Hello everyone, I'm Pablo Apiolazza and welcome to Perspectives podcast. This is the third episode and with me we have a fantastic guest. His name is Marco Franzon, and he's, an expert in cloud computing. Uh, let me introduce him properly. Today we will talk about a series of topics that are around generative AI and cloud computing, which is pretty much the the base of all computing that is relevant for, uh, human life. Actually in lately, right. If I'm not wrong. So first of all, I would like to introduce Marco. Marco is that engineer at the lab, a company specialized in high performance computing, and he's also the co-founder and CTO of dualistic, which is a new startup from 2023 that specializes in digital twin technology. He has an educational background in biochemistry and data science from the University of Trieste. And, uh, he's, uh, very keen into MLOps, MLOps, uh, machine learning and the transition of ML models from research to production. So. Well, without further ado. Um, how are you, Marco? -Let me. -Oh, fine. Thank you. To be to invite me here. I'm very happy to join you in this podcast. Um, yeah. I think that we have a lot of topics to discuss today. Interesting topics. So let's start. Let's do it. So okay, let's let's begin with, uh, with the most basic things because this, this podcast, uh, has the, the objective to, uh, get deep into subjects, but starting from, from the beginning. So. Talking about, uh, algorithms and processes and, uh, everything that makes the world go round today because so forth is, uh, uh, necessary and critical to every activity of the humankind. Right now, there's a big difference that has been developing in the last 20 years, and I'm guessing ten, 20 years, which is this paradigm between local and cloud computing. Right. Can you explain me a little bit about this? Yeah. So the main difference between local and cloud computing, I like the statement that cloud is a someone else's computer. So, uh, the, the main idea is that, uh, when you run something on cloud, you are deploying your application in a service in another, uh, in a cluster of computers and by someone else and, um, instead. Yeah, the main difference is that you, uh, give the, the the, um, I mean, the, the permissions to run your code to someone else. And also, uh, you had not to care about the status of the machine. And this is the main, the main thing important for, uh, for a developer, uh, because if you are the maintainer of the code 90% of the time, you don't want to maintain also the infrastructure. Uh, so if you use cloud for this reason, uh, you can avoid to think about, uh. Oh, yeah. My my my PC is on failure. Uh, so I had to, uh, update something or to fix some bug that, uh, can take a long time instead, if you are, if you run on cloud, you have. No, none of this problem. And this is one of the main point. Uh, why the because the the the cloud is so, so, so much hype and. Yeah. Is continuous developing. You. You work in, uh, in cloud, which is a company specializing in high performance computing. Right. So that was something that wasn't as necessary. Uh, it wasn't a popular application of of, uh, of, uh, cloud computing, right. So I would like to, uh, to explain to the audience what, what high performance computing is and what is the difference between, like, regular computing and this type of computing and how this, uh, is changing right now with these new -technologies? -Yeah. So high performance computing is a subset, I mean, of computing, um, which is um, um, I mean, regard a certain kind of application, uh, in general, um, we talk about, uh, HPC, so high performance computing, um, when we, uh, consider clusters. So, um, multiple, uh, computers connected to each other. And this is the fundamental of HPC because, uh, in general, your application had to be handled by multiple PC because the workload. So the computational time and the computational amount of calculation that you have to do is, uh, really huge. So a single computer could not handle this workload. And this is the the bottom line of HPC. Um, for the time being, high performance computing becomes relevant because, for example, in for machine learning task, when you have to train large models, you need to spread the workload, uh, on many different machine because otherwise you cannot handle this, uh, this large training. Uh, so uh, HPC comes fundamentals. Uh, also for this application, uh, this born for, uh, physical simulation, which requires a lot of, uh, computational, uh, time, uh, typically, but now it, it becomes relevant also for machine learning. This is something interesting because, um. If the paradigm of, um, you know, um, maybe we're skipping. We're going ahead to too much. But we will talk about this later. But. The difference between a, um. The computing of a regular algorithm and a machine learning algorithm is orders of magnitude more complex, right? Yeah. So if all these, uh, procedures are becoming machine learning inferences and or