II Dr. Ignacio Deza: Full Podcast Transcript (part 2.)

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

-

...

Dr. D: Yeah. Eventually, eventually we'll need to do it. But yeah,

M: uh that wouldn't be bad data. That will.

Dr. D: Eventually.

M: There will be bad data for.

Dr. D: The application.

M: Yeah. For that application.

Dr. D: The problem is bad. Sometimes data is broken.

M: And part of what we do as data scientists is to clean data.

Dr. D: Then what do you how do you deal with uncertainty? So I know I know you clean data but how do you make decisions with that uncertainty?

M: Because we're human. Humans make decisions. you have to say, okay, I have this thing and I have to make a call. And the call is um is something we do and that's part of what makes us valuable as a in a company. If you ask AI to make decisions either you give them a very set of decisions like a you can do it algorithmically like a for loop or whatever.

Dr. D: Yeah. But if you have some problem and it's a complex problem of very complex data wrong and you have to make a decision AI will struggle.

M: Then do you think we so you would say because we have the decision power or the ability of making that decision.

Dr. D: We should use it.

M: Whether it's wrong or.

Dr. D: Well I mean we do it all the time we we make decisions.

M: Yeah.

Dr. D: Computers struggle with that because they don't have accountability.

M: They can't they unless you put them in this reward.

Dr. D: Loop.

M: Where they do something for example and oh I have to do this and they have to do that and then they can more or less act but they're going to be very limited in their capacity of action.

Dr. D: But I don't I'm not talking about.

M: They just do that but they are very limited. Oh, I see.

Dr. D: But I'm not talking about AI. How do you recommend taking decisions when there is limited data or there is bad data?

M: That's experience.

Dr. D: You would say experience.

M: It's like it's the same as everything. How a business owner makes decisions when they don't know the full picture.

Dr. D: They they make a call and then just live with it. Well, there is a a good quote which is.

M: There is no good decisions. You just take a decision and make it the good one.

Dr. D: You just make your decision and you're sure and then you just stand by your decision and if it's wrong, you change your mind uh after the fact.

M: But then but you have to to stand by what you you did and and own it.

Dr. D: Yeah.

M: Um Yeah. But that's how how it works. And that's something again AI cannot do.

Dr. D: Okay. Then I know you were talking about models at some point.

M: What do you reckon makes a model useful? What makes it not useful a model?

Dr. D: You mean um AI model?

M: Yes.

Dr. D: Or mental model?

M: Well, both. Both are fairly similar.

Dr. D: No. No. They're totally different. No.

M: Okay. So mental model is basically your set of rules and set of values.

Dr. D: What define you as a person. So I believe in this.

M: Yeah.

Dr. D: And when I make a decision, I'm going to stand by my values and by my my what I believe is right and then I'm going to make decisions on that. Even if you for example I know for example I'm.

M: I wouldn't call a mental model that I would it's what guides your decisions. I would say it's a set of rules to make decisions.

Dr. D: Yeah, but those decisions are informed by your values. That's what I mean. So,.

M: Not necessarily. So,

Dr. D: let me give you an example. You see, imagine you have a business and you obviously want to make money.

M: Uhhuh.

Dr. D: But there's a line you wouldn't break. You wouldn't go to child pornography, even if it's really really profitable.

M: Okay.

Dr. D: Because you have your values.

M: Oh my goodness. I think we we'll have to bleep that out.

Dr. D: Okay.

M: Yeah.

Dr. D: Yeah.

M: You see what I mean? Yeah.

Dr. D: Uh that or you wouldn't murder for money.

M: Yes.

Dr. D: Because that and and that's important because uh if you don't teach that to a computer explicitly.

M: Then you won't.

Dr. D: You wouldn't know better.

M: You won't know.

Dr. D: Yeah.

M: So you.

Dr. D: But I would say that's a set of values that that is not a mental model. I would I would argue that the mental model is just a set of rules to make a decision.

M: Yeah. But how you make your rules.

Dr. D: But how you make your rules.

M: Because of your values.

Dr. D: Yeah. Because of the values. But the mental model is the continue like the next step.

M: It's the same.

Dr. D: But yeah, of course. Yeah. Yeah. Yeah. Yeah. So then what would it make? What would a good mental model or what would a mental a good model sorry.

M: What would a good model how do you say a model is good?

Dr. D: Um when it works. just when it works.

M: Yeah.

Dr. D: You wouldn't argue that the process is important.

M: Depends depends on on the example. So for example, if you're talking about data, you're going talking about analyzing data or or that things, it will works. If you're talking about having your own business, you may use um prioritize some things. you maybe want to grow or to make more money or to charge more or to whatever and your decisions are going to be informed by what you want and then that if it works then you're doing well keep doing that if it doesn't work I mean yeah I see so I see I see I see I see I see I see.

Dr. D: Um I would say I would argue a little bit different but I I understand your point though that what works is is the is what's actually used.

M: What would you say is your your idea?

Dr. D: I would say that the right process is the most important thing specifically in business.

M: But how you know if it's the right process?

Dr. D: You know it because it provides the most amount of value.

M: Because it works.

Dr. D: Not necessarily works now because in business you got this delay factor.

M: Yeah. Yeah. I'm absolutely talking about that too. So.

Dr. D: It's not it doesn't have to be working now.

M: Yeah.

Dr. D: But uh but it works at the end.

M: Yeah.

Dr. D: But you need to know or at least believe.

M: It's going to work.

Dr. D: Yeah. Now in business you got a weird concept which is providing too much value could also be your death.

M: Yeah. There was a bicycle thing.

Dr. D: Which lasted for a bicycle business that lasted for about 200 years per bicycle.

M: Yeah.

Dr. D: Without needing any repairs or anything.

M: What happened to them?

Dr. D: They couldn't sell.

M: Yeah.

Dr. D: So making the product too good also was a problem in that. But.

M: Depends on your goals. If your goal is to make the perfect bicycle, then you succeeded. If your goal was making money, then you didn't succeed.

