II Dr. Ignacio Deza: Full Podcast Transcript (part 1.)

Host: Maximiliano Fabres

Guest: Dr. Ignacio Deza

Location: Bristol, UK Youtube: https://www.youtube.com/watch?v=gCZnDDSviVg

Spotify: https://open.spotify.com/episode/4fRy4n010RBaoyM3JIAgLd?si=df3a9b92cd844320 

-

M: Welcome to Beyond the Horizon, the podcast for current and future leaders. For every episode, we explore a different field of knowledge, not only to collect facts, but to uncover the mental models behind every expert's understanding of the world. How do they make decisions? What counts as evidence? How do they reason through uncertainty? And what can leaders learn from the way they think? That extract you just heard comes from Frankenstein, which I found a fitting parallel to this week's podcast, artificial intelligence. Today, we're exploring the mental models behind AI and what it can teach us about leadership, decision making, and navigating through complexity. My guest today is Dr. Ignasio.

M: Thank you very much for coming.

Dr. D: Thank you for having me. First things uh that I want to do is in this first episode I really want to go through uh understanding what AI is and then how it emerged and the whole full backstory so we can actually understand it because at least uh I feel very lost um even though I did a ton of research um and also why should leaders care.

M: Okay.

Dr. D: So let's start with a very basic question. What is AI?

Dr. D: Well, AI is basically statistical computer system which was trained to predict a conversation. So, it's basically predicting the next word in any text. Some of those texts can be conversations. So, when I say something and you answer, if I say, "Hello, how are you?" What you going to say? I can predict you're going to say, "I'm I'm all right. I'm fine.

M: Okay. So, we humans do the same. We have an inherent prediction capability.

Dr. D: Uh sometimes it's easier to predict. Sometimes it's not. For example, if I say uh hello, my name is the the the prediction is going to be very hard because it has to find from a lot of different names. But sometimes I say um um I know uh some predictions are very easy because they are already part of the you're already for the ride. You already gave the information and the rest is just um fluff. So hello, how are you.

M: Today?

Dr. D: Okay.

M: Okay.

Dr. D: So I can show you later if you want some tools you can you have online which are prediction machines which are based on LMS and they show you how they uh work inside.

M: The but the the I I'll give you a couple of links.

Dr. D: Okay.

M: Uh we might put them in uh in the description below.

Dr. D: Okay. Perfect.

M: So that that's what AI is. It's simply a statistical computer system.

Dr. D: Okay.

M: Then the following question would be why what is it not?

Dr. D: Uh it's not a thinking machine. Uh it's not uh doesn't reason, doesn't um have values, doesn't think, doesn't make decisions. Uh it mimics many of those things very well and it passes what is called the touring test. So the capability of humans to know if the other person is a machine or a human by talking to them. So if you touring like 50 or 70 years ago, he said, "Okay, a machine should pass my test." Test means you have a chat. Uh well, it was a long time ago. So they was supposed to be somebody on the other side of the room.

M: Yeah.

Dr. D: The um and if they write something, the other person would answer. And you if you don't know if that person is a well that other is a person or a computer.

M: Mhm.

Dr. D: Then the computer passed the test.

M: So we're already there. We already passed the Turing test which was a benchmark which sat on challenge for almost 70 years.

Dr. D: Perfect.

M: I got I got a question though.

Dr. D: It's a mimic. It's not the real thinking we are used to as as humans.

M: I got a followup question on that and it's mainly towards a research.

Dr. D: That.

M: I think it was Gemini.

Dr. D: Mhm.

M: Where they basically tried to close AI an AI agent and it duplicated itself and try to hide inside somewhere.

Dr. D: Yeah.

M: Why do you reckon that happens?

Dr. D: Because the systems are trained to do that. When you train a system, you train them with reward systems. And the reward of the system is going to be to stay alive or the reward of the system is going to be to to do this type of task because that's the kind of way we humans operate. So if it's mimicking a human is always going to have some kind of self-preservation.

M: I see. I see. So if we would to input some other type of information then it would be different. That's what you would say.