trainings, there will be a stress in, in the network pretty much in the, in terms of computing, not in terms of, of traffic necessarily. Right. -Yeah. -So how do you consider this, uh, um, in the future, do you think there will be, um, differences in costs? Are there already differences in cost? How do you, uh, calculate this, uh, these cost differences when when developing? Yeah. Uh, this is very important when you plan to design, for example, a new neural network, you have to consider, uh, where the neural network runs and, uh, uh, which hardware it uses, uh, as much. Um, my personal opinion is that the future will be, uh, on new hardware. So it means hardware customized for, uh, specific tasks. Uh, for the time being. Uh, we can see, for example, the TPUs, which are, uh, a custom, uh, hardware that runs just for some machine learning and calculation. Um, so, uh, this is useful because using this specific hardware, you can reduce the power cost, uh, instead of use, uh, of using large GPUs, for example, which are more demanding in terms of energy and also cost much in terms of, in terms of money. Uh, so, yeah, I think that, um, in the future we had to move, uh, not only or to improve the algorithm in terms of, uh, uh, logic and software, but also to improve the hardware to have a low impact in terms of energy, energy usage. Yeah. -Thank you. -Can you explain me what are, uh, Kubernetes, Kubernetes and -microservices. -So in the in the logic of the cloud computing, uh, one of the fundamental, uh, uh, software tools, frameworks we can call it in different way, uh, is, uh, Kubernetes. Uh, it is uh, by definition an orchestrator, uh, of containers. So, uh, it is a sort of, uh, framework that handles for you a lot of services. Services are just the some applications that runs on the cloud. Um, the power of this framework is that you can have this orchestrator, uh, is that you have that you have a central dashboard, a central control plan to handle potentially thousands of services, uh, without carrying, uh, um, about the control of each one. So, uh, Kubernetes has some, uh, functionality in its, uh, in its core, uh, for example, it guarantees that each service can run, uh, um, can can stay, uh, forever in the, in the cluster. So if there is for some reason, a failure of the service, um, is um, Kubernetes. Kubernetes itself, uh, takes care to restart the service. And, uh, it, it is should guarantee you that you had not, uh, uh, downtime, uh, for the service. It is this is one of the main, uh, is the selling point of Kubernetes. So the reliance of, of, uh, of the service you put into. Okay. So, um, so pretty much it's it's, uh, correct me if I'm wrong is it is a serverless solution. So you have the, the computing one side and the content, which is the code and the data that is being processed, the input pretty much on another one. Right. So this is a way to scale up or down and also to um, like like you say assure uh, continuity of service. -Pretty much. -Exactly. Yeah. The other selling point for sure is the, the scaling of the service. If you have a service that, uh, as a tone of request, for example, I can think about social media, which can be visited by thousands and millions of, of people in, uh, in a short frame of time, you have the necessity to scale the service in multiple services, which are the same but replicated in multiple time, that guarantee you the continuity of the service. Uh, because a service as a, as a limit. So after a certain amount of request, typically it fails because it cannot handle, uh, infinite amount of requests. Yeah. So when this uh, cloud and microservices, uh, systems or uh, solutions are, um, a good idea and when -are they an overkill? -Yeah. Um, for the at the moment, I see a lot of, uh, inappropriate usage of, uh, microservices. Uh, sometimes, uh, when you have just, uh, a single application, for example, a classical web application, uh, and the usage of this application standard under a certain threshold of, uh, of request, typically 1000 of requests make it in a, in a, um, in a serverless mode. So, uh, put it in, uh, for example, in a Kubernetes cluster is for sure and overkill because, uh, you had to handle too much complexity for, um, uh, that not pay you because it's useless. I mean, if you have the same application deployed in a, in a server, you have the same performances without all the infrastructure, which is for sure a cost that you can avoid. -So. Yeah. -Yeah. Sorry. Yeah. -Please finish. -No. Instead, on the other end, you you had to put, uh, your services into a cloud environment. For example, when you have multiple services, many, many of them. And, and you had to be sure that there is no failure. So if you have a critical service for that, it has to be up and running 24 over seven. So, uh, solution uh, like cloud, uh, hosting is uh, is for sure a good, a good option, um, for example, in a Kubernetes cluster because, uh, it guarantee this, uh, this reliability that, uh, a server