Dr. D: Yeah. Then what when would you reckon a model should be trusted? Only when it the result is the valuable thing.

M: That's a very good story about a hotel business that they use this AI model to predict a number of bookings.

Dr. D: And the managers pay a lot money for the model. So they told everybody trust the trust the model. mall and then the hotel had to to basically go bankrupt.

M: Really?

Dr. D: Yeah. It was like 10 years ago before like a modern AI.

M: Yeah. the the ones that we have right now.

Dr. D: It was uh and it means you need to to make sure things work and you need to be able to pivot because the our world is so incredibly complex that you cannot just live for one set of rules.

M: Yeah. That's why I I talk more about values than about rules because.

Dr. D: I well the podcast is about the mental models which is the next step as you said the values come first and then.

M: Because uh if you don't if if you just try to set rules on basically automate.

Dr. D: Um the complexity of the world is too much it's too great it can be fantastic it can work even for a couple of years but then suddenly without knowing why um it just stopped working because some of the premises you build your system up stops working.

M: And you will know.

Dr. D: Yeah, I think um the best example towards that would be the fact that a lot of people date the same way that they were dating when they were 16.

M: Yeah.

Dr. D: Uh when they're 30.

M: Mhm. And well, it doesn't work.

Dr. D: Yeah. No,

M: that's that's the that's the real thing.

Dr. D: That's a perfect example. People uh learn how to do something at 16.

M: Mhm.

Dr. D: And then they just go dating mode, so they get divorced.

M: And they go back and try to do the same things that they did when they were 16.

Dr. D: They were cute at 16. They're they're not cute at 30. They're probably yeah.

M: They're probably not that not that cute at at 60 or whatever.

Dr. D: But yeah, so this is why changing the model or the way you approach or your set of rules.

M: To your approach.

Dr. D: Really is really really important. Um so going a little bit away from models then.

M: How do you reckon we so how do machines learn like I I understand that is by data.

Dr. D: And I understand you said it was by tokenization.

M: And they read all at once and then they find the language or the idea of the the text.

Dr. D: Statistical and they go to the memory.

M: And then they go to the memory and so on.

Dr. D: Mhm.

M: I still don't understand how do they get rewarded though.

Dr. D: That's the second step. So the first step is the what what you describe is the only the the word prediction machine.

M: Yes. Yeah. After that they go to the second step which is the um you get an LLM which basically teaches the LLM and.

Dr. D: So it's a two LLM process.

M: Talking to each other.

Dr. D: Yeah.

M: And they get rewarded. So for example they have a conversation.

Dr. D: Hello how are you? I'm good thank you and so on. and you just have um you just add them a a cost a reward and a cost to the model and the the system has to minimize the cost and because you're is programmed to that.

M: Okay. So it's programmed to.

Dr. D: To to.

M: Reduce the amount of negatives. Yeah. And.

Dr. D: And then it has to amend the amount of positives. You're.

M: Making it do whatever you want to do by this reward system. Okay.

Dr. D: So it's basically constraining it.

M: I see. I see.

Dr. D: Let let me give you an example. Okay.

M: The thing is it's very similar to how human work is very similar to how.

Dr. D: Very very similar dopamine and stuff like that.

M: It's more similar to the way you train animals than the way you train people.

Dr. D: I mean you should talk to animal trainers in my opinion.

M: Yeah. animal trainers uh uh either I mean old old school the ones who hate the animals.

Dr. D: Animals don't want to be don't want pain.

M: But even modern animal trainers who work through love and empathy and just making the the dog trust them you know they're also doing a reward system.

Dr. D: Well there is a very good book called pimp.

M: And it's from a guy that well was a pimp for all his survive got arrested at the end.

Dr. D: But he used the extremes.

M: Yeah.

Dr. D: He used to like beat the.

M: Prostitutes up.

Dr. D: And then automatically reward them and like entice them and so on.

M: Oh my god. That's that's so horrible. But that's how you train or that's how they used to train.

Dr. D: Well, for example, let me give you an example very very small. Imagine you have a warehouse.

M: Mhm.

Dr. D: And you want to build a robot.

M: Yeah. who goes to your warehouse and bricks a product and brings it back.

Dr. D: Okay.

M: If you just train how to go how to move forward backwards and you just.

Dr. D: Train like that.

M: Yeah.

Dr. D: And then you go go and fetch this.

M: Mhm.

Dr. D: The robot has no incentive to go and come back quickly.

M: That's a good point.

Dr. D: So what you do is every second you spend in the warehouse, you lose a point.

M: Okay. and they want to keep the most amount of points.

Dr. D: So they want to try to uh minimize that going for the most efficient route.

M: And coming on the most efficient route.

Dr. D: That is so interesting.

M: It's called reinforcement learning.

Dr. D: I love that.

M: NSD is together with machine learning and it's one of the types of AI. So these systems go through a uh LLM.

Dr. D: Word prediction and then they go to a reinforcement learning phase.

M: And then they go into a third phase with actual humans Nigerians who were able to to basically fine-tune.

Dr. D: The system.

M: Then how do you say humans learn different then.

Dr. D: Humans we our brains evolve at the same time we grow.

M: Mhm.

Dr. D: Okay. So, we are not the same person when you were born. Then now they are born with the all the knowledge.

M: Mhm.

Dr. D: They are trained with knowledge and then they are constrained with the reinforcement learning part. So once then they just stop learning because you cannot teach anything to an LLM.

M: When you interact with with a.

Dr. D: Gemini. GBT or whatever.

M: Uh, usually your prompts are saved.

Dr. D: Yeah.

M: Then some expert, some AI people are going to get your prompts. If they are good, they're going to use them for training later as examples of conversations.

Dr. D: Mhm.

M: But they're not going to it's not learning from you in real time like like we are.

Dr. D: Okay. So the gap of learning is different.

M: Training happens first. Um then.

Dr. D: Production or.

M: Mhm.

Dr. D: Inference works uh later.

