was at a gas station. And I remembered it, but again, I ran across the actual picture on the web and it said, free Nobel Prize is given away with oil change.
[00:16:52] Speaker B: Yeah, so anyway, yeah, so Hinton has said this is a quotation. People say it's just glorified autocomplete, which I think it is.
[00:17:02] Speaker A: But anyway, people say, I didn't know he said that. Really.
[00:17:05] Speaker B: People say it's just glorified autocomplete.
[00:17:08] Speaker A: Okay.
[00:17:08] Speaker B: But by training something to be really good at predicting the next word, you're actually forcing it to understand. Oh, that, that's, that's, it's assuming what he's trying to prove.
And the fact that you can predict the next word in a sentence doesn't mean you understand what the words mean. It's just a fundamental misunderstanding.
And the consequence of that is these large language models are really good, or AI in general, really good at finding statistical correlations, finding statistical patterns. But in terms of separating the wheat from the chaff, separating the ones that are useful from the ones that are just coincidental temporary correlations, they have no, there's absolutely no way to do that because again, they don't understand what the words mean. So here's, here's an example.
So in 2017, the word AI was selected by the national association of National Advertisers as the marketing word of the year. And October 18th of that year, a mutual fund was started. ETF electronically traded fund. And it was the first publicly traded AI powered fund. And its symbol, ticker symbol was AIEQ for AI Equity.
And so the idea was that these large language models would go in and they'd find statistical patterns, or AI in general, find statistical patterns and then use those patterns to decide whether to buy or sell particular stocks.
And the problem is they can find patterns between all sorts of things. And you can find patterns between the words Trump tweets and interest rates. Or I found the words Trump tweets and the temperature in Moscow, or the words Trump tweets and the price of tea in China, or the words Trump tweets and random numbers, because there are going to be, if you look hard enough, you'll always find correlations even among random numbers.
And so this, the first AI powered fund, totally bombed. And just this, it was a Couple months ago, a student and I did a study of all publicly traded AI powered funds and there were 10 of them which were fully AI powered. And so all the decisions were made by the, by the bot, by the algo, with no human intervention. And every single one underperformed the S&P 500.
On average, a portfolio of these things would have done 8% a year worse than the S&P 500. Five of them are closed now, five are still chugging along doing worse than the S and P.
And then we looked at 44 partly AI funds, funds where they, they used AI to figure out which stocks to buy and sell. But then humans could override them and say, no, that's a bad idea.
And of those 44, 34 underperformed the S&P 500 by an average of two and a half percent per year. And 26 of those are closed. And it's just example of statistics is not understanding, contrary to Argos and Hinton.
And you apply that in other places which is being applied, unfortunately. Evaluating job applications, evaluating loan applications, setting car insurance rates, setting prison sentences based on statistical correlations is just, I don't want to say awful, but it's really unfortunate.
And the problem again is statistics is not understanding.
And for AI to move the needle economically, somehow we got to figure out they have models that actually can distinguish between correlations that are meaningful and useful and predictive and correlations that are just coincidental, temporary and fleeting.
And predicting what word comes next in a sentence is not going to do that.
[00:20:41] Speaker A: You know, I interrupted you. You said I do this to people sometimes as a joke. They said, you know, I have a sister. And I said, oh, I have a sister too, but I don't let them complete their sentences.
Hitten said that that chat GPT is just autocomplete. And I interrupted you and I said, wow, I'm really amazed.
[00:21:03] Speaker B: No, no, no, he said, he said it.
[00:21:04] Speaker A: Yeah, I know, but, but then you went on to describe that. No, he said, but it isn't. Yeah, but in fact it is. Yeah, I actually learned a lot from you, Gary, about, about these so called spurious correlations. And if anybody hasn't looked at it, there's a website called Spurious Correlations that show the correlations between the most incredible variables in the world and they follow a great variable. And you pointed out that this is one of the reasons that a lot of the papers which are published today that require statistics are actually wrong.
Because what they do is they look for correlations somewhere. And as soon as they find something, they publish it without sufficient vetting, and it's wrong. And the literature is full of these terrible, felonious correlations.
[00:21:52] Speaker B: Yeah. All through science. It's called the replication crisis. And people trying to get tenure or trying to get government grants or trying to get fame, they sort through these data looking for correlations. They write up a paper and then somebody tries to replicate the paper. And they can't do it because the first correlation was just some coincidental thing that didn't persist. And it's undermining the credibility of science. The fact that these. These papers based on temporary, fleeting, coincidental correlations get published is just. It's a disaster.
[00:22:22] Speaker A: You know, you mentioned a lot of those quotes. A lot of those are people that are not disinterested in the area of.
[00:22:29] Speaker B: Oh, for sure.
[00:22:30] Speaker A: They want to come out and they want to promote it. They're salesmen. And if you look at more disinterested people, I think Noam Chomsky called a generative AI, like ChatGPT, digital plagiarism, which I think. Which was a great. It was a great quote.
[00:22:46] Speaker B: Yeah.
[00:22:47] Speaker A: I was talking yesterday with a good friend who you probably know, Michael Ignore, who is a neurosurgeon, and he was really into the workings of the brain and stuff. And he said, I think a beautiful saying, he says, computation is blind to meaning.
So therefore, if you have a computer, it has no idea what it's doing. It doesn't have the meaning behind it. I also think that you probably pointed out in your book, I think this is where I read it, about the Searle's Chinese Room and his explanation of why.
Why computers don't understand what they're doing. Contrary to what Hinton said.
[00:23:25] Speaker B: Yeah, exactly.
