Botond Seres, Dave Erickson - Guest: Noah Healy
Intro: Dave Erickson 00:03
How can you use AI to exploit global marketplaces and is there really a better way to trade products and services? We're gonna look at whether AI is the future of marketplaces or the end of humanity. Do we currently have the best global marketplace mechanics or is there something better on the horizon? We're going to discuss AI and marketplaces on this month's ScreamingBox podcast. Please like our podcast and subscribe to our channel to get notified of when next month's podcast is released.Is artificial intelligence or AI going to take over the world? Are humans becoming redundant? What will AI do to marketplaces? How will you be able to use AI to dominate global markets?
Dave Erickson 00:56
Welcome to the ScreamingBox technology and business rundown podcast. For this month I, Dave Erickson and my co hosts Botond Seres are going to try to peer into the Looking Glass to try to see our future with AI. By talking with Noah Healy, CEO of Core Disk. Early in his career, he used computational mathematics and software development capabilities to develop automations that boost internal productivity, drive client efficiencies, and improve regulatory compliance. Most recently, he founded Core Disc and developed coordinated discovery markets, also known as CDM, a technology product that has the potential to completely reorder the global financial system. CDM establishes a better exchange market for commodities by conducting all trades at market clearing prices with radically improved price discovery, lower total cost of transactions and the elimination of hedging needs. Today, we're going to delve into AI and the future of marketplaces and try to figure out if AI can make a better world that will hopefully still include humans in it. So Noah, did I get most of that, right?
Noah Healy 02:23
Yes, you did. That, that is the principal challenge today. The computers and AI and related technologies are the most exciting and impactful things that we have. And sadly, currently we're not really generating those things in ways that are conductive to having a functioning civilization.
Dave Erickson 02:47
And we all like a functioning civilization.
Noah Healy 02:51
Most of us depend on it for life. Yes.
Dave Erickson 02:54
CDM coordinate discovery market seems to be a core to some of the things that you're doing. Maybe you can talk a little bit more about what is a coordinated discovery market and how does it work?
Noah Healy 03:07
Absolutely. So the principal job of the marketplace is price discovery, that is figuring out what the deal structure that will cause interest on both sides of the deal to match. We have a complex civilization, many, many people are in productive roles, many, many people need the things that are coming out of these productive roles. And so by unifying the process of that selling pipeline, we save an enormous amount of time and effort and cost by basically creating these, these public sales funnels that create efficient trade mechanisms. What CDM does is it changes that process from an emergent process where essentially lots of people are putting out offers to buy and sell in the hopes of ensnaring a counterparty. And right now, something like 97-98% of the people doing this have no intention or even ability of buying or selling the underlying product, they're simply engaged in trying to find connections. So essentially, what they're trying to do is create two deals, one with a buyer and one with a seller, which they will bridge and they'll take the profits off the differences between how wide the gap they can form between those buyers and sellers actually are. In the CDM we directly publish and, and work towards a common deal that people can participate in. So rather than people basically trying to find the most desperate counterparties possible to get the best deals for themselves instead, they use their market knowledge to advertise that market knowledge through the CDM and take on a share of the total market volume that winds up being transacted through the CDM. So instead of kind of making your money by finding the most desperate stragglers that there are, you make your money by knowing where the deal points are going to wind up, letting everybody know that as far in advance as possible and then taking a smaller piece of many, many more deals than you're actually capable of engaging in yourself through the existing sort of picking off the suckers model.
Dave Erickson 05:53
Maybe we can use some kind of example of how that would work. So let me see if I got this right. Let's say we're in a common market and I'm looking to buy wheat and I'm basically going to go around talking to a bunch of different people, How much is your price for wheat? until I find somebody who, for whatever reason, is desperate to sell their wheat and so I can get a lower price. And then I'll buy it at that low price for that time. Sure. That's kind of how the current market works (Noah Healy): that is how the current market works, yeah), How would a CDM work? How would CDM work using that kind of example?
