WEBVTT 00:00:01.760 --> 00:00:35.760 Welcome to the Practical AI Podcast, where we break down the real world applications of artificial intelligence and how it's shaping the way we live, work, and create. Our goal is to help make AI technology practical, productive, and accessible to everyone. Whether you're a developer, business leader, or just curious about the tech behind the buzz, you're in the right place. Be sure to connect with us on LinkedIn, X, or Blue Sky to stay up to date with episode drops, behind the scenes content, and AI insights. You can learn more at practicalai.fm. 00:00:35.760 --> 00:00:37.440 Now onto the show. 00:00:41.280 --> 00:01:04.790 Welcome to another episode of the Practical AI Podcast. I am Chris Benson, your cohost. Daniel Whitenack is out on paternity leave right now, so he won't be joining us. We've got a good conversation coming up today, I know, because with us we have Lorenzo Sani, if I got your name correct there. And Lorenzo is a research scientist at Flower Labs. 00:01:04.790 --> 00:01:09.910 He is also finishing up his PhD at Cambridge in computer science. Welcome to the show. 00:01:09.910 --> 00:01:12.470 Hello, Chris. Hello, everybody. Thanks for having me today. 00:01:12.715 --> 00:01:25.035 Hey, absolutely. You have a really interesting set of research that you're working on. We're delighted to have you on the show today. I'm looking forward to talking about it. We got a lot of questions about it. 00:01:26.140 --> 00:01:50.615 But I wanted to actually start off with the fact that a little bit about your background. I know you came from the physics world and moved over, and we've seen that a fair amount in research scientists and AI. Daniel, the cohost, who's not here today, is in the same boat. I'm just kinda curious if you could give us a quick, you know, kinda how you arrived at what you're doing now having started on the physics side of things. 00:01:50.615 --> 00:02:02.490 Yeah, of course. You're right. I have a physics background. I've been studying physics for five years after my high school. I was in Bologna, amazing place, amazing people, and very good courses as well. 00:02:03.290 --> 00:02:33.445 And especially during my last year of Master's, I was looking into topics related to machine learning, and the group there was working on medical applications. We needed a way to kind of preserve privacy and looking into those privacy sensitive scenarios. And back then, Federated Learning was quite new, so I started looking into it. It was very interesting. I loved the motivations behind it and the way to carry along the privacy preserving component to it. 00:02:34.085 --> 00:02:59.260 And and there's a result. I was one of the very first user of the Flowr framework, and this brought me to meet all the Flowr people as well as the research group in Cambridge that was developing it. And things got along. Luckily enough, the professor, Nick Lane, here was looking for PhD students. I was very excited to to join this line of research and whatever was around it. 00:03:00.255 --> 00:03:20.970 And and I got lucky. I got a seat. I I started working with with the team. I started developing my own applications and methods. And over the course of the last few years, I got more familiar with the latest advancements of AI, I would say, like these LMs and everything related to it. 00:03:20.970 --> 00:03:29.610 Now I can say I'm very familiar with everything that goes from collecting the data down to serving the models and using them day to day. 00:03:29.745 --> 00:03:49.510 Can you talk a little bit about, as you dived into those areas, what were some of the specifics that really drew you in? What were some of the first problems that you addressed that kind of got you started in the area, and and what was... What what drove the passion regarding that? 00:03:49.510 --> 00:04:18.910 Yeah. It's a very good question. I've thought a lot about this after a couple of years that I took this path because, yeah, I wanted just to, you know, go back and try understand better my choices and why did I take them. I think the main reason was related to the complexity behind this technology. Because back when I decided to do physics, complexity was the most fascinating thing about studying physics and getting to know the world. 00:04:18.910 --> 00:04:45.310 And for this technology, it's mostly about we have a lot of data. We want to make sure that we have good machine learning technologies to learn from this data and get useful stuff done. But this data is very complex, and the training procedures are very complex. So these things excite me a lot. In the Federated Learning kind of area, the complexity goes into the distribution itself because you have all this siloed data. 00:04:45.470 --> 00:05:16.100 It's very scattered, very terogenous, very difficult to deal with in the standard way. But Federated Learning provides a good way to train, to train on this and get the intelligence that you want. It's similar to the direction I took when I started my PhD for the LLM pre training stuff that I've done. There, the question was, can we use the Federated Learning Technologies to train these language models? It was, like, something that nobody did before. 