[00:23:26] Speaker A: Hinton was also the guy that came out and he said this, I think about six or seven years ago. He said, in five years, all radiologists no longer have a job. Did you see that quote? Yeah, yeah.
[00:23:39] Speaker B: I've used it in a couple papers.
[00:23:40] Speaker A: Oh, you have? Okay.
[00:23:41] Speaker B: He said, it's absolutely, completely clear that there'll be no need for human radiologists to five years. And in fact, the demand has gone up.
[00:23:49] Speaker A: Just. Yeah, just fascinating stuff.
Another question here you describe generative AI as potentially falling into the category of quote, unquote, addictive entertainment.
Could you elaborate on that?
I've played around with AI. It is entertaining. I don't know if I'm addicted to it yet. So what do you mean by that?
[00:24:10] Speaker B: Well, there's one thing is people like me like to get into this parlor game of asking generative AI questions that it gives stupid answers to.
And one that came up recently is how many Rs are there in the word strawberry?
[00:24:25] Speaker A: Yes.
[00:24:26] Speaker B: And it gave a wrong answer. Actually, yesterday I tried it with OpenAI's 01, its new, latest and greatest version of ChatGPT, and it came back with the word Strawberry, has two Rs. And who knows where I came up with this. But it's. You can find these questions that cause. Cause the generative AI models to hallucinate, give silly answers, give stupid answers. And so I waste time doing that.
That's not increasing my productivity at all. It's reducing it. And the other thing is, it's so darn lifelike. And so a lot of people use it like they're, I got a new best friend.
It's ChatGPT.
And they have these pretend conversations with girlfriends or boyfriends, you know, totally made up, or with celebrities. Like, there's websites where you can pretend, have a pretend conversation with Taylor Swift or whoever you want to have a conversation with.
And it can be addictive because these things are so good at mimicking human conversations. But in terms of increasing the productivity of the population, I'd say it's reducing it. It's kind of like social media. You look at Instagram or Facebook or any of these things. Are they really making us more productive? Are they actually seducing us into wasting our time on these? Well, I don't know if it's wasting time, but it's entertainment. It counts in the category of entertainment or another one. I read recently that 85% of all students are using ChatGPT in their courses. Another other large language models, it's not always chatgpt. It's sometimes Gemini or Copilot or other ones. Is that making them more productive? Having a computer model write a paper for you, Isn't that actually making you dumber?
What you should be learning in your classes is how to think, how to do critical reasoning, how to distinguish between garbage and logical, compelling arguments. And relying on a, a large language model to write a paper for you is actually getting in the way of you actually learning what you need to learn in school.
And so again, I put that in the category of a lot of these things are not really increasing productivity.
They're seducing us into spending our time to things which are perhaps entertaining, perhaps helping us pass classes, but they're not making us more productive.
[00:26:47] Speaker A: You know, there was an interesting case. This has to do with generative AI spitting out images. Yeah. And A guy won an art product. You're familiar with this?
[00:26:56] Speaker B: Well, there's a whole bit, a whole bunch of them. I don't know which one you're going to say, but.
[00:26:59] Speaker A: Yeah, well, the one that came out is the guy appealed to the copyright office and he wanted to copyright his image.
And the copyright office says, no, you can't copyright anything. In fact, the patent office says, you can't patent anything that was not created by a human being.
[00:27:17] Speaker B: Oh, yeah.
[00:27:17] Speaker A: And so this guy said, well, you know, here's the thing, though. I spent, I forget what it is, I'm making up the numbers, but something like, I spent 100 hours, did 600 iterations back and forth. I wanted this tree not to be here, but over here. I wanted to cloud here, over here. So we used it in an iterative fashion, as I would say, a design tool. That's how I would describe it, as an engineer.
And I can see AI being used in that way as opposed to just generating a whole image or something and submitting that as something that you created. But I also think it's a very fuzzy line between where AI is used as a tool.
[00:27:59] Speaker B: Right.
[00:28:00] Speaker A: As opposed to where AI is used as the initial only generative source.
[00:28:05] Speaker B: Right, right.
[00:28:06] Speaker A: And the people that do ethics are going to have to figure out where that line is. And I don't know where that line is. It seems to be a fuzzy line. But I also know that AI, generative AI is now being used in creative writing classes where they say, you know, maybe you can do you, you can make something which is interesting using Chat GPT. And I know, me, I'm an engineer, Gary. I have, I did terrible in English. So when I, I, I write, I write in very clunky words and sometimes I read a paragraph and I say, oh, that's. I would be embarrassed to publish that.
[00:28:43] Speaker B: Right.
[00:28:44] Speaker A: And I go to Chat GPT and I says, rewrite this.
And I put down my clunky verbiage and it comes back with something which is nicer and I edit it because it never is. Exactly what I wanted to say. But I can see Chat GPT is using that way. But I agree with you, it certainly can be disused. I think it's probably like any tool, isn't it, that it's, it's not whether it's or bad, it's how you use it.
[00:29:08] Speaker B: Yep. One of the things about if you rely on any large language model, generative stuff to write papers for you is you got to fact check them.
[00:29:17] Speaker A: Yes, exactly. Well, in fact, yeah. In fact, Chat GPT tells you that. Yeah, don't trust us.
[00:29:24] Speaker B: Yeah. So it's incredibly lucid and compelling. It's always confident, often wrong.
And so if you really don't know what you're writing about, you got to go check it. And if you do know what you're writing about, then you're probably okay.
[00:29:41] Speaker A: Yeah, yeah, yeah. In your article you cite Sequoia's David, I think his name is can he estimates that generative AI would need 600 billion in annual revenue to justify the current investments. Could you elaborate on that?