M: Okay. So they have separated those two processes.

Dr. D: Exactly. That's why you can never teach it. So you can tell them you can tell them don't call me like this. It won't. But then next conversation is going to do it again.

M: I see. I see. Yeah. Yeah. Yeah. Yeah. Yeah. I think I think I understand.

Dr. D: Okay. So it has a context.

M: It has a context. Uh because a prompt Yeah.

Dr. D: prompt will say don't call me like that. Although don't for example I I hate when they flatter you too much or they order too Yeah. So I Yeah. Me too. I hate that.

M: I tell them always please be like don't don't flatter me.

Dr. D: I'm like please say thing like bad things hopefully.

M: And they they stop doing it because it's a prompt. Not because uh I change the behavior. I can't unless I am uh the owner of the M.

Dr. D: I see. I see. So, okay. So, that that is one thing.

M: And they're not learning.

Dr. D: Yeah. And they're not learning. They're just being reinforced which is way.

M: No, then they're not they just uh all the data is being saved.

Dr. D: And then a separate time when they update the version or whatever,

M: they can put all the data back.

Dr. D: Perfect.

M: Some of the data they want.

Dr. D: So, it's not the same as talking to a person. No, that for sure. So, one, they can't learn at the same pace. Any other.

M: Differences?

Dr. D: Any other differences? Yeah. On on how they how we learn.

M: So, yeah, because the the way we learn, we're babies. We learn how to to do things from from the beginning. We learn everything from scratch. We learn to see, we learn to hear, we learn to understand sounds, understand colors, uh sh shapes, uh, textures. Uh, then we go to like a infants, you learn to speak, you learn what means, you need voices, you need to understand, to recognize words. You see, our brain is developing.

Dr. D: Mhm.

M: And that part is actually the way we are created. We that's the most the part where you actually become you.

Dr. D: Uh, they don't have that. That's why they're so bland. is why they have personality. This why like um opinions or like a personalities are are created on that time. You usually inherit.

M: Uh your parents' quirks,

Dr. D: your siblings quirks,

M: you and that makes you yourself you.

Dr. D: Yeah. Unless we start like doing that with a AI which is we're very far away from doing that we will never have um any anything similar to a computer who is able to have a personality or an opinion or to stand by their word.

M: Yeah, I see.

Dr. D: They are tools.

M: Yeah, they're nothing but a tool. Mhm.

Dr. D: Then what since there are to tools since they are tools, what ways of thinking from AI can lead us borrowing them?

M: We have to use them. So it's a new technology.

Dr. D: Change the world.

M: We're going to be still talking about it. 50 years from now,

Dr. D: maybe in 50 years they're going to be far better. I think it's going to be incremental. It's not going to be like a another breakthrough every year or something. I think the main gain we got from the AI was the fact to be able to process text semantically instead of syntactically.

M: Mhm.

Dr. D: And to make use of all data we had.

M: Yeah.

Dr. D: And to be able to understand the meaning of that from a computer. Yeah,

M: that's the main value we can get from them.

Dr. D: Yeah.

M: Now we are still going to get better and better and better, but slower. It's like the difference between the first train.

Dr. D: Mhm.

M: And the trains we have now.

Dr. D: Yeah. Trains we have now are good.

M: Yeah.

Dr. D: Safer, better, whatever you want.

M: Faster.

Dr. D: But they are not incredibly better than the first trains.

M: Yeah. The first train.

Dr. D: They're not in incomparably better.

M: Yeah, exactly. But the first train was incomparably better than a horse.

Dr. D: Yeah, I see. I see.

M: Okay.

Dr. D: I I I like that. I like that description. So, we better use them and we better use them and get used to them.

M: They're going to start to be more and more and more part of our life. It's going to be part of our normal life. It's not going to change um.

Dr. D: Any.

M: Yeah the world's going to change. We're going to still have a societies. We're going to still have jobs and still have a life. Only.

Dr. D: It's just a more technology like Yes. If you were a dog, a horse vet, you probably lost your job when motormobiles came because before there were thousands of millions of horses everywhere.

M: Yeah.

Dr. D: And now they're not.

M: Yes, some jobs will disappear, but doesn't mean that.

Dr. D: Uh all jobs will disappear. Yeah.

M: So we have this tendency of exaggerating the good and the bad.

Dr. D: Thinking things like go they want.

M: I see. So one last question for this chapter. What do you reckon is the thing that we could take from decision making in AI?

Dr. D: Yeah. Um limitations of AI. So there are tools and you need to understand what they are able to do or what they're not able to do. M.

M: So you need to have this caveat in mind when you use them for decision making. So if you want to for example go analyze the competence uh read about the other people on your business use AI to get from the internet all the sorts of information about the other companies who work in your field and understand where they're going. AI is going to help you a lot to understand the different.

Dr. D: By the way.

M: For you guys other podcast I'm doing that on you.

Dr. D: So if you do.

M: Just stating that.

Dr. D: AI is going to be a fantastic tool for that.

M: Fantastic.

Dr. D: So everybody's going to be using it.

M: So.

Dr. D: So the complexity is going to be.

M: Complexity is going is going to go up everybody's gonna have the tool.

Dr. D: Not using it means it's like it's like going by horse when everybody has a a modern car.

M: Mhm.

Dr. D: I mean, you can do it for lacer, you can love it,

M: but it's going to be standardized.

Dr. D: But Exactly. But you can't use it for for the same job before because now everybody has the new technology.

M: Yeah.

Dr. D: So, it's going to make your life easier, but then the amount of um complexity is going to go up. So, on your extra time, you'll have to do more things.

M: Yeah. It's like what happened with the with the right home or the right whatever.

Dr. D: When you had productivity goes up.

M: Yeah.

Dr. D: But when productivity goes up price goes down.

M: Yeah.

Dr. D: And that that's the issue. That's what people don't see. For example, if I.

M: Uh for example now you can buy a a this glass.

Dr. D: For I know 5p probably in the China something.