Dr. D: So it's a statistical model. It's never going to give you the same solution to everything. You actually if you chat with a AI and you copy paste the same thing in different sessions, you're going to have different answers.

M: Yeah.

Dr. D: It's not uh.

M: That has happened to me a lot actually.

Dr. D: It's not something that that always is not what we expect from a normal computer.

M: Okay,

Dr. D: it's a statistical move.

M: Okay.

Dr. D: Uh the the the thing is they they are very complex.

M: And we humans have the call ability to anthropomorphosize things. So to give them human appearance. So you can have a stick with uh two legs and already you think it's a person and you can make it dance and It's very easy for us to just I can just get up a glass and I say hello I'm a glass and already we just think as as a glass.

M: Yeah.

Dr. D: Okay.

M: Well, we are storytellers as.

Dr. D: And if if I break the glass.

M: Cool.

Dr. D: People will suffer because I kill it.

M: Yeah, I see. I see.

Dr. D: So, we do the same with the LLMs which are able to answer our questions. It's very very easy for a human.

M: Yeah.

Dr. D: I want to go into LLM but we we will go there after.

M: Yeah.

Dr. D: So the the next one would be why is there so much confusion around them?

M: Well, two reasons. First of all, because it's a truly complex topic.

Dr. D: Mhm.

M: And secondly, because there's a lot of hype. There's a AI bubble. People need to to make it work. They need to sell as is a normal is a new technology. People invested billions of or trillions of of pounds in it. So they need to to to hype it.

Dr. D: Okay.

M: And selling the truth that it's just a statistical model which mimics human conversation.

Dr. D: Mhm.

M: Is much less sexy than saying artificial intelligence.

Dr. D: Yeah.

M: So would you say AI is a in artificial intelligence?

Dr. D: Depends what you mean with intelligence. is a cockroach intelligent.

M: I see. Okay.

Dr. D: You would compare it to a cockroach, right?

M: Yeah. It's about the same complexity.

Dr. D: Yeah. Well, we're talking about.

M: Yeah. The the number of connections, number of of a normal talking mole is about uh very a smart insect.

Dr. D: Okay.

M: Uh and yeah but what was what was the problem? The thing is LLMs are able to work with language and language is the operative system of our society.

Dr. D: Yeah.

M: So we live on stories, we live on text. So uh a machine which is able to manipulate text is able to manipulate society. So into like a Italian.

Dr. D: Italian.

M: Yeah. Italian blood.

Dr. D: Yeah.

M: So basically the that's the thing.

Dr. D: Okay.

M: So basically what they're doing is copying the language and.

Dr. D: Manipulating the language.

M: That's why they are so.

Dr. D: If you want can give you a really brief historical idea of how they came.

M: That's that's actually what what I was going to go after. I would love to see like what the history looks like.

Dr. D: Let me go very fast. Okay. I I'll start backwards.

M: Mhm.

Dr. D: So, we all remember November 2022.

M: Uh Chad GPT came out.

Dr. D: People went crazy.

M: Felt like Star Trek talking computers.

Dr. D: A little bit.

M: Uh it was amazing. everybody was. And then they very fast discovered they weren't as good as we thought they were.

Dr. D: Mhm.

M: And but still good, but they were like.

Dr. D: They're pretty decent.

M: They're pretty decent and they they're here to stay.

Dr. D: Okay.

M: Uh and they are built in three things happening together.

Dr. D: Okay.

M: So for artificial intelligence to arrive, we need three things. We needed all the technology. So data centers, computers, GPUs, uh things that were totally unrelatedly developed with for example GPUs were developed for gaming. Data centers were developed for storage of data.

Dr. D: Yeah.

M: There were nothing nothing to do with AI. 20 30 years of of building infrastructure. The internet was developed for talking to each other.

Dr. D: Yeah.

M: Then data avail which was also part of the internet and the way hard.

Dr. D: That was for marketing actually.

M: Yeah. That was probably well for defense before that.

Dr. D: Yeah.

M: And it was like um and the fact that data centers started to be much cheaper because before uh a hard drive was incredibly expensive and now uh then the the price per gigabyte went down. So you could start saving data.