Noah Healy 06:30
Right. So imagine this common market place, instead of being set up like that, where people are going to wander around, bump into each other, and you're sort of going to, you know, try to find your mall parking space, right, you're just going to look and try to find the best one. And maybe you'll skip the best one because you think you can do better and you wind up you can't. Well, what the CDM does is it says, Okay, we're going to do this tomorrow, we're going to do this next week, we're going to be doing this, you know, every day for decades. So what we're going to do is, for every single one of those days, we're going to publish what the price of wheat is going to be on that day. Now, it's a guess, we don't know what the price is next month yet, we don't know what the price is next year yet. But all of those guesses are sitting out there and you get to see that before you even show up at the marketplace. And the other thing you get to do is you get to negotiate for those future prices, you can go in and say Okay, I'm not buying wheat right now, but I'm gonna be buying wheat next week, or I know something about what's going to happen next month. So I'm going to bid on those prices in the future and I'm going to change those prices so that I participate when those prices are used, if my changes are still sort of resident in there. And so what happens is when people show up, the price that's, that's being offered in the marketplace has already been pre negotiated for days or weeks or months or years. And everybody already knows what it is. And so there isn't any desperate seller in there. Every seller is selling the wheat for exactly the same deal every buyer is getting that everyone already knew before they even showed up in the first place.
Dave Erickson 08:15
Is there examples of this working? Has a CDM been put in place and tried?
Noah Healy 08:21
Sadly, no. There are about half a dozen people that are working on establishing markets that incorporate the technology. You can, you can, it's a mathematical tool. And so you can build out a mathematical model, basically using the language of calculus. Effectively, it's an integral equation function and so the function of those integral equations can be mathematically verified. And actually, that's how CDM was essentially designed. By, by, by extracting that analysis from these functions. So that's been done.
Dave Erickson 09:03
So what, what is it going to take to put CDM in place, and to get people using CDM instead of the common marketplace?
Noah Healy 09:13
The principal challenge is regulatory markets have to operate within a legal system and legal systems get to decide what the rules are for operating within them. So people with the wherewithal to comply with a legal system for a novel mechanism and then attract customers have to be found. It's a, it's a big lift. You need about 50 customers to get the ball rolling as it were. And you also need to comply with whatever the regulations are in your districts. That can be extremely costly, which is why I'm in a licensing kind of scenario. And of course it changes from country to country. So, so basically every one I'm dealing with is actually in a different country somewhere in the world doing a different kind of thing and looking for more and, and seeing who manages to get over the hump first.
Botond Seres 10:13
So Noah, negotiating of future prices, it sounds to me something very similar to the futures markets that already exist. So how would you differentiate?
Noah Healy 10:30
The futures markets once again, while they're not doing that trading and present tense, like from our example, are working exactly the same way where people are wandering around looking for that future counterparty to establish those future prices. And so the existence of the futures market is actually the starting point. The fact is that pre establishing and pre negotiating price is something that's so valuable to people who are operating in low margin businesses with reasonable risks, that they're willing to give up some fraction of their potential income in order to assure their sales funnel. So what this system does is essentially federates that desire and allows them to much more efficiently allocate the capital that they that they expend to get those future prices, which means that they have to spend a lot less so whereas the futures market cannot control its cost overhead overheads, and also can't control its reward structure, this market can do both of those things, which allows us to lower the cost and improve the rewards for accurate forecasting at the same time, so you get better prices for less money.
Botond Seres 11:48
That's, that's a bit confusing. But if I understand you correctly, it is very similar to the futures markets. However, it is very different, because we don't buy and don't pay for the option to buy at the exact price. Rather, you can begin the negotiation on what the price is going to be.
Noah Healy 12:11
So what you're going to do is, rather than buying an option for you to transact at a future price, you're going to negotiate, forecast, speculate, there isn't a great word for this, because it's brand new, a price point for the entire marketplace. And what the system will do is take your bid on board along with everyone else's and integrate all of those together into a consensus price. And the degree to which your price moved things towards consensus is the degree to which you're now in the pool to be rewarded for having made a positive contribution to the information in the marketplace. And the extent to which your information moved away from consensus. That's, that's just gone. You had to pay the money for that movement. But you're not going to get rewarded for sort of moving people away from wherever one is winding up. So that won't gain in return.
Dave Erickson 13:19
So you kind of reward people who are good at predicting where the price should be, because that's where the consensus, consensus ends up moving the price and so you're rewarded. But if you're really off, and you always are consistently off in where you think the price should be, you don't get a reward, you just get that price.