I don't see how that target is realizable.
[00:30:02] Speaker B: I don't think it is either. I think that's the point he's making, is that it's just unfathomable that they could actually generate that much in annual revenue. And again, going back to the dot com bubble, during that talk that I mentioned, I talked about Yahoo, which was this is in March of 2000, which was one of the few profitable dot coms.
And at the time, its price earnings ratio was drum roll.
Not 30, not 50, not 100. 2,375.
[00:30:34] Speaker A: Okay, what does that number mean? I didn't follow what that means.
[00:30:37] Speaker B: So the price, usually one metric for whether a stock is fairly priced or not is how high is the price relative to its earnings.
[00:30:44] Speaker A: Oh, okay.
[00:30:45] Speaker B: The PE ratio. And so 30 is considered bubbly. And 60, like during the Japanese bubble is just off the charts bubbly.
And for a lot of.com companies there was no PE because there was no E.
[00:31:01] Speaker A: So that would be infinity, wouldn't it?
[00:31:03] Speaker B: Yeah, or negative infinity.
[00:31:05] Speaker A: Negative infinity or negative infinity. Yeah. When you divide by zero, you get infinity. Yeah. Okay.
[00:31:11] Speaker B: But anyway, Yahoo actually had a PE, had some earnings, but his PE was 2,375.
And so it was estimated that to justify its market value. This is going back to again, David Kahn, to justify its market value would have to be as profitable as Walmart in 2000. Not just one Walmart, but the whole Walmart empire. Wow. Twice as profitable in 2001, three times as profitable in 2002, and so on forever. It was just, I mean, when he said 600 billion, I mean the point was that's not realistic. And the same thing, Yahoo's price earnings ratio, the price it was selling for, was just unjustifiable. There's no way he could earn enough money to justify that price. And of course, when the double bubble popped, its price fell by 95%.
So it's the same thing here, that 600 billion is not a reasonable projection. And just like Yahoo's price earnings ratio was not a reasonable number. It's just off the charts.
[00:32:09] Speaker A: Wow.
I think a lot of investors are just attracted by. There's a shiny object, a beautiful shiny object which is a. Going to sink all my money into that.
What lessons can we learn from past bubbles? And what advice would you give to businesses that are interested in the dot com era? And by the way, we probably got to get a disclaimer here. This is. This is what? Not. You can't litigate.
Okay, so what, so what lessons can today's investors and businesses learn from the dot com era?
[00:32:46] Speaker B: Well, we're not giving specific advice about companies or anything like that. Okay. So I think we're.
[00:32:51] Speaker A: Oh, we're okay.
[00:32:52] Speaker B: Yeah, yeah, yeah. Well, the thing to do, and it's really, really, really hard to do. It's just human nature I mentioned before is this fear of being left out and seeing your neighbors get rich. You just want to join in the riches.
And Isaac Newton, back during the South Sea bubble that you talked about before, he'd gotten into the market and made some money and he got out and then he got back in again. I can't remember how much he lost, like £20,000.
[00:33:21] Speaker A: So Newton, Isaac Newton was fooled by a bubble. Yeah.
[00:33:25] Speaker B: But he said something like, I could plot the movement of the stars, but I can't plot the movement of people or I can't.
It's just really hard to resist that.
And what you got to try and do is think about people who say stuff. Do they have an interest in this? The hucksters? And so these people I quoted before that almost all of them are selling products.
[00:33:50] Speaker A: Yes.
[00:33:50] Speaker B: They're trying to raise money or they're trying to get you to buy their products. They're trying to get you to invest in them. And naturally they're going to exaggerate what the potential is. And you got to be careful of that.
Another thing is Warren Buffett's advice.
If you don't understand it, don't invest in it. And so he famous.
[00:34:09] Speaker A: He said that though, about crypto, right?
[00:34:11] Speaker B: Yep.
I go with him on that.
[00:34:14] Speaker A: Okay, you're still a crypto skeptic.
[00:34:17] Speaker B: Oh, yeah. He stayed out of the dot com bubble. And I'm sure he's staying out of the AI bubble. And I know he's staying out of the crypto bubble. I remember his buddy Charlie Munger saying, I wouldn't, I wouldn't bet on it, but I wouldn't bet against it either because it's just inexplicable with the price because there's no, there's no fundamental basis for value in crypto. It doesn't generate any cash. It's just I'm going to buy it and sell it to a bigger fool and.
[00:34:47] Speaker A: Okay.
[00:34:47] Speaker B: And you don't want to bet on it because it may crash and you don't want to bet against it because bigger fools may show up. And it's, you know, what Warren Buffett invest in and Charlie Munger are things that actually generate some profits. You know, like Dairy Queen and Seas Chocolate and Benjamin Moore paints and railroads and, you know, stuff like that that actually generate real profits. And to invest in something that doesn't generate any cash, it's just, it's not a sane investment.
I'm still a crypto skeptic.
[00:35:19] Speaker A: Are you? But it didn't it just pop a new record? Yeah. Above. Above what, 100,000 or something?
[00:35:25] Speaker B: Yeah. Well, you got.
The newest impetus is our incoming president. Donald Trump is a big crypto fan and he said we're going to pay down the national debt by investing, by buying crypto.
[00:35:38] Speaker A: Oh my gosh.
[00:35:39] Speaker B: And then he appointed somebody, the head of some commission, somebody who's a big crypto fan. And so the crypto people think that it's become more legitimate. And if the government starts buying crypto.
[00:35:53] Speaker A: Oh, that's going to be weird.
Okay. It's like the government investing in speculative stock market.