M: Whoa. Whoa. Whoa. Whoa. Don't dismiss my glasses, by the way, 95p.

Dr. D: Because they are.

M: In England.

Dr. D: Uh mass produced,

M: but if you have to ask somebody to make the glass by hand,

Dr. D: it's going to cost you a lot of money.

M: Yeah.

Dr. D: Uh because the increasive productivity made them cheaper.

M: So this going to be the same many tasks, many things we we used to do and charge a lot of money for it.

Dr. D: Mh. now are going to be cheaper because everybody can do them or a lot of people can do them.

M: Yeah.

Dr. D: And I know web pages or or or dashboards or um analysis.

M: Like for example right now writing an essay is is ridiculously easy to do.

Dr. D: I'm talking about quality but the fact of creating it is easy.

M: Is easy. Yeah.

Dr. D: So they're not valuable. Yeah.

M: So understanding that um is a technology you will have to use you have to use it correctly but that means it will free you time but that time is going to be filled up with more things that we'll have to do you will have to do in the future.

Dr. D: Yeah.

M: Is uh.

Dr. D: I think you should uh say hello to our new copywriter because we just.

M: They probably know that. Yeah, they probably do they probably uh know that as you said.

Dr. D: Sadly that's how it is.

M: But.

Dr. D: And the more time it passes.

M: Uh the more obvious it's becoming.

Dr. D: Especially with the AI automation which is happening now so it's not only you using a AI prompt on your chat.

M: There's other AI.

Dr. D: Using AI.

M: Um yeah.

Dr. D: Um.

M: Let's get straight into it uh And let's explore images and videos mainly because I really like to know.

Dr. D: How how they work.

M: Um yeah, how they work and what is going to be the next few steps towards a little bit of a better videography I would say.

Dr. D: So first let me very briefly explain you how they work.

M: Yeah. Yeah. Yeah.

Dr. D: Okay. So uh the the the AI of image generation started with images. Okay. Because videos are basically images stacked.

M: Yeah. So the the AI of image generation.

Dr. D: Started from a lot of different models which happened in the 90s and 2000 in the 2010s.

M: Which basically tried to solve the problem by doing many many kind of things and they were directly related with with machine vision. So machine vision is the art science of getting an image and understanding what it is image recognition.

Dr. D: Okay. and they tried to do to crack image recognition and image generation together. Okay, so image recognition.

M: Uh was basically tackled with normal um neuronet networks.

Dr. D: Uh but the amount of data an image has is huge. So if you talk about uh pixels, each pixel can be three colors and each color so it can be blue, red or or green.

M: Yeah.

Dr. D: And each color can be uh anything between zero and two 55. So that gives you a vast amount of data.

M: Mhm.

Dr. D: And a normal picture you can take with your phone has about 20 megapixels. That is 20 million pixels. So each picture has 20 million pixels times three per color. So it's 60 million pixels of information just for for a picture you take with a cell phone. So.

M: How are we possibly getting that into AI?

Dr. D: Exactly. That's that's the issue. So we were in the we were back in the 90s.

M: Yeah. 1998 like like Windows 98 was all the fury and it was.

Dr. D: Computer will still look like like shoe boxes and we were there and what happened was they discovered a way to basically reduce the size of these huge images.

M: Into something AI could manage that's called convolutional uh neural networks.

Dr. D: Okay. Basically what it does is a mathematical operation called convolution which.

M: Tries to extract information especially edges and.

Dr. D: The same babies recognize edges and and shapes.

M: Yeah.

Dr. D: Uh try to recognize that and gradually make the image smaller but it makes more of them.

M: Okay. So with time after um several layers of convolution you get very small images of maybe 10 time 10 or something like that.

Dr. D: Mhm.

M: But you get thousands but you can.

Dr. D: And then if you stack them together.

M: No, you just analyze one by one and you train your your AI with those parts of the images.

Dr. D: I see.

M: You know you know what I saw? I I saw something a little bit different uh while doing my research. I think it was uh Yeah, I think it was in your PDF.

Dr. D: Mhm.

M: Is it Is it the same uh thing?

Dr. D: It's exactly what I'm talking about.

M: Okay. So, basically the thing is I saw pixels.

Dr. D: At at the beginning.

M: You basically stack them together.

Dr. D: Yeah. It's not stacked. It's just one after the other.

M: Yeah. Okay.

Dr. D: You just don't you don't compress them.

M: Yeah. Yeah, you don't compress them, but you.

Dr. D: You just go then you analyze one after the other.

M: After the other after the other after the other and then it gets.

Dr. D: You use and then you use the AI to generate the for example imagine you have thousands of pictures of cats and dogs and you want to a computer to know.

M: Uh you just send a picture and the computer has to know if it's a cat or a dog.

Dr. D: Yeah,.

M: Typical. Mhm.

Dr. D: So what you do, you start to to do the convolution.

M: And the system will learn which parts of a cat or a dog like a type of eyes, show of eyes,.

Dr. D: Shape,

M: shape are more typical to of a dog or of a cat.

Dr. D: Yeah.

M: And after a while able to tell the difference.

Dr. D: I see.

M: Okay. So that was okay. It was 90s 1988 more or less they invented this.

Dr. D: Mhm. And more or less in 2012 2013 techn it got more complex it got better.

M: And it was able to to was a revolution.

Dr. D: Yeah.

M: And it got much better just in a couple of years and got superhuman. Now computers are better at recognizing things than humans mostly because they have infinite memory. So they know things.

Dr. D: Yeah.

M: We we don't. So I wouldn't tell I I know all the ra the dog types or the cat types computers do.

Dr. D: Yeah.

M: So that's why they're.

Dr. D: And they have every possible picture that has been ever uploaded.

M: Exactly.

Dr. D: Any dog.

M: So they know everything they can for example now nowadays and you can try it. You can go to any part of the world. You just take a picture of your phone. You just show it to Gemini Chpt. um whatever you you use and it's going to tell you where you are.

Dr. D: Any part of the world.