sometimes could not. So let's say you're developing -an app, right? -Yeah. And you have this, uh, well, of course, for an app, the destiny, if everything goes well, is to end up on a cloud, right? So how is the process, uh, when designing an app? Well, not the app itself, but the the, um. Um profitability of the of the app. To understand how much will it cost to provide a service uh, to the end client when they are, uh, -triggering an action. -Yeah. Um, in this, I mean, there is a different solution to provide an app, um, to provide an app in general. I mean, uh, I can think of three possible solution, which is the, um, I mean, the, the naive one, which is, uh, hosted in a server and you provide simply by the hosting. Second one is, um, putting, uh, the, the app inside, uh, uh, much more, uh, large infrastructure like Kubernetes cluster, uh, for example. And the third one is using its use a serverless, uh, approach, uh, which is a slight different from, uh, Kubernetes cluster. Yeah. Because the, the philosophy of the serverless is that you had you run just, uh, um, an application, you are just the function, possibly. Um, not, uh, uh, not a service entirely. So the difference is, um, I mean, um, is very small between service and, uh, and, uh, and function and application. A service in general is a little, is a little bit much complex because, uh, it requires some, uh, configuration when you deploy it in a Kubernetes cluster, for example, instead, uh, uh, an application, a pure application is literally just the code. And you can load upload this code in a serverless service and it handles for you the, uh, the run the runtime of this application. So this is the three possible way you can you can host an app, uh, service for uh, for for a customer. Okay. So your, your advice for uh because, you know, with the advent of GPT and all this revolution of, of AI, for example, there are millions of applications that that could be developed of these technologies. So there's a lot of, uh, sandboxing, uh, involved into this, uh, new, new potential, new services. So your, uh, what what would be your advice? Start with your own server and and make it fail and then move to the next one. Uh, uh, I think that a good, uh, good advice is to use a mixed approach. Um, at the moment, there are many different, uh, services that, uh, I mean, uh, website, uh, websites that gives you the, the possibility to host a serverless application. And I recommend to use them for, uh, um, computational demanding task, like, for example, hosting a generative AI model. Um, instead the classical, I mean, uh, web application apart. So the, the frontend and the rest API, uh, could be a good idea, uh, locally or in a, in another service, not serverless, because, I mean, you want that, um, the, the for example, the frontend you want that is always up and running, not, uh, um, stale in idle when you, when you have no request. This is a very interesting point that in serverless, uh, in serverless, uh, field, when you are not, uh, interacting with, uh, with the application, typically it, it goes in idle. So, uh, it was shut down and you had to wait the another request to, uh, start up and restart the process. So I think that this is a good approach. Put the put the computational someone on someone else's computer in serverless and uh, guarantee the the web application by your own. Yeah. Okay, so I'm moving into since you're talking about the generative AI, um, I think it's, uh, it's nice to talk about we talked about the infrastructure. So, um, you know, the computers, other people's computers, they are doing our, our, uh, tasks for us. And now we we we can talk about what they are doing. Exactly. So we as we mentioned before, we were talking about the difference between the classic, like I call it classical, but it's probably not the term the, the, the regular algorithms and this, uh, machine learning algorithms. So why are they more -computing intensive? -Yeah, this is a interesting because, um, some of the classical algorithms, its algorithms are even more compute intensive of of machine learning algorithm. But, uh, for sure, um, modern machine learning algorithms, algorithm like transformers and the I mean, which is the, the the fundamentals, the algorithm algorithm, the on which, uh, generative AI is based, uh, are really stressed in stressful in terms of computational demand. And yeah, the main the main point is, is the fact that they have to to do a lot, a lot, a lot of, uh, uh, computation in terms of multiplication of matrices. This is the, the, the, the task that, um, that an algorithm, uh, a machine learning algorithm to, uh, this is the, the very basic, uh, uh, task that, uh, it has to do to work. And this is the main reason why, uh, it performs so well in the GPUs, which are the perfect hardware to perform matrix multiplication. Um, so, um, uh, the, the, the the fact that the generative algorithm algorithms, uh, requires so much, uh, uh, time to, to train to inference