Dr. D: Um so we had a huge amount of data available. We had the computer power to process that and the only part what was missing was the architecture mathematical architecture for creating these things.

Dr. D: To actually put them together.

M: And then you put three things together and you get what the LMS.

Dr. D: Sorry I don't get the the.

M: Infrastructure.

Dr. D: Infrastructure data.

M: And the the mathematics the the model.

Dr. D: The model okay yeah.

M: Okay so the model, believe it or not, was the first thing.

Dr. D: Mhm.

M: Was started in about 1950s.

Dr. D: Okay.

M: With a model called the uh perceptron. It's a very 1950s name.

Dr. D: Perceptant.

M: Sounds like a transformer.

Dr. D: Yeah. Yeah.

M: The percept. Yeah. Like like Decepticon.

Dr. D: Decepticon. Yeah. The perceptron was basically a logistic regression which is an algorithm uh been there for at least 200 years.

M: Which basically gives you uh the probability of uh a number.

Dr. D: Okay,

M: it's a very simple formula been there forever. It's called logistic regression because it's been used in logistics a lot to calculate the probability for example of um of something arriving to your warehouse and all that. Mhm.

Dr. D: Uh just a formula.

M: Okay.

Dr. D: And they what they did was stack several of them and create something like a network. And they said that was mimicking how neurons worked.

M: Okay.

Dr. D: So it was all to mimic how neurons work.

M: Yeah. The idea was they had this um idea of how neurons worked. They were able they knew neurons were electric. So they actually work with electric impulses and they were able to open an animal and see how neurons interact like a very basic animal like a.

Dr. D: Yeah.

M: Like a cockroach.

Dr. D: Yeah. Well, cockroaches are very very sophisticated. It's like a maybe much much lower like a squid or something.

M: Oh, okay. Like amoeba or something.

Dr. D: Yeah. No, amas have no neurons.

M: Oh, my bad. Sorry, guys.

Dr. D: But a squid or something. And then they were able to to to measure impulses and they discovered that they basically behaved like that. So they did that in 1950 and say okay we cracked human intelligence. Everything is is a is a new one.

M: Good luck with that one.

Dr. D: And they did that. It didn't work. It didn't work. Um was a fiasco. they and AI basically was still born and didn't work. Decades passed.

M: Until about 199 like 40 years later.

Dr. D: Nothing happened and then started to develop algorithms all this theoretical about how AI would work.

M: What does what is an algorithm actually? Algorithms is basically a series of instructions to how to do something.

Dr. D: Perfect.

M: They have to be very specific. So a computer can do them.

Dr. D: Okay.

M: So you could do them blindfolded. It's like if I have to go tell you go to the to the fridge and bring some water, I have to give you every turn, every move, every how you how many degrees you have to move your hand.

Dr. D: Okay. So every time when we in the talk about AI, every time we hear algorithm is the same as a series of instructions of clear in instructions.

M: Yeah. But then nowadays in marketing they they call an algorithm an AI algorithm which is not an algorithm.

Dr. D: Okay.

M: When they say.

Dr. D: What is that then?

M: When they say the Facebook or um YouTube algorithm.

Dr. D: They don't mean that. They mean the a YouTube's AI system which is basically a reward system to reward you for doing some things they do.

M: They if you do the things like they want.

Dr. D: They they reward you with more viewerships or more recommendations. So.

M: I would I would argue a little bit being being in the in the space myself.

Dr. D: Or yeah I mean works both ways.

M: Yeah. I would say whatever the audience wants is going to get rewarded. Therefore,.

Dr. D: Yeah.

M: It's basically the same to say al YouTube algorithm than audience.

Dr. D: More or less or less because of they can make a lot of um changes themselves. For example, the their rules.

M: And you need to comply and you don't even know what the rules are. It's it's not an algorithm.

Dr. D: Yeah.

M: Okay. So basically that that's a problem. After that you have well so so the perception was there nobody liked it didn't work. Uh we got to 90s it developed all this kind of a mathematics background about how um AI works. Okay. And it was still it was useful. We started to make it work. It worked for some tasks. You could use for example character recognition. you could recognize num written numbers or words.