M: I mean of course if you take your picture of your shoe now but like a building or um.

Dr. D: I will try it out with with a street which we won't say because we're recording remember that.

M: But the but you go for example to a city center and you just take a a picture of the city center is going to tell you where you are.

Dr. D: That's crazy.

M: Yeah. that I find crazy.

Dr. D: Try it.

M: But but.

Dr. D: It's doing convolutions.

M: Yeah. I see. I see. Okay.

Dr. D: So, uh there was 2012, 2013, 2018.

M: And the the system got much better, but there were several.

Dr. D: uh technologies trying to do the opposite, which is creating images.

M: Yeah.

Dr. D: And they didn't work. I don't want to go there because it's a long long story, but I'm going to go into what worked.

M: Okay.

Dr. D: What it worked was a system called diffusion which basically they get an image,

M: they add a bit of white noise.

Dr. D: Okay.

M: Okay. And then um.

Dr. D: what is white noise in this case?

M: It's like the static of a television.

Dr. D: Perfect.

M: And then there's just a tiny bit and then they retrain the image with the noise.

Dr. D: Yeah. So the the computer know learns how to go from the normal picture to the picture with a noise. Then they do it again. So the picture with a tiny bit of noise to a picture with a little bit more of noise. And they train again. And they do it all the way until the picture is full noise. Looks like a like a totally scrambled.

M: Yeah. And has no way of recognizing.

Dr. D: There's no is totally noise. So they are able to to train the model to to.

M: Create.

Dr. D: To to see to know what it means to transform an image from an image.

M: To to noise.

Dr. D: But when the when the network was trained with thousands of millions of different pictures.

M: All step by step transforming into noise.

Dr. D: And several different ways you can you can go into noise so many times you have almost infinite training data. Yeah.

M: So, you have a huge amount of pictures, all of them trained with added noise.

Dr. D: All the way until they're pure noise.

M: Mhm.

Dr. D: Uh you get to a point where you can start from noise and get to any of your pictures.

M: That's so interesting.

Dr. D: Um and and that's what they did. The the problem is that they didn't have any control over it because you could start from noise and you could train and it was going to appear a cat, a dog, a building. You didn't know because there was no prompting. Yeah.

M: So they had to invent the prompting which is actually harder than the stable diffusion.

Dr. D: Okay.

M: And the prompting is basically an LM which has.

Dr. D: um.

M: recurren rewards.

Dr. D: A reward system.

M: Yes. So I love that day. Okay.

Dr. D: So they were able to to train it to basically do what the system says.

M: Yeah. Okay.

Dr. D: That's why every time you prompt a cat.

M: chat GBT image generation, J mini generation or stable diffusion.

Dr. D: Or like um they give you a different image.

M: Yeah, they do.

Dr. D: Because they start from different noise. Oh my god,

M: that is so cool though.

Dr. D: That is It's so cool that that Oh my goodness.

M: That's interesting. That's very interesting.

Dr. D: So that that's how they they do it.

M: And then well videos is just a compilation of the pictures. It's much it's more complicated because videos also.

Dr. D: They use the previous.

M: The previous.

Dr. D: Um.

M: oh the second one.

Dr. D: And then AI doesn't know they have to do that all at once.

M: They Yeah it's.

Dr. D: Because AI doesn't know time.

M: Yeah but they have to.

Dr. D: That is so crazy.

M: But they use a previous to inform the second and they use.

Dr. D: To inform the second and second. Yeah. So like Okay.

M: So although and.

Dr. D: So they only use two images basically at a time.

M: No, but the problem is that that's why usually they lose context. So it's hard to make long images in AI. That's why you cannot make a movie.

Dr. D: Yeah.

M: AI engineering movie.

Dr. D: Yeah, you do still need.

M: You need a lot of work because and it still doesn't understand physics. So for example.

Dr. D: The noodle thing.

M: Things appear and disappear. Yeah.

Dr. D: So, for example, if you want to tell if a if um picture is AI, you have to look for physics. For example, hair doesn't have strands.

M: Well, mine doesn't, but yours.

Dr. D: Uh have have strands. Uh AI hair is like a appears and disappears.

M: Yeah.

Dr. D: And and things like that.

M: So, okay.

Dr. D: Uh fingers were fixed.

M: Sorry.

Dr. D: Fingers were fixed.

M: Yeah,

Dr. D: that's a good thing. Okay, that's good.

M: Before fingers were a big problem.

Dr. D: Yeah. Well, but there there are a couple of AIs now that pass a noodle test.

M: Yeah.

Dr. D: Yeah. By the way, um the noodle test, if you could maybe.

M: No, it's just a test that it wasn't able to do it because they didn't have enough training data.

Dr. D: So.

M: it's just that it's like clocks.

Dr. D: Okay.

M: Most clocks are.

Dr. D: What was the test about?

M: Just generate uh that noodles. You're talking about the.

Dr. D: Yeah. and eating eating news. Yeah.

M: Yeah. But that's basically lack of training data. The algorithm has been the same all along.

Dr. D: Oh, okay.

M: It's not because they improved the algorithm. It's because they didn't have the data for that.

Dr. D: Oh, okay. So, they invented a test that they knew they could pass.

M: Was was famous. So, they they just for what you do.

Dr. D: Uh I the thing is I love marketing so I can't hate them for it. But I also now I'm understanding that this is all like a lot of crap.

M: It's a lot of marketing but that's.

Dr. D: it's a lot of marketing and then that's it.

M: Yeah.

Dr. D: They even can create stuff because they have well they have information about all the data and then they solve their own like their own made problem. It's I don't know it it gets weird.

M: Is progressing. Yeah, I know. I know it is.

Dr. D: But it's not by the metrics you you think it is.

M: So then how quickly is AI improving?

Dr. D: It's not quick. Um you won't see the same kind of um quantum leap you saw in 2022.

M: Why not?

Dr. D: Because there's nothing ripe for it. All the the changes we're seeing right now are going to be incremental. So, we're going to have a better LM, a better image generation, a better video generation.