00:05:16.100 --> 00:05:43.630 There was a few works from DeepMind that were very interesting, but we wanted to take a much deeper approach. And at the beginning, was challenging, but then, like, we realized, okay, we can do this stuff. We spent a lot of months in developing the tools, the basic tools to do it, and now we have plenty of papers showing that we dig deeper into the technology, into the science of these things, and we can get good results out of it. 00:05:43.630 --> 00:06:06.970 So for those who aren't familiar with it, can you talk a bit about what Federated Learning is, what the benefits of Federated Learning are, and specifically, as you talked about developing those tools as you were doing that, what drove you into that? Why was that the thing, and what was the benefit that you were able to achieve out of that? 00:06:06.970 --> 00:06:16.810 Yeah. Absolutely. So Federated Learning is a semi... Well, for the history of the world, it's a semi new technology. Now we probably, in the eye, it's pretty old. 00:06:18.615 --> 00:06:48.405 But the idea behind Federated Learning is the following. So data is collected in many different places at once. Right? We collect data from our edge devices like phones, or laptops, and so on and so forth. And all this data is usually then centralized into the data centers that are doing the machine learning pipelines to train models and to learn from them, and to provide useful insights, to construct applications, and so on and so forth. 00:06:49.205 --> 00:07:12.160 Now, most of this data is openly available. We know that you can scrape the web and gather as much as you can if intellectual property is preserved and so on and so forth. But still, there's a lot of data you can source pretty freely. But there's a massive chunk of data that is completely private, is very difficult to source, right? And it's private for a good reason. 00:07:12.285 --> 00:07:58.900 Sometimes it's your medical records, and you don't want everybody to look at it. Or some very personal information that you don't want to disclose. But this may be useful for training machine learning models, and of course we have regulations that are protecting our privacy for doing, from from, like, this data collection and whatever comes next. And Federated Learning comes into place in order to try to satisfy the regulations, protect your privacy while still being able to learn from your data and construct useful application for the users, for the people who are actually generating this data. And in short, Federated Learning tries to move the training pipeline, so all the computations that you need to train your machine learning model to where the data is collected. 00:07:58.900 --> 00:08:14.515 Of course you need to compute there, not just the devices that are collecting the data, but iterating these training procedures. Eventually you will be able to obtain a good machine learning model that can learn from the, that has learned from your data and provides useful insights. 00:08:14.515 --> 00:08:51.380 Could you talk a little bit about, there's a specific point there that I'm really interested in, that is if you're applying Federated Learning to private data, and to your point, there are aspects that you don't want to reveal necessarily to the public, but you want the benefit of the learnings that the data could offer, How does training at that source of the data keep it private? How do you know that the private aspects of that data are not taken into the model themselves in a way? How do you how do you achieve that balance? 00:08:51.460 --> 00:09:18.760 This is a very good question, and technically speaking, it's still kind of open for for research. There are bounds there. There's there's a lot of research that provides statistical bounds and guarantees of some sort, depending on the setting. The first ingredient for protecting your privacy is the fact that your data never moves from where it's collected. So no one can intercept traffic or actually look at your data. 00:09:18.760 --> 00:09:41.185 No one can open a tab on Chrome and just look at your medical records and stuff. So this is a very strong aspect. It's something that is fixed and it's guaranteed at all times. The question becomes more when you've trained the machine learning model, does it have, has it memorized something about yourself that you didn't want it to memorize, right? And that's where privacy can break. 00:09:41.665 --> 00:10:09.335 And there's techniques, a lot of techniques to train these machine learning models in order to prevent this from happening. Techniques such differential privacy and many other things are providing, these are providing like statistical guarantees that these models are safe, secure, and they are not breaking your privacy. But usually these depend on a case by case. Because in some settings you can apply these things very easily. In other settings it's very challenging. 00:10:09.780 --> 00:10:31.875 Other times, you just cannot apply them. There's plenty of industry who's trying to apply Federated Learning to their use cases, and sometimes they abandon this opportunity because they realize that it's not convenient to them. Other times this works pretty well, and they use it end to end from when they start collecting the data on the edge device, when they produce the final model that does the application. 00:10:32.355 --> 00:11:25.155 It raises an interesting question from my standpoint in that as you're thinking about industry building the models, you have these big frontier models that are private and closed, and then there's the world of open source, open weight models out there, Does the nature of those models being closed or open or anything affect how we should see the application of Federated Learning maybe from a trust standpoint or a utility standpoint. It makes me wonder if you're one of these giant global companies that's producing a large closed model versus some of the open models that are available, should I feel safer with one or another? To your way of thinking, how do you evaluate that personally? 