Noah Healy 13:43
So if you're off consistently, essentially what you're doing is you're paying the people who are good at these predictions to correct your errors. But since the people who are actually trading or trading at these, these consensuses that get arrived at iteratively. Even if you don't really have a good picture of where prices are going to wind up and you wind up wasting most of your speculative, you know, pennies, you're still going to get the common market price that everybody actually agreed on. When you go in to buy.
Dave Erickson 14:15
I guess the trick in that sense is in a CDM, the price is set for everybody. So nobody gets an advantage or a disadvantage. Everyone kind of gets the same price. Correct?
Noah Healy 14:29
Yeah, the, when you're just making and taking delivery, then there's no advantage to be had. You can, you could still differentiate yourself by being a good producer, and a good predictor and be able to beat your neighbor who's just as good at producing but not as good as predicting but, but we separate those two things into distinct buckets and then allow each of them to be separately rewarded and measured for expertise.
Dave Erickson 15:05
Well, you know, AI is really good at prediction models from what I understand.
Noah Healy 15:11
Potentially, yes, that's one of the, that's one of the considerations for generative CDM is that AI systems in existing marketplaces are actually a destabilizing force. Because when you're on that sort of wandering around looking for the most desperate person model, AIs are better at finding those desperate people, before they sort of come to their senses, and better at creating conditions to make people more desperate. So you can sort of show people a burning house and ask them if they want to be rescued, and, and, as soon as they sell you the house for fire sale prices, the, the illusion of fire that you've generated vanishes, and now you have a house that you bought for virtually nothing. So creating a system that rewards people based on how much better they're making the market operate is critical if we're going to keep having computer and AI systems available to people that are involved in the financial industry and I don't see a world where that's not going to happen. And I don't see any regulatory appetite for declaring that people in the financial industry may not have access to computational technology, so that, that can't happen. So we need to actually change the markets so that that technology doesn't destabilize them. Because that's what it currently does.
Botond Seres 16:44
There is one thing I can't quite grasp. If everyone has the same price, on the same day, what happens to the market depth? So when I go around and look at different markets, the first thing I check is, how do the buy and sell water stack up? Like, how far are people willing to go to acquire a certain product? Wouldn't a fixed daily price create extremely shallow markets, in a sense?
Noah Healy 17:24
Not really, because depth, while it's vitally important to the existing market design, is actually getting at the concept of what's actually going on in these predictions. And so by doing the predictions much more directly, depth actually turns into essentially the breadth of market information that's being provided to the system. So that is all now located inside the, the forecasting market that operates off of its successively higher priced pause and mutual integration process, that, that, that essentially takes over the entire task of information and knowledge processing for the marketplace, and makes existing models of depth, irrelevant.
Botond Seres 18:22
How about the possibility of collusion between predictors? So, I suppose many of these markets have extremely high data turnovers, such as the wheat market, for example. People need to keep the meals running at all times. So I will assume there is large amounts of money to be made by colluding and predicting either a much higher or a much lower price.
Noah Healy 18:49
Well, that's where things get interesting, because there actually is a lot of money to lose by predicting a much higher or much lower price, because the buyers and sellers are not, in fact, forced to utilize the marketplace. So if a collusive group manages to derange the market, what you actually do is suppress the trade through the marketplace. And what that does is one triggers criminal investigations, which are going to be very easy to prosecute, two, by driving the volume removes most of the value chain because you're getting essentially paid by the Commission's of the trades that occur. So by restricting the number of trades that occur, you're also restricting the commission stream that actually pays you off. Three, by reducing the size of the marketplace, you now actually make it cheaper to speculate within the marketplace. So people can come in and fix the mistakes that you put into the market very inexpensively. So the market essentially recovers from your nonsense almost immediately. So basically, It's an extremely expensive way to, to conduct yourself to jail to try to do that, if you're unfortunate enough to succeed, and success essentially requires you to out vote the broad marketplace. And in the short term, changing prices in the short term is extremely expensive, because basically people are about to use these things. However, not changing prices is very cheap. So as the negotiation sort of settles down, there becomes a dominating strategy to simply reinforce the current deal so that last minute players that decide to show up and try to queer the deal for everybody, effectively are simply transferring their money to sort of a protector class that only has to spend money, when somebody shows up to destroy everything.