M: But what about what about what we were talking just before uh just before we started the podcast, which is.

Dr. D: Sorry.

M: Yeah. Like what's.

Dr. D: um.

M: Which is so you you did talk about the three main factors that basically build AI.

Dr. D: The hardware.

M: Yeah.

Dr. D: yeah.

M: The mathematics and the data. So what if we get the mathematics or the equations or the algorithms right?

Dr. D: We did.

M: We already did.

Dr. D: We already went that we have a system that's able to understand pictures. We have a system that's able to understand text. We can't have a system is able to understand videos and make some AI. Uh we have a system is able to understand audio. We have a system that able to to generate videos.

M: We can generate pictures.

Dr. D: But maybe the same system won't work towards that quantum leap that is going to give us the Star Trek.

M: It's going to start getting better and better and better at what it does. But we already did everything. We have we created the the the steam engine already. There's nothing else to to create at the level technological level we have.

Dr. D: Of of the basic level.

M: Yeah. We have to wait until it gets much better and then going to invent the airplane.

Dr. D: So we just need a ton of engineers.

M: A ton of.

Dr. D: working hard on it.

M: I mean it's it's like it's like when you created the steam engine.

Dr. D: Mhm.

M: until you invented the airplane.

Dr. D: Or the rocket.

M: Yeah. Or the rocket.

Dr. D: Yeah. Uh.

M: I see.

Dr. D: there are discoveries that are ripe.

M: Yeah.

Dr. D: For example, between the there are people who were alive.

M: Yeah.

Dr. D: before the invention of the airplane.

M: Mhm.

Dr. D: Who saw.

M: the man on the moon.

Dr. D: Yeah.

M: So change can be very fast.

Dr. D: Yeah.

M: And we just went through that.

Dr. D: We just don't understand how momentous this has been. Yeah.

M: Uh it was.

Dr. D: Yeah. People don't understand that.

M: You.

Dr. D: this invention was.

M: Basically the the steam engine but for.

Dr. D: But there are things that you make you are able to do but then for example you can't keep for example the the space the space race we put a man in 69 will still not we're not colonizing Mars or even the moon.

M: Yeah.

Dr. D: because there are a lot of technical problems we can't do.

M: yeah of.

Dr. D: Uh, and we have to crack those to get to.

M: get.

Dr. D: Are there rewards?

M: There are lots of rewards. There are metroides like asteroids over there which are made of gold.

Dr. D: Yeah.

M: And diamonds and.

Dr. D: Yeah. and stuff like that.

M: But.

Dr. D: but technic.

M: technically we're not there yet. Same here.

Dr. D: We we just cross everything we could do at the level we are. We are really much higher than we were 10 10 years ago.

M: Is incredible. And this type of leap is not going to happen again at least in the next decade or or or more.

Dr. D: Maybe it happens tomorrow, but who know?

M: There's nothing else to do. There's nothing else you can say like um.

Dr. D: Okay.

M: You see what I mean?

Dr. D: Yeah. Sorry. Yes. Yes. I I do understand because now it's about incrementing.

M: We We can It's like can can computers understand when you speak to them? Yes. Can computers talk back to you? Yes. Can they understand what you show them? Yes. Can you understand? Can Can they generate pictures? Yes. Can they generate video? Yes. Can they understand videos? Yes. What? What else? Can they produce text? Yes. Can they understand text? Yes.

Dr. D: Well, now as a mix with robots.

M: now. Yeah. But then we're very far away from that.

Dr. D: Okay. So, talk talk me through that then. What what what does the future look like to you?

M: Future for uh.

Dr. D: and how quickly can we arrive there? So,.

M: so what what would you reckon is going to be the next one and then the next and then the next?

Dr. D: Okay. I've I I said that back last time in the context of losing their job.

M: Yeah.

Dr. D: That.

M: um is is incremental from now on.

Dr. D: Yeah.

M: So if people not going to expect computers to keep going so fast that they were going to be human like in.

Dr. D: um 10 years and everybody's going to have no job. Yeah. because it's is not going to happen because of the of of the speed things go.

M: It's gonna have a moment sooner or later when uh the next breakthrough will happen and the next breakthrough which we are not there is the fact that you going to be able to automatically have AI uh improve AI without supervision. How far are we from that?

Dr. D: Pretty far because uh AI can cannot understand the AI code the AI systems yet and depends on data. People are are there's a huge research on that.

M: Yeah. But usually the idea is that once AI is able to to improve AI.

Dr. D: Uh AI is going to improve exponentially.

M: Yeah.

Dr. D: And there's where the real AGI ghost happens. Okay. Because then humans are out of the loop.

M: I see. And that's basically the.

Dr. D: As long as they still need data from us.

M: Not only the data but the training and the and everything.

Dr. D: Yeah. Reinforcement etc etc.

M: Uh but nowadays if you have clot code and you make them uh try to fix AI models and things like that. No, it is not it's not working. They are trying and you're gonna find a lot of scientists hyping up their research because they need funding.

Dr. D: Yeah.

M: But they.

Dr. D: I love that. I love that you say that as a as a researcher.

M: Do you do that?

Dr. D: Everybody does that.

M: Uh but you do that with yours.

Dr. D: Honestly, it's not something you.

M: Yeah. Okay.

Dr. D: You do. Uh it's not We're not there yet. Okay.

M: And if you we and it's not going to happen something people going to know about it.

Dr. D: Okay.

M: So let me let me rephrase that then.

Dr. D: It's basically a but it's going to be something similar to the nuclear bomb.

M: So that that kind of a huge thing countries or companies are going to race to get it.

Dr. D: Yeah.

M: And that was basically fuel the AI bubble. They thought they were near that.

Dr. D: Yeah.

M: And they.

Dr. D: and now they realize they're not but they keep need to keep going.

M: To.

Dr. D: so yeah.

M: to sell it to sell it to people basically that's why they changed the ruling for.

Dr. D: for when IPOs can happen or.