00:11:25.155 --> 00:12:01.630 Yeah, this is a very tricky question because we know very little, right? Like we don't know how these closed source models are constructed. We know a bit more about the open source ones, but not 100% of it. I tend to be very optimistic on this side of things. Like I hope and I kind of trust everybody to do their job and to try to enhance the models and improve them in order to maintain privacy, to maintain the data they can collect and they use for training private, and not to leak anything and to be safe, and all of this. 00:12:01.950 --> 00:12:48.635 Of course, there's techniques to probe models, even though they are closed source. Like, some techniques do actually work by just, you know, doing API calls and just probing the model directly, get the outputs, and see if they are actually containing information that shouldn't be there. But with open source models, it's much easier because the amount of techniques that you can use is just many, many more. I would say that if I had to develop a privacy heavy application, I would prefer to use an open source model as a starting point, especially because I have more control over it. And in case there's something broken out of this open source model, I know there's a way to fix it, and I can pull it off. 00:12:48.635 --> 00:12:56.100 Otherwise, for closed source models, it's very difficult to say, Okay, 100%, if there's something broken, I'm gonna be able to fix it. 00:12:57.860 --> 00:13:32.670 Can you think back to a time when you wanted to share your expertise or maybe an open source project or your resume online in a very compelling way that looks professional. It can be super overwhelming to get to that point. I know it has been for me over time with my personal website or things that I've put up online for Practical AI. I want those things to match what I have in my mind. And that's why I love what Hostinger is doing with their AI builder, which lets you get to a working version of your dream website fast. 00:13:32.670 --> 00:14:21.730 It allows you to feel confident preparing for your next job interview or hearing from prospective clients of your consulting business. Whatever that is that you're trying to share online, Hostinger's AI builder can help you get there and share that. And now through December 14, they're running their Black Friday and Cyber Monday pricing where you can get a premium plan for just $1.99 per month. Go to hostinger.com/practical and use code Practical for an additional 10% off Black Friday and Cyber Monday pricing. That's hostinger.com/practical and use code Practical for an additional 10% off Black Friday and Cyber Monday pricing. 00:14:23.570 --> 00:14:52.260 Alright. Well, in the first segment, we were talking a bit about kind of the background on Federated Learning and some of the issues, privacy and stuff around it. Could you, to turn a corner a little bit, could you talk a little bit about how you're applying Federated Learning, some of the use cases, maybe how Agentic fit into the framework, and and some of the the use cases or applications that you've been diving into. 00:14:52.260 --> 00:15:11.765 Absolutely. So in in my first experience, research experience during my Ph. D, we started to develop Federated Learning Technologies for training language models. Back then it was pretty challenging and fairly new. Basically, the idea behind it was that we were able to source much more data than what was available. 00:15:12.060 --> 00:15:52.600 Right? So the idea there is I'm not using Federated Learning just to preserve privacy, but also to be able to tap into new data sources that could be very valuable. So in the context of language models, right now, what I've seen applied and what I've applied myself is mostly related to this kind of outputs. Like, I really want to use the Federated Learning Technologies so that I could use this massive amount of data that otherwise it's not available. And hopefully, I'd be able to construct much better models, or models that can reflect much better what users produce in terms of data. 00:15:53.635 --> 00:16:57.695 When it comes to the most recent kind of advancement for these LLMs and AI technologies, yeah, we've seen a lot of use in Agentic systems. These Agentic systems are much more complex than the single model application, because you have interactions, you have multiple users, you have interactions with external services, and you have these kind of trajectories. It's probably a word that everybody has heard recently more and more. And there, the privacy component and the Federated Learning component comes into place because all these trajectories do not just contain outputs from the model, but also the prompts that the user type into the chatbot or the applications they're trying to use. Or maybe some third component, third party component, like MCP server, rack systems, all these things that are interacting with the language model in order to provide it with more information regarding the prompt that you just passed. 