and so on, uh, is due to the fact that they have to recognize, uh, uh, billions of, uh, of features in some input data, um, for, for doing this, it requires, uh, uh. To do, uh, an exponential number of multiplications of the, of the difference input data. And so and for this reason you had to, uh, you need uh, the, the, a lot of computational power. And, and for this reason we have, uh, we see some, uh, uh, some interesting things like, uh, uh, these days we, we, we see, for example, the one of the main, uh, uh, discussed model, which is a sorta this, uh, this great, uh, this disgraced, uh, uh, model that can generate video from text. Okay. But, uh, under the hood, to train this model, uh, you require pretty much an infinite amount of, uh, of computational power. Um, and for how these algorithms are designed, um, the the main difference between, uh, um, I mean, uh, a result or another result, uh, is just related to the training time. More features, more, uh, input data a model, uh, has in input. Uh, better will be the performance. So, uh, I'm pretty much sure that at the moment, the, the main difference between, for example, an open source and a closed source solution is the, the, the possibility to, to run on large cluster or, and at a large a huge amount of -computational time. -This is a perfect segue for another topic, which is, uh, business because of course, uh, if everything is the way that it looks like, it looks like, uh, every company, regardless of their area of influence or, or or sector, uh, is going to have to implement at some level, uh, artificial intelligence in order to stay competitive. Right? Yeah. So. Given that everything you have to do has to go through computational power. My question is. Is it feasible to be competitive or stay competitive, or try to develop an innovative idea if. At the end of the day, regardless of what your idea is, you will need a lot of computing power and hence a lot of money to to develop. What do you think? Is it feasible? Is it? Uh, are we moving into, like some sort of huge monopoly of, uh, libraries or. Um, uh. Applications that cannot be, you know, um. Contrasted with your own applications, because simply you don't have this -infinitive computing power. -This is, I think, one of the m a billion question. I mean, you know, it's very difficult to to answer, uh, because there is no answer, no, a certain answer, because the main point is that we have to to see, um, how the, the researcher, uh, change in the next month and years, for example, uh, there is some something encouraging, in my opinion, uh, that uh, basic the how, uh, these algorithms are changing. I mean, um, at the moment, for sure, the answer is yes, in my opinion. Uh, for the next three, uh, four years, uh, uh, I'm sure that the monopoly is on the large company, because for sure, they have the possibility to run on a large cluster and have a lot of computational power. But what is interesting is that, um, the research is moving faster. And, uh, the there are a lot of new application mathematical tricks, uh, that gives you the possibility to, um, for example, uh, inject something, something in the, in, in, in this large neural networks and, uh, change some behavior. So, uh, if we continue to start in this direction and move in this direction, possibly this algorithm will change and becomes more feasible and more, uh, uh, scalable. Uh, possibly, um, to run also on, uh, cheaper, uh, hardware and in parallel if we, uh, move also the research in the, in new hardware much more affordable and uh, uh, more custom for uh, uh, machine learning tasks. The, the some of these two things probably will end up in a more equity, uh, machine learning, uh, um, distribution and not -just on big companies. -Okay. So, um, so there's hope you're saying there. Yeah, yeah. I'm really. Yeah. I'm trustful on the future -because. -You know, for example, and this is an example that touches me directly because we are we are doing some, some tools to, to create, um, to apply, um, technologies that are because this is something that maybe the audience is don't know. But, uh, once you get inside this, uh, this sector, you realize that what is being marketed, uh, a lot of times, uh, as, uh, some things that are completely magical, etc., are actually research that are open for everybody, and everybody can, can actually, uh, use it and apply it. Of course, not full scale, but you can do a lot of things with the with the papers and the code examples and, uh, and the models that are, uh, open source and available and uh, there's uh, this phenomenon called like, uh, boxing libraries where the you just get an interface that, uh, gives the service that is pretty much free because you can you could run it on your own, but instead, uh, you're being charged for pretty much the, the handshaking. Right? So what I'm noticing is that this, uh, for example, for, uh, creating video, it's a hugely, uh, uh, computationally intensive and, uh, and for example, there's