M: Yeah. Oh my god that is going to be such a big crash and I I just know it.

Dr. D: Yeah. But that but okay. But basically the.

M: thing is I don't know how to prepare for that beyond holding cash.

Dr. D: Yeah. More precious metals and your in your backyard.

M: Allen cash precious metals.

Dr. D: Yeah. Prep and save food and.

M: Yeah. stuff like that. You know.

Dr. D: the thing is because it's the biggest IPO of all times and it will be worth basically.

M: one word. Well, thing is it's not that IPO.

Dr. D: The problem is that Chant GBT is also in in that a lot of the a lot of the money raised for uh companies like Google because of Gemini is because again their AI agent and so on so forth.

M: So basically don't don't do it. Don't invest in that. the IP big IPOs don't pay after a well you know that if you invest in the IPO you think you're getting in but if if you know it's hype.

Dr. D: Then eventually maybe not in one day but maybe in one year the companies will end up going down to their original value.

M: well companies always tend to go to their to the to their minimal value and if we keep surviving. So the best test for a company is always time.

Dr. D: Yeah.

M: And the reason why that is is because.

Dr. D: what a company does is literally.

M: add elements and give value that is more than just the sum of the parts. Uh but but you know in the past uh IPOs were a huge opportunity because you could get a company the stocks directly from the company a cheap.

Dr. D: And then resell market.

M: Now companies learn that so they overhype themselves.

Dr. D: Yeah.

M: Uh sell them expensive and then the market adjusts and you lose unless you are a trader and you.

Dr. D: the thing is Yeah. The thing is the market changed because of angel investment.

M: So angel investment all that all that pre All that pre-investment before IPO very hyped and then.

Dr. D: and then the the company they say is worth X and it's actually worth a T.

M: It's I mean the thing is what a company is worth nobody knows because nobody can predict its future value. AI as it stands now AI has.

Dr. D: um not a lot of value. It's is not zero uh is very valuable but the technology is valuable not the.

M: company.

Dr. D: Mhm.

M: So it's like the now we know how to make LLMs.

Dr. D: It's open source. You can download your own transformers from the internet and create your own.

M: That that I find really really cool. So the fact that everybody can make their own AI.

Dr. D: And that and that's my prediction for the future. My prediction for the future is and I I'm not a predictor, okay? So I I I'm not claiming I know the future, but my prediction is um AI is dem going to democratize. So the cost of building your own AI is going down very fast.

M: Yeah. So before it was something only Google or like OpenAI could afford.

Dr. D: Yeah.

M: Uh now is getting into a couple million pounds.

Dr. D: Mhm.

M: Which is not something like.

Dr. D: not something crazy like Coca-Cola could do their own.

M: Coca-Cola could definitely do it. But I mean.

Dr. D: but even like a small building company now.

M: small. Yeah. So it's not incredible. They can't. It's still better for them to just pay. But if they for example have want to have um let's say um expert system which all the information all the manuals all the things they have.

Dr. D: And they just want to be able to chat with them on their phone to get uh the correct things or to show them something and to know if the norms are correct or things like that they can't build that.

M: And it's going to cost them like say 10 million.

Dr. D: Yeah.

M: Uh but in 10 years probably cost them like 10,000.

Dr. D: Yeah. So, and then everybody's going to have their own. So, that the business model of using your data is gonna crack.

M: Maybe you're going to have like a like a battery, so to say, in your backyard that has your own AI.

Dr. D: Yeah.

M: Your own agent.

Dr. D: Yeah. I mean usually big companies are going to try to force them theirs.

M: But it's not like social networks where there a network factor where everybody wants to use the same network.

Dr. D: Because otherwise.

M: if if I don't want to use uh WhatsApp want to go to you know signal.

Dr. D: Yeah.

M: but my family is on WhatsApp.

Dr. D: Yeah.

M: I need to to stay on WhatsApp to be able to talk to them.

Dr. D: Yeah.

M: Uh.

Dr. D: here is you can talk to your own AI and nobody you don't care if there's no network effect.

M: So people will probably start to create their own companies will start and then when it goes down and down and down.

Dr. D: uh people will start creating their own. So and then they're going to be suited to you to your own personalities. They will have your own data.

M: Oh that's going to be so cool though.

Dr. D: It's going to be.

M: And that's in 10 years.

Dr. D: So like.

M: I I can't tell. doesn't ask me because I I I don't have a crystal ball but I think I believe direction is there because I feel the.

Dr. D: the.

M: but it's it's not in five years.

Dr. D: I don't know. I don't know.

M: You don't know. Okay.

Dr. D: What I'm telling you is that the the technology is getting cheaper and more.

M: And cheaper fast.

Dr. D: and more efficient. So you don't need to buy your own data center for that. and data centers are going to be excess capacity. The the original CH GPT was very un uh inefficient.

M: Yeah.

Dr. D: And now they can make them much better.

M: But okay, I know you're not predictor uh but I do want to know what does an AI researcher do? Not not do you what what are you what are your thoughts.

Dr. D: about what.

M: on when it's going to be?

Dr. D: I have no idea.

M: I have no idea. It can be if you tell me.

Dr. D: Could it be like a year from now?

M: I mean right now.

Dr. D: Yeah.

M: If you have a spare computer.

Dr. D: Yeah.

M: Like let's say a good computer with a you say you're a gamer.

Dr. D: Yeah.

M: And you have a nice gaming computer with a good GPU.

Dr. D: 3K computer. Yeah.

M: Yeah. with a GPU, very big fat GPU,

Dr. D: you can build your SML, so small language mode.

M: Which would do an AI agent or something like that.

Dr. D: You can you can have your own AI agent. You can use it for for I know doing your usual task. You can use it on Telegram. You can text.

M: Can you use it in WhatsApp?

Dr. D: No, because WhatsApp is No, doesn't allow you. Can you why not?

M: But Telegram has a bot.

Dr. D: so you can program your own bots. So you can chat with your own with your own agent.