00:16:58.415 --> 00:17:25.515 And all these trajectories are super useful to improve these AI systems that are relying on agents. Why they're useful? Because trajectories may get messy. They are very long. Models can make mistakes, and these mistakes may come from just, I don't know, a word that is misspelled, or a word that is used twice to refer to different things, or a tool that was a little bit broken because the implementation was not that good. 00:17:26.715 --> 00:18:00.780 And whenever you have this kind of trajectory, this kind of usage, you want your models to be better at dealing with this, and they can be better if you collect enough data to train them in these adjunctive systems. And this is what Federated Learning can provide, a way to train on this distributed data that is very private because it contains your prompts. It contains what you like about the work you do, or about the conversation you like to have with LGBT or Claude, right? All these things are very private. Sometimes, sometimes they're not. 00:18:00.780 --> 00:18:17.215 You're maybe just asking about, oh, tell me the news, what's today's news? But other times they are. Sometimes I ask my agents about, you know what? I've been doing this thing and my knee now is not is not feeling good. Should I be contacting the doctor or not? 00:18:17.455 --> 00:18:22.495 These sort of things. And you don't want this data to be available to anybody else, right? 00:18:22.495 --> 00:18:23.135 Sure. 00:18:23.375 --> 00:18:35.100 And if you want to improve your agentic systems, the tarry line on this kind of trajectory is Federated Learning is a natural way to train in a distribution these AI systems. 00:18:35.100 --> 00:19:16.045 Could you talk a little bit about, like, We've dived into some of the complexity of introducing Agentics and all those trajectories, I think one of the things I'm trying to grasp right now is it's e... When what we were talking about just the model before the break, it was easier to grasp the private and stuff like that. As you're talking about introducing agents at some scale, and they are producing these trajectories, and presumably you have many parallel trajectories at the same time. What does that mean? How do you put that together with Federated Learning in such a way? 00:19:16.045 --> 00:19:32.900 What does that mean in real life? How do you see that? Because it seems like it gets quite complicated quite quickly. Can you bring that together and lay out what that seems to you? Because I'm trying to follow you down different paths, and trying to figure out how they fit together. 00:19:33.155 --> 00:20:11.785 Yeah, so things get very messy very quickly, and you're right on that, especially because agents, what's the difference between a single model and an agent is that the agents can take actions, right? Like, that's the definition of an agent, like the simplest definitions that you can get. Taking actions means that it can read things, it can scrape things, it can modify things, it can remove things. And every of these steps relies on the fact that these agents execute in an environment. And this environment contains part of yourself. 00:20:11.945 --> 00:20:42.045 It contains, I don't know, information about your laptop, your PHY system. It may contain information about your camera, your images, what you're doing, and also your props. All these things compound. So whenever you have a long running trajectory, the amount of information about yourself that retain, that you may think they are private, increases more and more as the agent goes along, right? Because it can take more in, understand more about yourself to produce a final output. 00:20:42.230 --> 00:21:38.020 All these things are basically constructed as you have the initial prompt, you let the model do its thing, then maybe you have two calls, you have calling sub agents, so something that branches out. You can think about like a line that branches out and gather information, you can imagine that I'm getting down from a little database that could be your file system or some webpage or I don't know, your medical records or something like that, And this informs the model, constructing another part of the context it's operating with. And then as you go along, this information get always compacted into the main line of work to get you to the final outcome. So this is kind of the way people are usually thinking about these long running trajectories. So you start with a prompt, then at some point you may have branches out more and more, it's like a tree, and then things get back into the main line when you get to the final answer. 00:21:38.020 --> 00:21:46.260 Gotcha. That helps a lot. I appreciate the context on that. Can you talk a little bit about some of the things that you've done? Any interesting projects? 00:21:46.435 --> 00:21:55.715 Any fun things that you've done with this? Things that have caught your attention? I know before the show you relayed something, and I was hoping you might share that as well. 00:21:55.715 --> 00:22:07.260 Absolutely. I was looking forward to this question. So, I mean, we are we're just after Halloween. Right? Halloween was on the thirty first. 00:22:08.140 --> 00:22:16.615 And I decided to give a party in my house with my flatmates. Hopefully, the landlady doesn't know about this. But anyways 00:22:17.095 --> 00:22:22.615 She... It was a small It on if she watches Practical AI. We hope she does, but, you know, 00:22:22.615 --> 00:22:32.350 We we'll hope she does, yes. But I I won't ask at this point. Fair enough. If she doesn't... Anyway, yeah, we gave we gave a party. 