Dr. D: Those thing that technology started to come in the in the 1990s you were able to OCR things documents and transform them into into texts.

M: Mhm.

Dr. D: That existed and that was AI. It was very basic and it was basically there but it was usually cheaper and better just to hire somebody to type it because the If the data was complex, the system couldn't cope.

M: Perfect.

Dr. D: Then the 21st century comes 2000s.

M: And the technology starts to advance very fast because we start to have synergies with the data. The data was going to was starting to go up. data was more available was uh cheaper and there people were doing OCR of books.

Dr. D: There was more data.

M: Many many books may maybe articles many things that were available in paper now we're starting to.

Dr. D: Google academics started to.

M: Yeah papers were started to be online things move online so the internet and people moving things online.

Dr. D: When YouTube was still an app.

M: Yeah you you could basically transform all of that into uh into training data.

Dr. D: Yeah.

M: So you use people use that and then people start to generate data all the time because interacting on on the internet makes you are generating data.

Dr. D: You are chatting, you're commenting, you're speaking, you're clicking, you're doing things. So.

M: You're doing stuff. Yeah.

Dr. D: Yeah. So that data starts to to be generated.

M: Yeah. You start to interact with the platform and therefore you get more. So the amount of data you had in the 1995 and the amount of data you had in 2005 was totally different.

Dr. D: It's a totally different world.

M: And then.

Dr. D: And then uh that's what happened. The the amount of.

M: Information.

Dr. D: Information was fed into the systems systems started to be able to handle it.

M: Okay.

Dr. D: And they became more useful. did in that.

M: I was there so you just asked me I remember.

Dr. D: Okay.

M: So oh my god yeah thank you so.

Dr. D: Um then that means that was the what was in that moment then the the ceiling of of the three things that you mentioned.

M: Processing.

Dr. D: The processing power so.

M: No I I haven't got there I had we had the data.

Dr. D: Okay.

M: We had the algor algorithms were basically the ones we had in 90s.

Dr. D: The same ones.

M: No, I'm going to go back in months.

Dr. D: But kind of kind of the same.

M: And but they were very similar.

Dr. D: Mhm.

M: And then about 2005 they invented something called um recurrent neuronet networks.

Dr. D: Which we were able to take in in account the passage of time.

M: Before that they weren't able to think about time. there was just you have the data and you make classification. You have the data and you make a prediction.

Dr. D: It wasn't about uh data first then then this happens then this happens then this happens.

M: Okay.

Dr. D: That's more complex. That's more complex.

M: I wouldn't know how to go about it. But they implement it time.

Dr. D: Yeah.

M: That's that's the important thing. But they have a big problem which is that these systems don't remember um things that happened before. So if I say the cat.

Dr. D: Mhm.

M: Who was lisurely sitting on the couch on this beautiful summer Sunday.

Dr. D: Mhm.

M: Was white. If you say that a human will know the cat is white.

Dr. D: Yeah.

M: The the those systems after all my introduction of about the leurly sitting and whatever.

Dr. D: Will probably think the couch is white.

M: Or the Sunday was.

Dr. D: Or the Sunday was white.

M: You see what I mean?

Dr. D: Okay.

M: So they lost.

Dr. D: Track of the.

M: Track of so the context.

Dr. D: Yeah.

M: And it happens to everybody only. We have a bigger context window. So if I say tell you something and then I ask you tomorrow, you won't know what I'm talking about.

Dr. D: Yeah.

M: If I answer you just what's up tomorrow saying yes, thank you. You will know what part of our conversation I'm talking about.

Dr. D: Oh, I actually do know. I actually do.

M: Yeah.

Dr. D: But that that will be the the idea. Okay. Yeah.

M: Um so, uh the this algorithmic part was sold in the in n in 2017.

Dr. D: 17. Okay.

M: Yeah. With a paper by Google.

Dr. D: Called uh which invented something called transformers.

M: I do. Yeah, I do remember investigating.

Dr. D: Transformers are the basic of how LMS work now. And what they do is instead of going step by step in the future.