uh, two models right now. One is, uh, you already mentioned it, which is sort of. And which is the closed source, uh, uh, model from OpenAI. And then there's, for example, it's an open, open source counterpart that looks, uh, a lot less performant, but it's available and you can actually use it, which is called, uh, stable Diffusion video and, uh, trying them, you see that it takes a lot of time and a lot of GPUs to, to get results. And, uh, the question is, how is it possible that, uh, services like OpenAI can, uh, give these, uh, uh, um, services for, for so cheap when if you actually have to pay for the computing power, it it -it costs a lot of money. -Yeah. This is interesting. And permit me to, to make some, uh, point on the discussion because, uh, it's different the computational demand between training and the inference. Uh, and this is the selling point, in my opinion, because the training is the hardest part. I mean, it was the, the main part in which an algorithm, uh, should uh, uh, do the, um, this huge amount of computations instead, the inference is much more affordable, and it is the part in which you can make more tricks to reduce the computational times. So, um, for sure, uh, companies like OpenAI optimize the inference of their model. And this gives the possibility to, uh, make them, uh, make the cost of the APIs so, so cheap. So, yeah, I think that that is the, the main point because otherwise, uh, I mean, it's very difficult, uh, try to, to, to figure out how generate a video could take just a few few cents. -I mean, yeah. -Do you think it will cost just a few cents? Or do you think the if and when? Uh, I don't know what's going to happen, but, um, when soda is available for the public, it will be kind of an expensive, uh, service. Ah, I think that probably, as the other, uh, services provided by OpenAI. It will, it will have a very detailed pricing. So in terms of, for example, frame rate, duration, uh, uh, number of tokens, uh, that you can send this or the length of the prompt. For example. Um, but yeah, for a short video, probably with that resolution, I can say that, yeah, some sense could be -a reasonable amount of money. -All right. Right now we're we're paying a lot more for generating. Yeah, but I mean, better resolution and very short. We have a yeah. Yeah, we have a. -Lot more control as well. -So that is a very important point that I, I like very much. That is yeah. These models are fantastic I mean incredible, but I think that. It is important to consider that you have no control on the result, and if you want to change a small detail in the result is not so easy because you have to redo the inference and the result will be different from the previous one. Uh, also on another, uh, kind of detail. So of the, of the results. So yeah, um, I think that this aspect of uncertainty will be the drawback of these -incredible models. -What? Um. Do you think? Uh, let's talk a little bit about sustainability, right? Because, uh, uh, we in, uh, in this, uh, project called Azimuth Media that you are actually watching right now or listening if you're on a, on Spotify, uh, there's, uh, this, uh, the focus of this podcast is talking about, uh, you know, the Sustainable Development Goals of the UN, which is a framework to define what should we do to not, you know, go extinct in the in the next 30, 40 years. So in this sense, this this presents a challenge in terms of, um, uh, of course, uh, computational power means, uh, more electricity, more energy, more, uh, more water to cool servers, etc.. So we are kind of stressing a lot, uh, the, um. The system? Pretty much. And, uh, something just to not be bleak or anything because, you know, the, the, the outcomes, uh, the interesting paradox about these technologies that it could solve very complex issues. Right? So, uh, without going macro scale and going to the into the little, the little person, what can we suggest people to do or not to do, for example, to be more conscious, uh, environmentally, for example, not, uh, you know, uh, doing an inference to open the fridge, -for example. -Yeah. This is a very good point. I mean, uh, a lot of these model, uh, are used for, uh, useless tasks just for fun or just for, uh. For. See how this works. But, uh, due to the fact that they are really demanding in terms of, uh, of power, uh, will be better use them with more consciousness. So, um, I think that more most of the of the task that we, um, we the demand on this models could be done in a very in another sustainable way and, and a lot of services that are arising. For example, uh, one of the main, uh, um. A selling market is the of this model is the generation of images. For example um and generate images is very demanding and uh I, I see on, on internet, uh, an infinite amount of websites and services that generate images. And not only the generation of these images is demanding, but especially the storage of an indication of these images is demanding. So, uh, this is