M: That's so crazy. And then you can use it for for you can give it tasks and then uh it can ask you can train it to do things for you like sort your mail or.

Dr. D: or do your your.

M: I do want to I do want uh to start doing that sorting sorting mails and.

Dr. D: some messages that are repetitive basically.

M: So hey guys do you want to be in my podcast?

Dr. D: You do that and and that's something everybody can do now and it's quite cheap. I'm talking about if you want to rent either you have your own computer home sitting home doing it for you.

M: Yeah.

Dr. D: or you can just have.

M: have.

Dr. D: some know some uh cloud.

M: pay pay for some compute.

Dr. D: and it's your model it's your data I'm not talking about using cloud or using jptt I'm talking about.

M: your own my own data system.

Dr. D: when your own computer you can do that now but or you can uh you can train from scratch your own SLM as as I do.

M: And you can train for some tasks for example I'm using it for predicting time series predicting the stock market for example using uh SLMs.

Dr. D: no it's research I don't I I don't pretend to to really make money from that, but it's for.

M: Yeah. You told me that you you you forgot about doing something.

Dr. D: No, because it's uh there's a lot of luck involved.

M: Yeah.

Dr. D: So, and and even if you do it, you can you wouldn't need like a maybe several tens of thousands of pounds.

M: Yeah. Or you can have some automatic trading system running, but you're going to make like maybe 50 pounds a week a month or 50 pounds a week. Um,.

Dr. D: yeah.

M: Um, so one of my last questions is about their advice, your advice to future leaders.

Dr. D: Future leaders with AI.

M: Yes. What do you reckon is like the main advice that you would.

Dr. D: AI cannot lead? So that's your job. It's not the computer's job. Automate whatever you want with AI. Use it. It's a tool. But uh leadership and charisma are profoundly human things. You need to cultivate people because that's the thing AI won't do.

M: You can outsource reading, texting, writing, whatever you want. But the human touch just texting somebody being there.

Dr. D: at the right moment.

M: at that right moment that uh leading understanding your employees or like a customer's problems that is something AI will never give you. You can try to make your own very lean uh funnels for clients, but they're going to feel fake because they are.

Dr. D: Yeah.

M: Uh and that's a big problem I see. People are using AI wrongly.

Dr. D: Yeah.

M: They are trying to automate the human part and instead of automating the computer part.

Dr. D: That that's the that's my point. the if you automate the computer part and you get uh all the I know numbers and accounting and the many many things out of the way automatic emails and things.

M: It has.

Dr. D: all the all the part but you keep the human thing.

M: then you're going to be much valuable much more valuable than than if you do the other parts. So it's it's not about getting rid of the accountant, but it's getting rid of him doing the accounts and him in putting him to talk to the client.

Dr. D: or doing have having one account and not 10.

M: Yeah.

Dr. D: Or having accounting made easy. So making making it easy for for the accountant.

M: Yeah. So you can uh be compliant very fast or having for example AI CRM which allow you to understand exactly what is the each client wants.

Dr. D: Yeah.

M: And but then acting directly so using as a database to understand what the client wants and then acting from there.

Dr. D: Yeah. I I got to say uh CRM are customer relationship management systems.

M: So uh.

Dr. D: I just got to say that because.

M: I worked in sales and I know well I understand it but like it's not.

Dr. D: it's not Yeah.

M: Yeah. I don't I don't know if other people know it. Yeah. So basically that that's the the thing.

Dr. D: if you want to lead you remember that um leading means uh you have to why people choose a leader and people choose a leader because a leader uh can make a decision and also being part of a community which is a community under the leader.

M: Under the leader. Yeah. And that uh is something uh profoundly human.

Dr. D: under the vision I would say of the leader.

M: And the and the and the values and whatever and the of the company and the morals and whatever you you're part of a team. You're part of a pack.

Dr. D: Yeah.

M: And we are pack animals.

Dr. D: We're mammals.

M: and this pack animals we are uh being part of a group um makes you feel less anxious, makes you feel better. That's why people follow people. Leaders need to uh cultivate that because people if you're following somebody who's worth following that makes uh that makes sense for people and that's not going to change. That's not going to change with AI. It's not going to change with anything. It's that what we are.

Dr. D: Yeah.

M: And and if you want some like a like a nice uh goalpost to that's it.

Dr. D: That's great.

M: Really love that. Ignasio.

Dr. D: Yeah.

M: Dr. Nacho.

Dr. D: Ignasio, thank you very much for being.

M: Thank you so much. I think we've all learned uh quite a lot. Yeah, it was well it is an incredibly difficult thing that for sure.

Dr. D: It's uh new but I think in some years it's going to be more democratized.

M: M you have a doctorate in physics.

Dr. D: People are going to to see it better. I mean it's going to stop being new.

M: Yeah.

Dr. D: Many things happened in the last 10 years.

M: Uh and most of these revolutions were silent. So we crossed thresholds and I knew it because I saw the papers.

Dr. D: Yeah.

M: I I was there. I saw the paper in 2017 which created Transformers. I read it in 2017. I understood in 2017.

Dr. D: You read it there?

M: Yeah.

Dr. D: That's great.

M: I didn't I didn't understand it was going to be an impact.

Dr. D: Yeah.

M: Well, Ingasio, thank you very much for being on the podcast. Uh guys, thank you very much for listening. uh to this week's episode. Um for all of you out there that have stayed until uh this bit, please um follow, comment, like, share, all of that. Put a fivestar review if possible. uh that really helps us uh the podcast grow and be shared uh to other people who might love this content h at no extra cost to you. So, thank you very much and um see you next time.

M: Thank you for listening to this week's episode. Beyond the Horizon is a podcast about how experts think and what the leaders can learn from the way they see the world. If you enjoy this conversation and want to support the show, you can do that through our Patreon. The link is in the description. That's also where we share deeper breakdowns of each episode and where you can join our monthly team meetings to help us improve the podcast and the quality of our conversations. Thank you for listening and we'll see you next time.