Botond Seres 20:56
I mean, that sounds pretty good, in theory. I'll be honest with you, it sounds pretty amazing. But I cannot gloss over the fact that market manipulation has long been a part of our current systems and that is an integral part, part, whether we like it or not. Like, at the moment, we have many giants who do this for a living, I don't want to bring up any names. But I'm pretty sure everyone can mention at least one certain company, who is not called dark stone, but something very similar. And I'm pretty sure that independent actors are not going to have the financial power to overrule them. And that way, they would still get the rewards, even though they have completely maliciously modifying the marketplace.
Noah Healy 21:48
Well, the thing to understand is that there aren't rewards for maliciously altering the market price. That ,that when you maliciously alter the marketplace, you actually turn down the value. So you actually lose money by doing that. And two the, the amount that you can move the market, the amount that you can spend, is dependent on the amount that you intend to move the marketplace. So if you, if you want to spend the sorts of money that governments or extraordinarily large financial players could spend in order to manipulate the market, you would have to change prices in a completely insane degree. So for the wheat market, you'd have to go from, you know, sort of like $8 to $12 a bushel to like $10,000 a bushel in order to spend that sort of money. And again, that would, that would basically involve them cutting the check for 10s of billions of dollars, then having that money transferred to people at pennies on the $100 after a day of no trade because nobody's buying wheat at $10,000 a bushel. And once again, a very, very easy prosecution to make, as well as a very, very easy civil case to make, that they were clearly attempting to harm people by doing that, because there's no reason to say that it could have worked, and it didn't work. So it's not much of an issue and the reason it's not much of an issue is that the existing system because of its sort of murky emergent nature of price discovery means that manipulation is an intrinsic part of that system, there's, there's gains to be made. And so it's going to happen. And so we have regulations and a certain amount of winking going on to to cope with those things, but by doing things in a direct and open fashion, with a CDM, that all goes away and and these large players can basically destroy themselves through bankruptcy or just stay out of those markets, because they're not offering anything of value. And they are, they are the primary losers of the adoption of this technology.
Botond Seres 24:30
Can you help me understand how the market would correct itself in this case? Because I understand that Yeah, okay, everyone pays a bunch of money to move the price, no one, no one trades and then the next day, what happens? Does it like reset, or was it a slow, methodical process?
Noah Healy 24:53
So, so we have, yeah, so we have, we have a price go crazy and so trade restricts a great deal. So what the CDM says is it says, Okay, I'm not doing, for wheat, let's say, let's say it's doing like a billion dollars a day, per four days a week, which is about right for the wheat market. So one day, this crazy thing happens, they trade, you know, a million dollars worth of wheat, and it says, Oh, I'm not a billion dollar a day marketplace, I'm a, I'm a million dollar day marketplace. It doesn't cost x like I thought yesterday, to change prices in me it costs 1,000th of x, because I'm a crappy, tiny marketplace that barely has any deal flow. So, people can now change prices for 1/1000 of what it used to cost. Of course, it's not the kinds of numbers I was just quoting, it might be one 1 millionth, or one 1 billionth of the of the price. And so now, the people who come in and say I want to move the price, you know, back from $10,000, to you know, $10, like, like it was the day before yesterday, they don't have to stroke a check for $100 billion, they have stroke check for $100. So it's sort of this big lottery basically of who actually gets to do this. And, and anyone can buy in for basically any amount that they want to. So somebody can flip in a penny to do some of it and they might get lucky enough that that Penny actually gets to do significant amounts of money. And then that Penny is going to have, you know, 1,000x or 1,000,000x payoff, basically, because the market’s getting back to sensible. So that's, that's how we moderate the CDM, monitor how much business it's doing, and then decides how valuable it is to be able to change itself depending on that business. So when you suppress the business flow, you also suppress the cost of changing.
Botond Seres 27:02
I absolutely love that thank you for going into so much detail and they'll appreciate it.