make an impact all the chain as an impact. So not only the generation, but the storage and the retrieval of these, um, of these, uh, media created. So, uh, um, I think that, uh, one possible aspect is, um, as any other tool is use them with a little bit of more, uh, consciousness. So, uh, if you have to generate just 100 of images just to generate them, you can probably generate just one image and not an hundred to see some, uh, uh, little detail, different little details because. Yeah, uh, if you multiply this, uh, this, uh, approach by 1 million of people, it, it becomes really, really impressive in terms of computational demand and power, energy and so on and so forth. So a little bit more consciousness in when you use. Yeah, let's go for it. Um, okay. Wrapping, wrapping it up a little bit. Um, what do you consider are the most uh, let's go to the macro now instead, since you, you are actually working probably on, on projects that have a huge impact on, you know, uh, on people's lives, for example, when, when it's, uh, about healthcare or, um, or for example, the industry. Um, what is the silver lining here? Like, uh, what are the applications and that, you see that can actually improve people's lives and are also able to be -done with this type of technology? -Yeah, I think that this technology are great for, um, an aspect which is the, the fact that it can accelerate the process of doing something. Uh, one of them, that technology I, I found most interesting is the retrieval augmented generation, which is raga for, uh, for the context. And so it is an, a one of the newest technology which merged together the generative aspects of the models and the local data you have and the data you have, and blending them together, um, gives you the possibility to, uh. Make the task that you have to do more easily or more easy. But also you can do it in a, in a fewer time. Uh, for example, let me do an example. You have to, to read the, uh, 20, 30 uh, papers in PDF and you have to summarize them and create some slides and so on. Um, this is quite, um, boring task. And sometimes also, uh, time demanding task. So using a technology like this, you can do it in probably half on the time or even, uh, in a small amount of time. So, uh. Why? Because you can ask to a generative model to help you to elaborate. Uh, these documents, uh, without uh, uh, the the the cost of, uh, uh, train the model, uh, with this, with this document. So it's no longer need to, uh, show at the model, uh, the data. But the data are shown, uh, during the inference, uh, and during the creation, during the generation phase. Um, and this is extremely interesting because you reduce the you have not time to, to spend in training or the other stuff, but you can immediately use the model for your, your own task and on for, uh, tasks that other people had, uh, and think about. So, uh, in general, when you train a model, you focus that that model, uh, will do a specific task with this approach. You can, uh, um, you can use a model for many different, uh, task, many different users. You can generate, uh, answers for biomedical research papers, for agricultural research paper. You can ask to comment an image and extract information from an image. You can extract information from a video, from an audio and collect all this information together to elaborate something new. And this is extremely impactful in my opinion, and useful. And this is the real usage. I, I can see, uh, with this kind of models. Is this kind of technology the one that, uh, projects like, like lung -chain are using? -Yeah, exactly. Lung training is one of the most used framework to build, uh, racks. Uh, but you don't need to use them. You can realize the, uh, rags. Also, by your own, I mean, uh, what is interesting is the logic to build this, uh, this application, uh, which is a very interesting, uh, trick, uh, because you, in a certain sense, inject your data during the generation phase. Uh, the model is, during this generation, retrieve the most interesting, uh, the most relevant information in your data to generate the response. Uh, in this approach is a game changer, in my opinion, in the usage of the generation models, the generative models. What about because you talked about the, um, the, you know, the the inference and the training and, uh, the difficulty of training. Uh, talking also about the evolution of, of, of business is that we're moving from a model that that was used to call the SAS software as a service and to a model that is not well, you know, software as a service was, uh, originally, of course, without the maintenance. But, uh, let's say that you develop an app that solved the problem. You did the code once, and then you just pretty much cashed for anything. -And. -You know, and that's the explanation why people like Jeff Bezos or Mark Zuckerberg became billionaires. Right? But now the training is becoming a manual process, which is which means that you have to continuously, uh, put manual