M: Like other recurrent networks do, they process time all at once.

Dr. D: So.

M: Everything is happening.

Dr. D: At the same time.

M: Oh my goodness.

Dr. D: So if I give you a text like a prompt in a lm.

M: Mhm.

Dr. D: Uh the system is going to read the whole thing at once and then it's going to try to see which are the most relevant links between the texts. This is so crazy.

M: And that's called attention attention mechanism because it's basically paying attention to the links.

Dr. D: And that solved the biggest the big problem of losing context with time.

M: Perfect.

Dr. D: And that's why uh the tech the this um algorithmic problem was the first to come.

M: And the last to be solved. I mean it's quite a complex way of solving it I would say.

Dr. D: And it's.

M: You just need a huge processing power.

Dr. D: Yes. And also it's not the way humans work.

M: Not at all.

Dr. D: Because when we read something.

M: Is actually counter counterintuitive.

Dr. D: The way we we we don't read a text all al together. We read a word by word and we are able to keep the context.

M: And we know that the way we're doing it with computers is not the the way we do it. Nobody said there's no there's just one way but we know we don't work the same way.

Dr. D: Okay.

M: Okay.

Dr. D: Yeah. Yeah. which is is very profound if you think about because that means the LMS are are not us not even in the in the at the core.

M: At the core.

Dr. D: I would say they do build the connection all at once which is what we do.

M: But they're not follow they're not chaining they're not creating this chains of.

Dr. D: They just they just work.

M: Story lines.

Dr. D: They just work in a total different Yeah.

M: And more akin to saying they are mimicking us.

Dr. D: More than that they are thinking.

M: So at the core AI is completely different.

Dr. D: Yes. And when you and just back into 2017.

M: They the system they discovered wasn't a chatbot. Remember chatbots came on 2022.

Dr. D: Yeah. 23 almost.

M: Five years after.

Dr. D: Yeah what happened in those five years in those five years what they got in.

M: Centers.

Dr. D: Yeah what no the centers were already there.

M: Oh okay.

Dr. D: Uh now they're better but I remember very well that in 2017 there were so many.

M: Okay.

Dr. D: What the what happened there was it wasn't a chatbot It was a text complexion machine.

M: Sorry,

Dr. D: text prediction machine is is what it is at the core. So basically predicts the next word or as they call it token.

M: The next word or token.

Dr. D: Token is what they call basically a word.

M: Why do they call it token?

Dr. D: Because that's how linguist call it. Oh, token is the the word for a single word.

M: Or a part of a word and they just took that word from linguistics and now they are charging per token and token became a basic basically a currency.

Dr. D: A currency that is actually very useful to know. So token is a word basically.

M: So basically they now they're charging per token when you when you use an LLM.

Dr. D: That's crazy.

M: But there's basically a word or a part of a word. All right. LLM is launch language model.

Dr. D: Yeah.

M: Which is basically transformer.

Dr. D: Which has been fed with a lot of language.

M: Transformer being.

Dr. D: This uh 2017 technology which was able to think to see everything at once instead of.

M: Moment by moment.

Dr. D: So if we translate everything would be a large language model.

M: Yeah.

Dr. D: Is a transformer aka a model. Recurrent neuron network able to understand things all at once and use the attention mechanism instead of the time mechanism.

M: Mhm. Basically very difficult. Yeah. And you get charged by by token which is basically a word.

Dr. D: A word. So if you prompt it with a million words, it's going to be more expensive to run.

M: Yeah.

Dr. D: Than if you prompted with hello, how are you?

M: Perfect. Right now as it stands they giving us all the this for free.

Dr. D: And most of the companies are losing money.

M: Yeah.

Dr. D: But they uh they are trying to to impose the technology and and so on.

M: Yeah. Now I saw that Changd is doing ad revenue.

Dr. D: It's going to I mean the the free ride is going to end sooner or later.

M: Yeah. Um.

Dr. D: Sadly,

M: unless they they get some way to capitalize it um.

Dr. D: By selling data maybe.

M: That's probably what they're doing.

Dr. D: They're already doing.