Dave Erickson 27:08
Well, that's a good explanation of it, I think it's pretty clear. The big challenge, I think, is going to get the global economy to start using it more. And obviously, people are afraid of new things. On one hand, they're afraid of new things and so they're very cautious about doing something like a CDM. On the other thing, we love shiny new toys like AI and are willing to basically let it control everything. So I wish the world as a whole would have a better balance when approaching new things. Obviously, AI will have a big role in CDM, you know, but it depends on where AI goes and how much of AI society is going to allow it to participate in. I think that brings us to this topic of AI alignment, as well as you know will AI take over the world. So maybe you can go a little bit into your thoughts about that.
Noah Healy 28:12
I see AI as one of the technologies that allows computers to, to exhibit superhuman abilities. And I see superhuman abilities as something that we we see as intrinsic to human society, human civilization, from bows and arrows and their ability to you know, make a make a stick with a rock on it fly further than a human being throwing it to the steam engine or the internal combustion engine, being able to produce more power with less, less cost than human muscles. We now have the ability to generate thoughts, control, judgment, writing and many many other things in ways that are either less expensive or more, more accurate or more valuable or in some cases both than, than human minds can generate. So in, in one sense, that's fantastic. If we can exploit these super judgers, super forecasters, super remembers in ways that can benefit us then, then we will greatly benefit just as we greatly benefit from the ability to hop on a train and cross a country or fly across an ocean. If, on the other hand, we can't find ways to make these capacities helpful to us, then we will, we will not really be able to keep the, the ball rolling on the way that we're currently doing things because the amount of noise that's being produced by computers today is overwhelming our financial and political systems and without laws or our economy is then we're gonna have a hard time with, you know, food and shelter.
Dave Erickson 30:20
Yeah, I think one of the aspects of AI or artificial intelligence is that, you know, AI is specifically computer programming and data analytics, that take a bunch of data and reconfigure it in a way that answers questions, that predicts outcomes and that has language models that allow for communication with humans, and to make their request better. But that's not consciousness and that's not creativity and that's not inventiveness. The question is, are those things needed for AI to become dangerous, like I hear all these doom and gloom prophecies about AI taking over the world and humans are eliminated. But on the other hand, if they're really just computer programs at our disposal, I think until an artificial consciousness is, is part of that, I don't know how great its danger really could be.
Noah Healy 31:31
The example I like to give is that the danger is not a Skynet or Frankenstein style danger. The danger is, if you built a car that didn't have brakes, or a steering wheel, or an exhaust system, and so you just put your foot on the gas, it vented carbon monoxide into your face, and at some point, you ran into something and then you stopped quite abruptly, the the issues that we're seeing is that, like, for example, with financial markets, we just a month ago saw an electronic bank run, that was enabled by the kinds of communications that computers enable and, and that meant that a bank could be targeted in drained faster than our existing regulatory structure thinks that banks can be targeted and drained. And that's, that happened in a very specific instance, to a bank that was incredibly vulnerable to that. But we can expect that thing to happen more and more often. And as for example, the Wall Street Bets thing demonstrated, it's perfectly possible for large scale social media networks to form that would allow that sort of behavior to happen to quite literally any bank on the planet, even extremely well capitalized, highly diversely capitalized banks, like JP Morgan, which is currently number one. You could imagine the existence of what amounts to a global public conspiracy by people who basically just want something to do on their Friday afternoon, deciding for, for giggles, to destroy that bank. And if there were 100 million ish people in that conspiracy, they would have the trivial power to do so. Now, I don't think that's going to happen tomorrow. But we are going to continue scaling up the availability of communication, structured communication, and so on and that's going to hyper-accelerate the rate of noise production. And so we're gonna see more and more agents doing things on behalf of people, whether they're conscious or not and the interactions of those agents are going to become less and less predictable, the more and more of them that there are, and that will cause the sorts of financial calamities that we see through flash crashes or these bank runs, that will cause political polarization and and other things. So right now, politicians respond to polls, they also respond in many cases to, to, you know, personal addresses, emails aren't really valued that much, personal letters are valued more, calls are valued even more. Well, deep fake exists. It is, it is well within the current, the current technological paradigm for anybody with a few weeks to waste and a couple $1,000 to generate 1000s 10s of 1000s of constituent calls to every single member of Congress, all all plausibly sounding somewhat like people having vaguely common interests around something. And well, that's not a, that wouldn't work that well, today, it would work pretty well. And, you know, call me back in a week and see how much better it will work then because that's what they're working on over open AI. And…
Dave Erickson 35:26
Two years from now, it should be really good, right? Just like, four years ago, it was really bad and now it's kind of okay and two years from now, it'll be really good, right? It's just the pace of technology.