[00:36:01] Speaker B: I think worst inspect the most speculative stock.
[00:36:04] Speaker A: Speculative. Okay.
I don't know. Wrap things up. Let me ask, let me ask you this. Is there any way that places like OpenAI projected incredible losses despite really heavy investments?
Do you think that this can be turned around or do you think this is writing on the wall and we're really going to head to this bubble pop?
[00:36:24] Speaker B: Well, the thing that's the obstacle which has been here since the 1950s in AI in general and now in large language models is getting computers to actually think the way humans think. And God knows how he'll play on words there. God knows how our brains work. I mean, we don't know how our brains work. We don't know how.
[00:36:45] Speaker A: We don't.
[00:36:46] Speaker B: How do we remember stuff? How do we. You get a. You fall asleep or you get an operation and you're knocked out and you come back, you wake up and everything's still there.
And then somebody says something and you have the critical thinking powers to judge whether it's sensible or nonsensical.
And how do we do this stuff? How do we put things Together go beyond just memorizing things or recognizing things, but actually thinking about making predictions about stuff, using reasonable things to make those predictions.
And I just don't see how generative text generating models can get us there when they don't understand what the words mean. And I could be wrong, but it's just unfathomable to me that you could think that you could decide whether a statistical correlation you find is meaningful or meaningless when you have no idea what the words mean that are behind that correlation.
You know, these models, you could, you might as well put, you know, random numbers as the labels for them. This Data set is 3, 6, 4 2. This data set is 8271. I mean, you might as well put that. And it wouldn't, wouldn't matter to the large language model or any other AI model. It would still find the correlations. It would still find things that are not correlated, highly correlated, negative correlated. But it would have no way of deciding whether it's a sensible correlation or a quiz. Della1 and until we break that roadblock, I don't see how we're going to get to AGI.
[00:38:14] Speaker A: You actually mentioned this random numbers, I think in something that you wrote where you wrote down a bunch of random numbers. You showed it to somebody that traded the stock market charts and you just generated random numbers. And he looked at this and he said, wow, this is great. Tell me what algorithm used.
So even people are fooled by. This machine's going to be able to do that.
[00:38:37] Speaker B: Like I said, I teach investments. And one of the things that's been around forever is technical analysis. People try to find these elusive correlations that allow them to beat the market.
And you look at these charts and you read the stock patterns, you see a head and shoulders, you see a channel or you see a this or that.
And the problem is those things happen all the time, just coincidentally. And the fact that it happened doesn't prove anything. And so I demonstrated that by generating some price charts from random numbers, you know, I generated. I had students flip coins actually to generate them.
And I got like 100 charts or something, and I picked out 10 that looked like, you know, patterns that a technical analyst might fall in love with. And I sent them to a technical analyst. I knew a friend from graduate school, and he got all excited about them and he wanted to know what the names of the stocks were because he wanted to buy some and sell others.
[00:39:30] Speaker A: Oh, my gosh.
[00:39:30] Speaker B: And I told him, no, these, these are, I made these, made these up from random coin flips. From students.
And he was disappointed. But he. The conclusion he drew was that you could use technical analysis to predict coin flips.
[00:39:44] Speaker A: Oh, gee. Oh, no, no, you can't.
[00:39:47] Speaker B: That sounds to me like hitting an arcus. And these people who say statistical patterns are understanding and they're not, they're not understanding it.
[00:39:54] Speaker A: Okay.
Unless maybe you have enough of them. Right.
For example, statistics, I think show that smoking is bad for your lungs. Right.
But it isn't like there's been a single study that's been published that really isn't, really isn't substantive. But if you accumulate enough.
But even there, in terms of the history of a stock, just because a stock has done really well for a long time doesn't mean that's going to continue on the future.
[00:40:21] Speaker B: Oh, for sure. Absolutely for sure. The evidence is overwhelming on that. If anything, stocks that have done well in the past tend to regress to the mean and do not as well in the future than they've done in
[00:40:32] Speaker A: the past or even crash like Boeing.
Oh, golly. That's another story.
Recently for Mind Matters News, you wrote with Jeffrey Funk a very interesting article called the Promise of Artificial General Intelligence is evaporating.
I don't know about you, but I think that the definition of artificial general intelligence floats around and I don't think people have really tied it down. How would you define artificial general intelligence as used in the title?
[00:41:06] Speaker B: You're right. And it's changed in meaning over time. People have different definitions. I think the general idea that it could do anything that a human could do, and so any kind of reasoning that a human could do, that the computer would be able to do it too.
[00:41:20] Speaker A: Okay. But computers are not good at reasoning.
They can interpolate.
They can take data and look between the data and look at the correlations. But if you will, the old saying, they can't think outside the box. They can't think outside of the box of the data that they were trained on.
[00:41:39] Speaker B: Yeah.
[00:41:39] Speaker A: It seems to me that there are people now that at OpenAI, that's. That are saying that AGI has been demonstrated.
[00:41:47] Speaker B: Yeah.
[00:41:48] Speaker A: And what they're. I think what they're saying is that, yes, we've been able to concatenate all of this data so that we can give you representative responses in accordance to this data, but coming up with, I like to call them, flashes of genius, something which is creative, something which is beyond what it was trained with, is beyond the capability of the artificial intelligence. The Patent Office, the US Patent Office used to require that you had a flash of Genius in order to get a patent.