work into, into this, uh, solution. So, uh, are there technologies to that are helping in the, in the, um, you know, um, selecting of, of data, pruning of data and, uh, making this, uh, process, which is a lot very manual and a little bit easier. Yeah, yeah. There are many different tools, uh, for data labeling, for example, that, uh, uh, um, process pre-processing and so on. Uh, but one of the interesting aspects in, in, in preparing data for training is the fact that if your data, uh, as a have a bias or an error or something else, that then the model propagate this, this error in its, uh, in its logic. I mean, so, uh, this will remain in the model forever. And this is very interesting in my opinion, because at the end of the day, the performance is and the efficiency of, of a model, uh, always start with the, um, purity of the initial data, uh, and not also the, not just the purity, but also, uh, I mean, the how much the data are, uh, realistic and representative. Uh, we saw many cases in which, uh, uh, data are uh, uh, some, uh, kind of data and are under represented representative. And for this reason, the response of the model are in. Is always biased and and respond with just one group of data and on the other one. Uh, so we can use tools to improve the data preparation and cleaning and so on. But at the end, I think that, uh, the manual process is, uh, is necessary because, uh, we are the final users of these models. So we have to feed them, uh, with the data that we want to, uh, to see. Also, at the end of the process, uh, um, there is some examples in, in which, for example, you use a model to feed another model. So uh, you can use, uh, GPT to generate text labeled with some labels and use this data as a data set to train another language model. You are. Uh, for example, in this case, you are. What you are doing is, uh, um, give it a. The. At just chat GPT the possibility to introduce some bias in the new in the new model. So when you uh that when you will use this new model, you can encounter some biases and you are not able to, uh, um, make the inverse process. So you see this bias and you can propagate in the future this bias. And in my opinion, this is a problem you have always to be, uh, to have the control of the flow. Uh, also in this, uh, in this complex model. So the and you can have the control, uh, knowing which are the, uh, the, the initial -data. -So yeah, there's a funny example with Gemini, right? Right now, which, uh, they ask, uh, who did more harm if Hitler or Elon Musk. And he's like the answer. So it's a different difficult question. And and so this is an example. Yeah. So so training is a very delicate, delicate, uh, aspect of the, of the process. Pretty much. Yeah. Yeah. And. And, um. It's also kind of dangerous because, you know, when you the expectations of people when, uh, doing an inference, which is nothing else than, uh. Uh, asking for the most probably probable result of a statistic. Pretty much. Right. Uh, it's taken by people like, uh, like an oracle. Right? Like, oh, this this thing has all the answers. -And, uh. -It can help me with my -life or whatever. -So. So it is a bit dangerous to the, the to forget that we, we as people are actually telling the the machine what to say, uh, in some way. Right? Yeah. It's dangerous also because, uh, uh, I think that is G is dangerous. Uh uh, taking as true it every, uh, news that you read that you read on, uh, on internet. So imagine that this, uh, fake news, for example, is used to feed the model and then after a Bayesian process. So a statistical process, uh, this, uh, news is used to generate the answer of, uh, of the, the answer at your question. And yeah. So you, you create a chain of, uh, of, -uh. -Uh. Of things that are not real. But if you trust without any doubt, uh, yeah. -You know, you have a problem. -Oh my God. So what's the perfect theme to, um. To create the data set. Like what people should be hands on into the data set because all only that the scientists or maybe philosophers or, -uh, what do you think? -Yeah, that data preparation, in my opinion, is one of the most the most relevant aspects when you are approaching to create a machine learning solution. Uh, for sure is not just upon a data scientist, uh, because a data scientist has not, uh, all the knowledge of the domain knowledge of each, uh, aspects clearly. So, for example, if you have to create a dataset for a bio biomedical, uh, task, for sure, you need, uh, someone that is an a domain expert that can help you to understand which kind of data are good, which kind of data, uh, should be labeled in a certain manner, and so on, because, uh, as a data scientist, you adjust the statistical tools to evaluate the data set, but without the domain, uh, the domain knowledge, you cannot you cannot identify some aspects of the data that are further, uh, that are beyond the, the statistical, the statistical analysis. I mean, so you need for sure, uh, the, the domain expert and why