Noah Healy 35:38
Exactly. So a system based on human to human communication, is going to find itself inundated as the computers manage to figure out what about human communication that we recognize as valuable, and then become capable of generating whatever that is in indefinite profusion. And so we need systems that can identify what's actually, what the actual value they can provide and create deliberate and structured ways to provide that value, at which point, we can take the fire hose and plug them in, and then the systems just work better. But without, without that, that part of the work, that, that sort of receiver end of things noise will eventually pollute everything.
Dave Erickson 36:35
Well, Botond, you're an expert in noise. Anything you want to ask about this?
Botond Seres 36:42
Sure. I mean, I would argue that noise is already drowning out any kind of useful communication. So I'll be honest with you, no, I don't see how it could get worse, but I am certain it will. One of the fun things I saw quite recently is you can now set up a small, AI bot on your phone to keep in touch with your friends, because maybe you're an introvert and don't realize that people feel that it's important that you send them a message every once in a while. And so I firmly believe that at some point, it will be just these bots writing to each other, having the best conversation of our lives with our friends without any of us actually participating. And I have felt like this has been happening in the financial markets for a good 5 to 10 years, since high frequency transactions have been attained for ages. But with the proliferation of AI that is happening right now, I feel like it's gonna have a much, much scarier role in poll determining price and interest.
Noah Healy 37:59
I would, I would agree with that. Yes. And that's, that's why, to some extent, the pitch for CDM gets better and better every day, because the existing markets are, are sort of being more and more spectacularly bad at doing their jobs constantly. And, and my contention is that it's being driven primarily by these technological changes, and not the incompetence and malfeasance of the actors within those marketplaces, although, as the stories demonstrate, that's not absent, it but it's, it's also, I think, incorrectly, being set forth as the primary problem. So it's not, the issue isn't the fact that we've globally network communication and human propensity for faddishness, suddenly, we can drain a bank before anybody can do anything about it. The problem is that these specific banks were just run by people who were overconfident or dumb or criminal, and all the other banks which might not be run by people like that, are obviously fine until one of them fails. And then it's Oh, right. We forgot about that one, which is also one of these bad but we're all still on the good side. But, you know, I what I'm looking at is this relentless tide of technological expansion in the communication information space, banging into cultural and social structures that were innovated between 700 and 3,000 years ago. And I would be shocked if we discovered that most of those, those institutions were up to the task of dealing with the internet. Something that would be difficult to even describe to the people that, that actually came up with these things in the first place.
Dave Erickson 40:08
So in, in financial models and in marketplaces, where our current system, there's a lot of opportunities for opinion, opinion sways prices, opinion affects things. Whereas in your description of a CDM opinion has less of an effect, there's less opinion points. And I think that that's where the communication and AI kind of come in, because you can use AI. I'll give you a very good example. I learned how to make 300 short YouTube videos using AI in about an hour. Now, I can make an opinion, short video that says, hey, I went to my bank, the XYZ bank, they have problems, they're short on money, you should go pull your money out, right? And make 300 of these videos and stick them on YouTube, whose algorithm is set up to promote the stuff that gets traction. And these rumors are fulfilling the fantasy of conspiracy theorists all over the place. And then all of a sudden, all the traffic is going to 300 because there's so many of them. Or you could just say every day, I'm going to put out 500 or 1000 of these videos. And just by sheer number, enough people will see this and say, Oh, and this is my opinion, I should go pick my money out of this bank, right? Or a financial institution where the price of wheat is set by, oh, there was a disaster in the Ukraine, and nobody heard about it. But then all of a sudden, there's 5000 videos out there showing a grain silo on fire or something and the wheat, the price of wheat is going to go up. So you should, you know, buy futures at this rate. Right? That kind of manipulation because much, right? Yeah, right.