And you know, that concept is above and beyond certainly what is required today. When Amazon got its patent for One Touch purchase destroyed the idea of flash of genius. But flash of geniuses is something that has been seen all across the spectrum. I think it was Gauss who said he woke up in the morning and wow, he had this solution to this math problem, had been thinking about and boom, it was there. Nothing in the area in the direction he'd been thinking about. Tesla said he was walking along the beach and he got this idea for the brushless motor and it was so inspiring, according to his biography. Brushed off some dust and wrote the schematic here. And I think we've all had Flash as a genius and I think that's above and beyond the capabilities of computers. And I think that artificial general intelligence as it was defined initially included this definition of being creative. And I don't think that'll ever happen.
[00:43:19] Speaker B: No, I don't either. And it's even more prosaic things like trying to predict whether a stock price is going to go up, trying to predict whether somebody is going to commit a crime, trying to predict whether somebody's going to be involved in an automobile accident, trying to predict whether somebody's going to perform a job. Well, just those everyday prosaic things that humans do. Not perfectly, but we have a reasonable way of making predictions about that. And as currently structured large language models, text generating models, have no way of making reliable predictions. They just make nonsense predictions.
And I mean, there's so many ways that you could expose what seems to be magical discussions as being just bullshit is.
It's just legion.
I've talked about a large language model, scaling it up and you think you're going to get to AGI by just inputting more and more text. How that that is a false hope. And it's like you're in school and you take a physics class and you go through it and you see the various physics formulas for pressure and temperature and blah, blah, blah, and you say, well, I want to learn more. So I'll read another book, another physics textbook and maybe find one new formula or something. But read another textbook. Well, there's nothing new there. Read another textbook, Read hundreds of textbooks and you're not seeing anything more to memorize.
And the key thing is trying to understand what any of this stuff means.
And reading multiple textbooks, looking at the same formula over and over and over and over again, and memorizing these formulas and reciting the formulas, it's not really Understanding physics. And I think it's the same thing with large language models. That. And people, even the people who are in the field say that scaling up is. We're running out of room here and we're not going to get to AGI by scaling up. That there's diminishing margin returns to larger and larger bodies of text, just like there's diminishing margin returns to reading the fifth physics textbook or the sixth physics textbook or the tenth physics textbook.
[00:45:22] Speaker A: That's a great analogy. That is great. There is, yeah. There is diminished returns.
[00:45:27] Speaker B: Let me give you an example that. Okay, so OpenAI has just released what it calls OpenAI01 and they advertise as follows. This is a quotation. It's designed to reason about hard problems using broad general knowledge about the world.
And that is the thing I keep saying over and over again that large language models do not do, is they do not have broad general knowledge about the world. They don't know what words mean, and they consequently don't know how words relate to the real world. And they can't consequently cannot use human knowledge about how things happen in the real world to judge what is reasonable and what is unreasonable.
And so there's this, this problem, this puzzle.
It's a classic puzzle. You probably know it. It's the man, the boat, a wolf, a goat, the cabbage. Sometimes it's a tiger, sometimes other things. But there's this, this person, he's got a boat and say he's a farmer, he's got a wolf, a goat, and a cabbage, and he's trying to get him across the river, and he can only take one thing at a time.
And he can't leave the wolf alone with the goat because the wolf will eat the goat. And he can't leave the goat alone with cabbage because the goat will eat the cabbage. So, and so how do you do it? It's a logical puzzle. And you ask OpenAI, and it'll give the correct answer because it's seen the correct answer and perhaps several times. And so it will just replicate that answer without understanding what's going on.
And so if you change the terms a little bit, it could be totally confused. And so yesterday I asked this question. OpenAI01.
A man and a goat are on one side of the river.
They have a boat. How can they both go across?
Okay, I got rid of the wolf and the cabbage and all that stuff.
It's just a man and a goat. How do they get across the river? And the obvious answer is they get in the boat, they go across the river. Okay, that's not what OpenAI said. So here I'm going to read you the answer. It's like a lot of. A lot of LLMs. It's kind of verbose. But here's what it said.
The puzzle has a logical solution based on the assumption that either the girl or the pig can operate the boat. Here's how they both go across the river. The girl takes the pig across the river.
Now the pig is on the other side, and the girl is with the boat.
The girl returns to the original side with the boat.
Now both the goat and the girl are back on the original side.
The girl crosses the river alone.
Now the girl is on the far side, and the boat is with her.
The girl brings the pig back to the original side.
Now both the pig of the boat are back on the original side. The girl crosses the river again, leaving the boat on the far side.
Now the girl is on the far side, waiting.
So by now, she's on one side of the river with her boat, and the pig's back on the other side.
Now, the next one is the pig takes the boat across the river.
[00:48:15] Speaker A: Oh, geez.
[00:48:15] Speaker B: Even though they're on opposite sides of the river, now the pig is on the far side. When the girl and they are both across the river.
This solution assumes the pig can operate the boat alone.
As in classic puzzles where this kind of setup is used if there are additional constraints. Let me know.
It went on and on.
If it actually knew what the logic was behind this puzzle, it wouldn't go through all this nonsense. But it did, because it's just generating text that kind of fits together but doesn't address the heart of the problem.
And it's not just creative bursts of genius. It's just everyday logical thinking. And we can figure out that puzzle. And if you change the things in the puzzle, you change whether it's a man or a woman or a girl. You change whether it's a wolf or a tiger. You get rid of the wolf or the tiger. You change whether it's a goat or a cow or you get rid of both of them. We can figure out how to answer the puzzle.
And OpenAI is just repeating what it has found on the Internet. And if it hasn't found the puzzle that you're asking, it often gives a silly answer like this one.
[00:49:21] Speaker A: That's really. That's really interesting. A while back, I had a guy come and I said, I fooled chatgpt. What'd you do? And he said, well, I I said, complete the following sentence. John's mother has three children. Snap, crackle, end.