not? Sometimes I feel as if, uh, could be a very good, uh, help. It's interesting because this, this paradox is probably the most, uh, important one when talking about, uh, um, job replacements. Right? Like, uh, there's a lot of, uh, of talking about, uh, well, it will replace our jobs, which is a lot of times there's a there's a company of truth on it, uh, because it looks like for people that it's not in the domain, that the on the first impression that the, the lamps can perform tasks, uh, with a lot of competence. And also the data shows it. Right. But also there's the. So. Uh, and correct me if I'm wrong, but this is an opinion and probably they will come a storm of, you know, layoffs and, uh, and people, uh, losing their jobs because it looks like the, the, the GPT or the LM or the tool can, can do what the person is doing. But then, uh, you have this paradox that you actually need a domain expert to validate the results of the, of the tool. So, uh, they will be they will be more, even a bigger need of humans to be on the loop. Yeah. Of these -technologies. Right. -And also another aspect is relevant, in my opinion, that the fact that, um, the, the products generated by this kind of models, uh, is not perfect. And sometimes you want to change some small aspects or some tune. I mean, the response and uh, doing this sometimes is very painful, uh, just by prompting and trying to optimize the, the way you ask something to this model. And sometimes I find more, uh, uh, time saving do it by my own not asking to to the model to, to generate, uh, a short email. But probably sometimes if you write by your own is better. And and this is extremely relevant. For example, for the code, there are many different models that generate code for you by asking generate a function that doing blah blah, blah, blah. And that is very, very dangerous. And trust the the code that a model generate for you in my opinion, is uh, is wrong because no one has the control on that snippet of code and could be potentially a risk for all the code base. And uh, uh, you need always to, to, to read it and document it, because otherwise it is just, uh, I mean, it is useless if anyone knows what this, this function, this class, this module do in your code base. So in my opinion, yeah, the job replacement is uh, is a wrong, uh, uh, consequence of this, uh, of the advent of this model. I mean, this model should be used as a tool, as any other tool to to doing the job, not substitute people that doing the job. Exactly. Let's do a little bit of futurism to to finish the the meeting. Um, right now open a Sam Altman is, uh, has announced that he wants to create, uh, uh, start with the building of, uh, GPUs in, uh, business. And, uh, the Nvidia CEO is trying to get into the LM, uh, business. What do you think that will happen? And do you think. -Who will win and. -Who has better chances? -Yeah. I mean, uh. -Nvidia is more, uh, historical. Uh, I mean, uh, knowledge about GPUs and for sure they are, uh, the, the, the right choice if I have to bet between a OpenAI, uh, Nvidia. Uh, but in my art, I hope that, um, someone else will come and with another idea which is not new GPU's, but with new hardware, much more, uh, friendly for for the nature of, I mean, for the footprint -and. -We will. See. Uh uh, more uh uh uh, hardware for machine learning task more affordable, um, with low computer power, need of, uh, of energy. Power. And this is my my my hope, I mean. Do you think that, uh, non AI development, uh, will will continue to strive? Will it be guided by AI or or not? Are we going to depend of inferences for the rest of our lives? And the point is that the is very comfortable use using the AI. So, uh, I think yeah, we depend in future we will depend even more. Or on the AI. Uh, but it is not bad. Uh, in my in my view. I mean, if we use it just, uh, as a tool, as any other thing that we use, I mean, it has no problem. The problem is, uh, as we said before, we trust the, uh, without any doubt on, uh, uh, the AI says or right or generate. That is the problem. Great. Okay, Marco, thank you very much for this. Do you have something else to to, to add? Um, or something that I'm starting to do is, uh, if you could bring someone here who who you bring and, uh, what would you like to talk about? -Uh, yeah. -I think that a good boy could be Sam Altman and and ask you, please make more affordable, uh, uh, hardware. Not another kind of GPUs. -Okay, great. -Do you want to have something that, uh, maybe you want to specify something that we missed? No, I think that we have covered mostly all the topics that we have. Uh. Dutch, so I -think we are done. -Thank you very much again for being here. And thank thank you all for listening. You can find uh, the this podcast in video format on uh, video Dot Asimov, dot media or uh, you can find it in, uh, Spotify if you look for Asimov perspectives. Thank you once again, and I hope to see you soon with another episode of perspectives. Bye bye. Bye.