M: Yeah, of course. Everything you tell them it's for sale.

Dr. D: I mean, it's the same as everything you.

M: Yeah. Well, it's the same. But I think a lot of young people in particular talk a lot about their emotional state.

Dr. D: Well, now it's now it's public information. So well yeah so okay we basically got to 2020.

M: 2017.

Dr. D: 2017.

M: And then the three then they had to now we have to go from 2017 to 2022.

Dr. D: And what's the.

M: What happened there is they you went from a prediction machine.

Dr. D: So I say I want eggs milk and it's going to say butter or bread.

M: You see.

Dr. D: I'm a terrible predictor I wouldn't be I wouldn't be a good AI.

M: Yeah, you you are you are.

Dr. D: Uh I am okay.

M: Yeah, I mean humans are excellent at prediction.

Dr. D: Okay.

M: And if if you just get one of those systems which are available online, you can download one and use it.

Dr. D: For prediction of text. Uh they for example the one you have in your phone when you're writing and predict the next word.

M: Yeah.

Dr. D: If you if you have one of those, they are uh they are not able to to chat with you.

M: Perfect.

Dr. D: So those five years were that how to get from A to B.

M: And how did you how did they did two things? First they did a reward system. Reward system basically forces them to give the answer somebody expects. Mhm.

Dr. D: Remember this trend with all the information in humanity. So everything we we know not all the not all of the information they well famously they didn't ask for permission. So many of of the information they use was copyrighted and things.

M: So we can expect this information to stop being.

Dr. D: In in future training data.

M: Because they're going to take it out probably because of copyright.

Dr. D: That's going to be a good lawsuit. Yeah, there are there are many of them.

M: I know.

Dr. D: So, but we can expect LMS in the future to be less uh capable. Then after that you have the the second thing is that the these LLMs are going to they were just predicting machines. So you have to give a reward system to make them interact like as human as people.

M: And they were trained with a lot of conversational data. So they probably just went into all your WhatsApp conversations and were able to to mimic how people talk to each other.

Dr. D: So how did they solve it again? So what was the key thing?

M: They train it with a lot of conversations.

Dr. D: So just data.

M: Yeah. And then they were able so it can complete conversations.

Dr. D: Mhm.

M: Okay. So it's not not the same thing completing a text or essay than completing a a conversation. No.

Dr. D: So it needs to have examples of conversations.

M: Well,

Dr. D: so what I'm guessing this is what is happening the looking at all the text that you wrote all at once plus making those connections of those words.

M: How what's more important.

Dr. D: With what's more important and so on. finding the context.

M: And then.

Dr. D: Goes to the memory.

M: And then search it all at once.

Dr. D: With those.

M: Goes to the memory finds the most relevant information about the the text you wrote.

Dr. D: And then answer on a way a human would.

M: Okay.

Dr. D: So answering as if was a conversation instead of answering just completing because if I say if I am if I if I say for For example, I want to go to Paris.

M: Mhm.

Dr. D: Uh modern chat GBT will say, "Oh, okay. Give some information about Paris."

M: Yeah.

Dr. D: If I I get the basic 2017.

M: And I say, "I want to go to Paris." It's going to answer, "I want to go to London. I want to go to Berlin. I want to go to."

Dr. D: Because.

M: Related things,

Dr. D: it's going to just complete.

M: Not not a not a.

Dr. D: And then it has a a training reward system.

M: Training reward." And it was it's a totally different training called um a reward system.

Dr. D: Where where this is the system is trained and if it's.

M: Doesn't give the the good the correct answer it's punished.

Dr. D: On some tokens and if it's if it gives the correct answer it's rewarded.

M: Okay.

Dr. D: So that's part of a training and that's what we were talking before. I'm very I'm very interested in how do they actually uh reward and so on. We're moving forward with the conversation but this time we are going to understand how an AI researcher thinks how they operate how they actually see information and digest it I'll call it and digest their information. Basically what we're going to talk about is about the mental models. So let's jump straight into it.

M: Predictions. AI research is a lot about predictions as we already talked about in the first podcast.