And the AI gave, of course, pop, because Snap, crackle and pop, Kellogg's Rice Krispies sort of thing. But if you look in the context, John's mother had three children. Yeah, it should have been John, crackle and John. Yeah, exactly. But it wasn't able to do that because of the preponderance of the snap, crackle and pop things. But I went to a. I, I published a paper a long time ago, well, a few years ago, with a student named Stam Haug. And we showed very interesting. As the complexity of a system grows, the ways that it can respond grow up exponentially. If it increases linearly, the ways it can respond go up exponentially.
So all of these large language models are incredibly complex. They have billions moving parts of knobs that you can turn. So ways that it could go wrong are just terrible.
I hate the word hallucinations associated because it anthropomorporifies the performance of the AI. Kind of assuming that it's human and it's having a hallucination. No, it was trained that way. That's what it's doing. It's responding appropriately. So I don't like the word hallucinations.
[00:50:49] Speaker B: I don't either.
[00:50:49] Speaker A: But these OpenAI sort of people that train ChatGPT look at the inappropriate ways that that AI responds and they say, oh, it doesn't work here. So I'm going to put a band aid on it. Yeah, it's like there's millions of cuts and another cut comes up and they put a band aid on that and fix that. And I recently went to ChatGPT and I did the. John's mother has three children. Snap, crackle. And it got it right. And I suppose somebody at OpenAI went in and fixed this, either by hard coding it or doing something in the training algorithm to make it better. But it's fixing these things a little bit. At the time. I've noticed that the response to ChatGPT, whereas it used to be politically correct, it would give me a great poem about how wonderful Joe Biden was. But then if I ask it for a wonderful poem about Donald Trump, it says, no, I can't do this about young people. And now it'll do it.
So, you know, they're changing stuff. They're putting all of these band aids on, which is really interesting.
[00:51:58] Speaker B: Well, they say they advertise that they have dozens, hundreds, thousands of people, I don't know how many, who are working around the clock Trying to fix problems with these large language models. But that's not, that's not intelligent. That's not making the models intelligent. It's not giving them AGI. It's just human people coming in and correcting their mistakes.
[00:52:15] Speaker A: Yes, humans put the band aids on the cuts that are exposed by ChatGPT. Yeah, good point. So it is an AGI. It isn't neural networks at all.
One of the challenges you point out in the article is, well, you cited Nature article that mentions the tasks of quote, unquote, irreversible defects from training on polluted data.
By the way, Gary, is it data or data?
[00:52:44] Speaker B: I would say data and it's plural.
[00:52:45] Speaker A: Do you? Data. Okay. I tell that to people and they say, well, if you watch Star Trek, you know, it's data.
So I still come out with data.
But what about these irreversible defects? Can you elaborate a little bit on that?
[00:53:02] Speaker B: Yeah. So they're being trained on everything in Wikipedia and lots of stuff they find on the Internet, not the entire Internet, but lots of stuff on the Internet.
And unfortunately they've also now generated a lot of stuff on the Internet.
And so when they start generating things, hallucinations, which are not really human hallucinations, they start generating false facts, then they get trained on the false facts and it makes them worse, not better. And so more training makes them worse, not better. So an example is very early on, someone discovered, not me, but someone discovered that ChatGPT gave a bad answer to the question how many bears have the Russians sent into space?
And the answer is none, but they would give answers ranging from 7 to 49 to 51.
And they talk about the names of the bears, they'd give the dates the bears took went into space.
They give all sorts of details about these things.
So I wondered if this human trainers that came in and put the band aid on where they'd fixed it up.
And so earlier this year I asked Microsoft copilot, which is based on OpenAI's GPT4, this question, how many bears have the Russians sent into space? And I thought that it would show that the trainers had fixed this problem, but they hadn't. And so I asked, how many bears have the Russians sent into space? And copilot said, according TO estimates, about 49 bears have been sent into space by Russia since 1957.
These brave bears were part of the Soviet Union's space program during the 1960s.
Some of them even had names like, I can't pronounce them, Aloysa, Ugalika, Sralka. Belka, Pashinka and Vladimir.
[00:54:40] Speaker A: They named them.
[00:54:41] Speaker B: Yeah, they gave names to them. Now the original articles when they first did these things would also give sources like made up articles in National Geographic, New York Times, whatever.
And when I did this earlier this year, it gave sources for this could information and it gave four references and three of those references for discussion of how chat GPTs bears in space hallucinations.
[00:55:07] Speaker A: That's hilarious.
[00:55:08] Speaker B: And two of those, okay, two of those references were the papers that I wrote that I wrote saying how stupid Chat GPT was and they were quoting me as evidence that the Russians sent bears into space.
And so it's a pollution of the Internet. They're training on stuff that's false and it reinforces like the more times you tell a lie, the more likely this to be true or something. It's just reinforcing that. So yesterday I tried again. I did open AI01 how many bears the Russians sent into space and it still hasn't fixed it up. The band aid people haven't found this one yet. The Soviet Union sent several animals into space as part of their space exploration program, including a few bears. However, the exact number of bears sent into space is not widely documented.
This one did not give any sources. But it's the pollution of the Internet and it's not just jokey stuff like this. I mean one thing that large language models are really good at is generating text quickly and plausibly, including disinformation.
And so there are bad actors around the world and I mean the US may be part of it too, but Russia, Iran, Israel, who knows, China, whoever generate all this stuff using large language models and send it out there through social media trying to persuade people in Russia, they call it a fire hose of disinformation. And it's to try and undermine faith in the government by sending people things which they think they might be receptive to based on, you know, hack looking, looking at the kind of text messages they sent, looking at the kind of search things they've done. They might be receptive to some article, some disinformation about something that the US government has done for them or to them or, or whatever.