Noah Healy 42:03
Exactly. Yeah. One of, one of the things that I've been thinking about is that information is actually almost perfectly physically equivalent to entropy. And so entropy is very related to energy. And, and in our common parlance, we divide energy into two separate concepts: work and heat. And the difference between these concepts isn't physical, it's social. Work is energy that's essentially being used for something that human beings want. Heat is all the rest of the energy in the system. I think we have names already, noise and signal to separate information in a similar kind of way. And right now, our institutions are not built around the concept that there is a separation between noise and signal. And so they treat all information essentially, equally and so as the systems that we develop, crank up the noise level, what we see is noise comes to dominate. And because noise is defined as things that human beings don't find useful, is definitionally useless. As most of the channel gets consumed by things that are useless, we see these breakdowns occur. And so this needs to happen at the philosophical level, first, the recognition of what signal even means in the context of some particular application and then there's a computational engineering standpoint, where you figure out how to measure and isolate and promote signal over noise. And then there's the cultural aspect, where people need to participate in a system that's rewarding them for providing signal and, and punishing them or just simply ignoring them when they provide noise. And that, that's sort of my top end paradigm for how we can navigate these problems like alignment and AI in the world that we're going to be occupying, you know, today and tomorrow and for the next centuries.
Botond Seres 44:35
So what do you think about old AI systems that are attempting to improve the signal to noise ratio, the ones that can recognize AI generated content, essentially, I mean, in my personal opinion, I think they are going to become essential tools, like maybe even browser extensions like we have ad block today, we are gonna have AI block in a few years from now. I wonder what's your thoughts are on this?
Noah Healy 45:04
I see those in sort of two branches. One, the construction of these things actually makes AI more effective and, and increases its capacities. So for example, chess right now human beings can't play chess effectively, that we have got these AI systems, they're like 1000 points better than the best of us are. Humans, more or less can't, can't compete with computers in chess. But we can compete using computers in chess. And the two top teams, there's an open source team called Stockfish that has a little bit of AI in it, but mostly is a dedicated team of high skilled programmers and chess masters, designing the very best chess engine they can. And the other one is this project called Lila, which is, chess enthusiasts donating computing time to a self teaching AI system. And Lila is a little bit worse than Stockfish. But one of the things that helped Stockfish along is a Leila side project called Spearfish, where instead of Lila teaching itself how to play chess, it teaches itself how to beat Stockfish. And so what that does, that program allows the Stockfish developers to see what their weaknesses are, and redevelop their project in ways that can defeat that version of Spearfish and then iterate up. And so I see I see these these sort of counter generative AI, you know, distinguishing technologies, as the beginnings of what might be part of a social cultural immune system, where, you know, the human body has sort of a White Label, Black Label approach, to knowing what it is and knowing what it isn't attacking what it isn't, and then developing sort of super attackers for specific known threats. We might see something like that, or we might see these technologies essentially subsumed into getting over that Turing Test bubble. And getting us into the point where we really can't distinguish between the two of them, because we solve that problem, and I, I would prefer us to work on the first one, but the projects that I've seen so far, largely being dedicated to that second effort,
Botond Seres 47:51
Right? I mean, I honestly was not aware that there were such efforts in place already. I do wonder a lot.
Noah Healy 48:01
Oh, yeah, yeah. One of my favorite papers, and there's several different ones of these, but they're known as the pixel attack in sort of the early days of AI when they were developing things that could sort of tell the difference between a dog and a cat, or, you know, is this a picture of a mountain or whatnot or something like that? There were developed counter AI,
Botond Seres 48:27
everyone's familiar with them, they all helped train those AIs.
Noah Healy 48:31
Yeah. So the, there was a, a development of AIs that would analyze those AIs and develop what was known as a pixel attack, where this AI could change a single pix, pixel in a picture being given to a classifier and cause that classifier to be incorrect. And so I could have saved the screenshot of Dave and show it to the classifiers, I'd say, Yeah, that's a human, that's a human man. And then the pixel attack goes in, changes a single pixel, something that you, I, we couldn't even see and it says, No, that's not a human being, that's not even a man, because that pixel is different and it turns out that that pixel has a, you know, strong enough effect across the space that it can it can destroy that particular AI's ability to work.