Dr. D: So at its core, what do you reckon is the problem um that AI is trying to solve? AI is trying to mimic the human brain. So that that's the the goal. Okay. It started in the 50s and is still we're still going there. We every time every step we're going near that and the the goal is eventually in the future gen create a thinking machine. If you guys want to understand a little bit more about the history, just watch the last episode.

M: So that's that's the thing. So uh we can argue a lot. I'm not going to do that about the implications of creating a thinking machine.

Dr. D: But right now we are basically in the early steps and it is giving us a lot of value now. So we don't do it for the future um AI like a taking over revolution. We're doing it for the actual value creating the systems is systems are incredibly simple compared to our brain.

M: They are not by any means similar to to human brain.

Dr. D: All right. So the mental model is uh we live in a world which is extremely complex. We don't have all the information.

M: We don't have all the variables. We don't know what's going to happen. And even worse, especially in business, our decisions can affect the outcome.

Dr. D: Yes. I think I think that's the toughest bit of business.

M: So the problem is it's not like so if you're predicting the weather for example is a very complex system. It's very hard to predict. You can see perfectly that even advanced systems get it wrong. Mhm.

Dr. D: But at least the weather is not going to change because of your prediction.

M: Exactly.

Dr. D: But in business or in social social life or many other things, your prediction is affected. So for example, if I say a candidate is going to win an election, maybe saying that some expert predicted that makes some people change their mind.

M: Yeah. And actually makes it more likely to.

Dr. D: Or more unlikely.

M: More unlikely. Exactly. don't know.

Dr. D: Because they don't like you for example.

M: Or yeah.

Dr. D: So that that's a problem and that's called a complex system.

M: Complex system it doesn't mean like a complicated system complicated means something which is okay many variables many things complex means something that you cannot split in parts and keep the same um behavior.

Dr. D: I think the best uh the best system to understand that is a stock market as soon as you buy a stock it goes up.

M: Yeah.

Dr. D: As soon as you sell it, it goes down.

M: Exactly.

Dr. D: So your action whether it's good or bad.

M: Or positive or negative.

Dr. D: Will have a response.

M: It will have a response.

Dr. D: So if you want to predict that even your predictions if they are public.

M: Will have a response.

Dr. D: Everything you do will have a response. So it's such a complex system that we're trying to predict that.

M: Requires another complex system.

Dr. D: You cannot predict a complex system using simple rules. It's impossible. It's like predicting uh but what you can do is to get to know the system well enough.

M: Yeah.

Dr. D: And then say what is ca capable to do.

M: Yeah. Okay. And that's the way many econometrics and many things actually do. They know what the system is able to do.

Dr. D: So it's not able to do anything. It has some rules.

M: Yeah.

Dr. D: Some inertia and things.

M: I do remember the the a class where my economics professor talked about the market and why people haven't cracked a code for this. And it's because what you're saying that you do need a complex system to understand a complex system and that's at the very beginning. Once you have those two, you still are going to have probabilities.

Dr. D: Complexity is going to go up.

M: Yeah.

Dr. D: So if you study the stock market from the 50s,

M: 1950s,

Dr. D: you're going to see it was much simpler than the behavior of the of the stock market.

M: Well, there were many less actors. Also actors were slower because you needed to be there in the floor of the place to buy and sell or the ticker.

Dr. D: You it was slower and it was less complex. The moment you add more technology and you add algorithmics and you add AI and you add things the system is always going to be complex more and more complex and usually out of reach barely out of reach.

M: Okay? Because if it's not out of reach, somebody will play it.

Dr. D: And win it basically.

M: And win it and break it because.

Dr. D: That's exactly what they uh found because there's there is one uh thing that did uh that did crack I I don't remember the company but it's one of the richest men in in the world and they had this uh thing where they had to place a bed at a certain level. They couldn't pass that level because if not they broke their algorithm, their own algorithm and it wouldn't.

M: It's not the algorithm. It's it's trust. So you invest on the stock market because you believe is not rigged or if it's rigged is not too much ripped.

Dr. D: Okay. If you know it's completely manipulated and you're just have no chance of winning and they just throwing away your money,

M: Then you wouldn't.