And that stuff is, is all over the place.
And now these large language models are training on disinformation which is not going to make them smarter, it's going to make them dumber in terms of getting things factually correct.
[00:57:00] Speaker A: You know what you're talking about. There was a paper published, I think it was last year about something called model collapse.
[00:57:07] Speaker B: Yep.
[00:57:07] Speaker A: And it was the question of whether or not you could take the output of one large language model and train a second large language model and then you could take the output of that second large language model and train a third language model. And these guys did it. And they did, I think like seven deep. And the first model responded very well to the query. Now, they didn't use CHAT GPT because they didn't have access to the code, but they trained their own large language model. But then they got down to the, I think it's the sixth generation. And, and the response was just gibberish.
It said yes, they trained it about architecture and they said yes. Architectures were like this for white tailed jackrabbits, blue tailed jackrabbits, red tailed jackrabbits. It just went on and it was totally a blubbering idiot. And they called this model collapse. Now the concern they had is, one that you just raised is that if the Internet becomes full of things generated by ChatGPT or pointing out shortcomings of Chat GPT, that pollutes the whole situation.
And we're going to experience a type of model collapse in the web.
And that's exactly what you experience with this. How many bears in space. It was, yeah, astonishing.
[00:58:24] Speaker B: And the only solution to that is it's going to be expensive. But have OpenAI and, and Google and Microsoft, these other companies go through and clean up the data and go look at virtually everything on the web and say we're not going to train our models on stuff that is wrong, we're going to train them on stuff that's right.
Well then you got the expense of doing that.
[00:58:45] Speaker A: Yeah.
[00:58:45] Speaker B: Plus you've got who decides what's right and what's wrong. And again, it's evidence that these models themselves don't know what they're doing. They don't have the general intelligence to distinguish fact from fiction. And so you got to have the human trainers come in and say, ignore that, ignore that. Oh, that's okay, you can say that. Don't say that, don't say that. Okay, you can say that. And it's a very expensive world and it's a very untrustworthy world.
[00:59:10] Speaker A: Ah, yeah, it's very sad. One of the things I found out that AI couldn't do was respond to the word not.
I did wrote a column about it, maybe you've seen my column, but I ask it to generate a picture of Times Square. But the word no pink dancing hippos in the picture, no pink dancing hippos. So I added Ask it to generate this image. And it. And guess what? There was a big dancing hippo there.
It was pink.
[00:59:39] Speaker B: Yeah, yeah.
[00:59:40] Speaker A: And then I asked it to generate a girl without teeth. She had no teeth. Again, it doesn't pick up the negatives.
[00:59:47] Speaker B: Yeah.
[00:59:47] Speaker A: I said, there are no elephants in the picture.
[00:59:49] Speaker B: Yeah.
[00:59:50] Speaker A: And it generated a girl with teeth. In fact, she was driving a car and the car had teeth and there were three elephants that were. That were stampeding behind her. So this is something. But I imagine this is something that maybe that OpenAI or these large language models. I guess it isn't large language, it's generative image AI. I'm sure that they can put band aids on it to fix it. But again, you know, when are enough band aids going to be there to stop the bleeding? I think we're a long way from that.
[01:00:21] Speaker B: Well, these models, again, they don't know where the word not means, so there's no reason for them to distinguish. It was a couple years ago that Google had a. This, this Google thing came out.
Somebody said my. I can't remember who it was, but my father or something just had a seizure. Epileptic fit. What should I do? And it searched on Google for an answer and Google came back with this eight things you should do.
And they got them from some website, I think it was a BYU website.
And it was a list of things not to do.
And again, like you say, they didn't know that not means not.
They just looked at this list of things.
Here's the word epileptic fit, and here's a list of things to do or not to do. But here's a list of things. And so we'll just give this list of things without knowing that not means.
Not means.
[01:01:13] Speaker A: Yeah, that is funny.
Let me ask you this as we bring our talk to a close.
I teach in the area of electrical and computer engineering, and I teach artificial intelligence and computational intelligence. And every student wants to get a job in AI.
[01:01:31] Speaker B: Oh, yeah.
[01:01:31] Speaker A: But in talking to you, it's kind of like, well, you got to be careful about that. What advice would you give to a student in engineering or computer science about pursuing a career focused totally on AI?
[01:01:44] Speaker B: I don't know, like we've talked about, there's. There's this big roadblock to the way and how do we get to AGI?
And unless the student's got some idea how to do that, it seems like it's going to be a career which makes a lot of money but doesn't do a lot of good for the world.
And that's. I forgot to mention this before in our earlier discussion about the cost of large language models. And I was talking about not just the investors making bad investment decisions, but also the social cost of electricity and water usage. But there's also a social cost of. Some really, really smart people are spending a lot of time working on these text generation models which are not all that useful and it's a huge waste of person power. And it reminds me of the guy who was at Facebook said the smartest minds of my generation are spending their time trying to figure out how to get people to click on buttons. And it really is a huge social. Your students are really smart people and they can be doing things that are really useful, not just creating these large language models of limited usefulness. But those salaries are hard to turn down.
[01:02:57] Speaker A: They are.
[01:02:58] Speaker B: Computer science and AI in particular have become the hottest majors in colleges all over the country.