Botond Seres 49:21
I had no idea those systems can be so fragile. That changing a single pixel can have a butterfly effect.
Noah Healy 49:27
Yeah, yeah. Well, it's it's hard to produce unfragile or anti fragile decision structures and, and existing
Botond Seres 49:39
That is the ultimate goal isn’t it?
Noah Healy 49:43
Well, sure, yeah, I hope so. These questions become very difficult to deal with. There's a, there's also something known as the, the hungry donkey's dilemma. I think, I think that’s the thing I’ve usually seen them under. Basically, any discrete classification of a continuous by continuous function must have a undecidable input. So the classic example is you have a hungry donkey and you present him with two piles of hay. And so he goes to the bigger pile of hay. And then you make the same size, so you go to the closest pile of hay. But there's a position that you can put that donkey in with two piles of hay, where he won't be able to make the decision, and he'll just loop forever and starve to death. Now, this hasn't been biologically replicated, but it's, it's one of these sorts of stories to demonstrate a mathematical fact that no matter how clever you are, if you have a continuous choice function, making a decision among a discrete set of outcomes, you have to have one or more of these inputs that will cause the choice function to not actually produce anything.
Botond Seres 51:08
I'm sure you're always also familiar with the prisoner's dilemma. Absolutely. Just for our listeners, uh, very simple thought experiment. We take two people and both of those people get the same deal. They give up, they give up the other, get reduced sentence, or they don't, and they might get the medium length sentence. So I do wonder if, if you think that there is always going to be at least one AI willing to rat out the rest? It’s over simplifying it quite a bit.
Noah Healy 51:49
Well so, Right. So one thing to understand about the prisoner's dilemma is that the prisoner's dilemma changes behavior under iteration. So if you're playing the game, once that the game is structured in such a fashion that defection or betrayal is, is in the individual interests of the participants in the game. But what if there's a tomorrow if there's tomorrow, then the person you injured today might injure you back. And the structure of the prisoner's dilemma is such that being injured back is even worse than sort of taking your own medicine. So in an iterative prisoner's dilemma, where there's always a tomorrow or there might always be a tomorrow, the sensible behavior actually shifts to being less tratorist, less betrayal oriented. But that's not the only kind of strategic situation that people can find themselves in. Coordinated discovery markets is based on coordination games. So things like the stag hunt model, where people can either hunt for deer or rabbits. And hunting for deer means that you eat longer, but you need friends to help you hunt the deer. So what CDM says is, we'll publish what people are going to do and if there's a better option out there, people can pay to get other people to come on board with a better option. And if that option turns out to be better, then the rewards of going to there will be paid back to that person for paying to make everybody's lives better. And if it's worse, then that person will now have paid the cost to other people to try out their new idea and they can change back to what they used to be doing, which is better. There's also for opposition chicken and chicken might be an even better model for, for putting AI into, because the prisoner's dilemma actually creates collusive behavior, again, if there's going to be tomorrow. So you, you effectively create the incentives to create cartels around collusive behavior. Whereas chicken offers people who want to, want to sort of go for the brass ring, the double loss where conflict is resolved by destroying both parties. And so depending on how frequently other parties are actually willing to care about what it is that they care about, it becomes less and less sensible to dispute them unless you also care enough to sort of put it all on the line. And so if we have concerns about AIs generating collusive behavior that can counter our interests, the way to fight against that would be, we put them in strategic situations that make compliant behavior advantageous and collusive behavior, suicidal.
Dave Erickson 54:52
Noah, thank you so much for taking the time to let us know that there's still a chance that humanity might survive AI and that we may all also have a better option on how to make a secure marketplace like a CDM work in the future. For our listeners, please join us in the first week of the next month for another ScreamingBox technology and business rundown podcast. And until then stay on the right side of life and have a good day.
Dave Erickson Outro
Thank you very much for taking this journey with us. Join us for our next exciting exploration of technology and business in the first week of every month. Please help us by subscribing, liking and following us on whichever platform you're listening to or watching us on. We hope you enjoyed this podcast and please let us know any subjects or topics you'd like us to discuss in our next podcast by leaving a message for us in the comment sections or sending us a Twitter DM. Till next month. Please stay happy and healthy.