Dr. D: Then you wouldn't use it.

M: Yeah. So that that happens for example a casino if they they're very like a mafia kind of run casino which are obviously people just stop working people just stop going there.

Dr. D: Yeah.

M: Then um the next thing that I actually want to talk about is because you manage large amounts of data and it's one of the three uh things that we talked about.

Dr. D: That create this AI framework. How do AI researchers then view data? How do you think about it?

M: Oh, data is the most similar to natural resource you you have right now. So.

Dr. D: You see it as a natural resource.

M: Yeah. It's like a oil or like a.

Dr. D: That's crazy. Okay.

M: I mean the all computer systems work on data.

Dr. D: I've never thought about it that way. It's great. But yeah, yeah. Yeah. Now, now I understand your reasoning. Yeah.

M: So, that's why people steal it. That's why it's so valuable. That's why people give you things for free uh for your consent or your data.

Dr. D: Yeah. It's literally like a barrel of oil.

M: Yeah. And uh without data, without private data, sadly. And there's there's a white hat and a black hat. You know what white hat and black hat is?

Dr. D: Good. hackers. Yeah. Good. Good. No, like.

M: White hack hackers are the ones who actually work for the companies and help them.

Dr. D: Oh, yeah.

M: And black hats are actually the criminals.

Dr. D: Okay.

M: So, uh you use it as a so good guys and the bad guys.

Dr. D: Okay.

M: So, the there's a lot of black hack bad um data companies who basically steal personal data and use it. And then may even normal companies who need for example to to make chat GPT or those companies to work and to talk like humans.

Dr. D: Had to use human conversations.

M: So they just steal it.

Dr. D: They had to get them somewhere. I don't know how they did it. Twitter.

M: Nobody knows how.

Dr. D: Twitter is probably free but people don't talk on Twitter. People usually use things like chat like a WhatsApp or that kind of of um instance.

M: I don't know how much data they they used. I have no idea. I have no claim. So I'm not Don't sue me please. But their data has to come from somewhere.

Dr. D: And.

M: And that that somewhere.

Dr. D: Companies.

M: Then somewhere maybe old RAC channels back in the day where people use RA for chatting or maybe is uh paying people to to to give them their data. I know. So, um most I know at least Chad GBT was developed with help of Nigerian.

Dr. D: Yep from some Nigerians or something.

M: What they did is they needed humans.

Dr. D: To be able to tell if the answers were humanlike or not. Mhm. So the Nigeria, India and a couple of other countries have two conditions. They are people you can pay very little and who speak English.

M: All right.

Dr. D: Yeah.

M: So that that's the.

Dr. D: That's the equation.

M: Yeah. Because uh if you want uh of course now it works for every language,

Dr. D: but I'm talking about the beginning.

M: Yeah. Perfect. Um, so then what separates good and bad data?

Dr. D: You mean good and bad data in the in the sense of good data like or or morally?

M: No, as in good data. Not not not morally. We're not we're not making the moral claim right now.

Dr. D: So data data is data. Okay. Uh if data is contradictory data is has been for example imagine I have a sensor and I'm checking temperature in the room.

M: But the sensors of this the sun.

Dr. D: Mhm.

M: So it gets overheated. It's not representing temperature of the room.

Dr. D: It's going to say here we are and you know 10 degrees more than they actually have. If I'm recording that and one year later I say oh that was a very hot day.

M: That's bad days. But it wasn't. Okay. So something that is not real so to say or doesn't represent reality.

Dr. D: Can be an error can be missing data.

M: Can be misrepresentation can be data has been manipulated and broken like for example you're manipulating some text and then you break.

Dr. D: Twitter basically. No, you break your, for example, some languages have accents and then those accents uh get transformed into some strange characters.

M: Yeah.

Dr. D: Um.

M: Okay.

Dr. D: Or you know, many many things.

M: Something like that. Yeah. Something in that sort.

Dr. D: Many things can happen.

M: It's like taking Chilean Spanish or Argentinian Spanish text to create a model in Spanish. Spanish.