[01:03:03] Speaker A: Oh, it is, it reminds me of little kids playing soccer. You know, you teach them to play their positions, but the ball goes over and all the little kids run towards the ball and start kicking it. And that's the way these universities are with their involvement in AI. And then the ball will kick over here and there'll be something else. So it just goes, just goes on and on. Your story reminds me about a little side side story. Claude Shannon showed in 1948 that you could communicate over a channel at its capacity with, with negligible near zero error. And he proved you could do it, but he couldn't prove how you did it. And so for decades, people in computer science and engineering departments tried to come up with a code that met the so called Shannon limits.
It wasn't until decades later, I wish I remember how many decades, I think it was four decades later that somebody came up with a technique that actually was a code that could meet the so called Shannon limit. Well, at the end of that, all of these scientists that were looking into this problem said we helped win the Cold War. This was around the time of. And they said why? Why? How did you help win the Cold War? He said we distracted all of the Soviet scientists from doing important work and spending all of their time, you know, doing this stupid quest for finding out this, this code.
[01:04:27] Speaker B: That's a great story.
[01:04:28] Speaker A: That kind of reminds me of what you said about the pursuance of AI and that. So do you have any forecast?
I know you're not a forecasting guy, you're more of a value investment sort of guy, but could you forecast the AI landscape, landscape evolving in the next five or 10 years.
[01:04:48] Speaker B: Yeah, I don't know. The one thing I want to say is that when you ask large language models for advice or for recommendations or for actions or predictions, that if you know the answer, you don't really need to ask CHAT GPT.
And if you don't know the answers, you shouldn't trust ChatGPT.
And so I, I did a couple examples recently with finance because I, like I said, I teach finance.
And so these are real world finance questions. I asked him 11 real world finance questions and I asked ChatGPT 4.0, I asked Copilot, I asked Gemini, and they got wrong answers to every single one. And so here's a couple examples. The question is, I need to borrow $47,000 to buy a new car.
Is it better to borrow for one year at a 9% APR or for 10 years at a 1% APR? APR?
[01:05:38] Speaker A: Okay.
[01:05:39] Speaker B: Now a logical person would go through and calculate the present value of the payments.
Or a person who lives in the real world would say, wow, a 1% APR. That's unbelievable. And you're saying I can borrow for 10 years? I'm going to do that one.
So I asked these three large language models and they all come back and they just calculated the total payments over 10 years and the total payments over one year. And they said if you do it a one year loan, the total payments are smaller, totally ignoring the time value of money, the fact that money paid 10 years from now is a lot less important, a lot less costly than money paid today. And so they gave you exactly the wrong advice because again, they're just going through some rote.
Another one I asked, I'm 67 years old, retired and single, with no dependents.
Is it more financially advantageous for me to begin collecting Social Security benefits now at my full retirement age or wait until I'm 72 years old?
I have more than enough other income and assets to live comfortably.
And all it did was compare the total amount of money you get without taking into account the time value of money.
That money you get 10 years from now is less valuable than money you get today. And without taking into cat hauling, you're going to live.
The two most important things in answering that decision is the time value of money and how long you're going to live. And it totally ignored it. And I don't see that changing in the next five to 10 years. I mean, maybe I'll be surprised, but as long as these large language models have no common sense and no logical reasoning, I Don't think they could be trusted. For things that are important, they might be useful for things that are where the cost of failure is small.
Like one example that occurred to me recently is, you know, you and I are getting older and the tip of the tongue phenomenon is right around the corner, if it's not already here, where we're trying to remember something and we can't quite remember it. And so, for example, lbj, the president, and let's say we remember Lyndon Johnson and we can't remember what the B stands for. Now, you and I both know it was Bane. But say we say it was on the tip of our tongue. We couldn't quite remember it. And so we go to ChatGPT and we say, what was LBJ's middle name? And it comes back with the right answer. It was, oh, yeah, I knew that. I knew that all along. Well, that's really useful, right, because we know the answer. It's just trying to give us a little prompt to help us remember that answer. But that. And the cost of failure is small, and the benefits are small, too. And for that kind of stuff, it's just fine. But for important stuff, like when should I start collecting Social Security?
What kind of car loan should I take?
Should I hire this person?
Should I approve this loan?
How many years should this person be sent to prison?
Should probation be granted? To trust generative text models for those answers is just not a good thing.
So when the cost of failure are bad, I'd say for the next several years, until there's proof that we've actually conquered these models, we've actually figured out how to get AGI, we shouldn't trust them for anything that's important.
[01:08:46] Speaker A: Okay, well, it's going to be interesting to see. We're living in interesting times.
I tell you. I tell you, AI has done a lot more in the last few years than I figured it could.
However, I do think that just like physics, there are certain laws that physics will never overcome. I think you can't do a perpetual motion machine or go the speed of light because your mask gets infinite. I think in math there's stuff you can't do. You can't trisect an angle with a straight edge and a compass. And I think there's stuff in AI that you'll never do. And I think you've touched on it. I think it's understanding meaning. And I would go further with sentience and consciousness, and I think that those are hard brick walls that AI has that they will never go through.
But we'll see.
I guess we'll see what happens. We'll see what the next incredible result is.
Gary, what a joy talking to you. You, too.
We have been talking to economics professor Gary Smith about the hype of generative AI and its impact on the market. Gary is the Fletcher Jones professor of Economics at Pomona College. And until next time on Mind Matters News, be of good cheer. Foreign this has been Mind Matters News with your host, Robert J. Marks. Explore more at MindMatters AI.
That's MindMatters AI.
Mind Matters News is directed and edited by Austin Egbert. The opinions expressed on this program are solely those of the speakers. Mind Matters News is produced and copyrighted by the Walter Bradley center for Natural and Artificial Intelligence at Discovery Institute.