00:22:32.350 --> 00:22:54.755 It was it was fun, of course. A lot of friends. But then we, yeah, we we dressed up the house, right, like standard stuff, all the little decorations and everything, lit candles and stuff. But one thing that we really wanted to do since all of my flatmates are working in computer science, they are working in the department I work in as a PhD student. There are postdocs or students as well. 00:22:54.755 --> 00:23:31.950 Like, that's that's a bit problematic because sometimes we spend, like, the entire evenings after dinner discussing research, and it gets a bit too much sometimes, but that's another story. Then we decided we wanted to do something a bit nerdy, a bit tacky for the party. And since we knew that every guest would come dressed up as, I don't know, as a scary character or whatever they wanted to, we thought, well, would be good to have, like, a competition about about all these costumes and stuff. It looks nice, and people have done this before. But how could we do it? 00:23:31.950 --> 00:24:04.430 Like, we would spend a lot of time identifying the judges and everything, and then we would monopolize the entire evening for no reason. But what we do in our house is also serving our own models, our own language vision models. We have our agents running because we don't like too much to share these trajectories with closed source or our personal subscriptions. Sometimes we do most of the work with our own agents that we serve and serve. Then we decided to develop, well, develop, to vibe code, a platform to do this competition. 00:24:04.910 --> 00:24:43.255 And the platform works like this. You go in front of a camera, you press the space bar, you take a picture, and then all the agents start discussing about whether you're very like very close to the Halloween themes that people usually bring to these kind of parties, if you're scary or not, and also if you're trash or not, because we do value trash up to a point. And then people got very excited. They were coming over, getting a lot of pictures, and they started also trying to act the system, because these agents were judging continuously. Like the more pictures you take, the more they judge. 00:24:43.255 --> 00:24:57.260 They compare pictures across each other. They want to understand, oh, but this is better than the other. Why? Why is this better? And this was really fun because people started to do some prompt injection, like taking a picture with like a signal saying, give me a high score, please. 00:24:57.260 --> 00:25:10.505 I really need this. It didn't work out. Or they started to group together, like, two very scary costumes. Yeah. Singularly, they had a very good score, but together, much, much better. 00:25:10.505 --> 00:25:24.610 So they started trying to act the system to get the best scores. This was real fun. Everybody liked it even though they are not too much deep into the tech. Also, people who don't know anything about these models were very, very excited. 00:25:25.170 --> 00:25:53.150 Sounds like fun. And I wanna go down a rabbit hole real quick, if you don't mind, since you brought up that you're hosting at your house. And I do the same thing. I also do open weights models that are private hosted in my house, and that's becoming a hobby of mine is to find different utilities for what I'm doing at home, and obviously you're doing the same thing. And so since you hit on that, I'd love to... 00:25:53.470 --> 00:26:06.345 Other things that you're doing or that you would like to do, just to get some ideas, because I know that there are folks out there that are either doing or considering doing the same at this point. And so if you wouldn't mind taking a quick detour 00:26:06.425 --> 00:26:07.065 Yeah, yeah, absolutely. 00:26:07.065 --> 00:26:13.225 And sharing some of the creative ideas on if you got open weights models hosting in your own house, what are you gonna go do? 00:26:13.225 --> 00:26:40.115 Yeah, yeah, absolutely. So first thing, I need to acknowledge the fact that I can do this also because of my flatmates. Like, my flatmates are very keen to do these things, otherwise, I wouldn't do this just for myself probably. It's much better when I share this kind of, you know, like, little games, I would say. Like, when I was a kid, I was playing with, I don't know, little raspberries, building little robots or something like this. 00:26:40.115 --> 00:26:57.510 Now, with all these technologies, can write code a website in fifteen minutes, right? I can do all these applications real quick. And when you back them with AI agent, what you can do is also take... Allow them to take actions. So some projects that we're planning to do are... 00:26:58.470 --> 00:27:26.340 Some some of them are related to, like, house management. For example, reminders about the rubbish bins or reminders about the eating. Sometimes we leave the eating gone overnight, and we wake up at 4AM sweating in our bedsheets. That's a bit problematic. But also, another interesting thing that we're planning to do, but this is not just there yet, would be enable a live recording for all our evenings. 00:27:26.340 --> 00:27:47.915 Because as I said earlier, we end up discussing about research and cool ideas over the evening. We have a little whiteboard in the kitchen. This is a bad choice if you want to, you know, like kind of diversify your time over the day, but eventually we like it. But one thing that we realize is that we have a lot of super interesting conversation. None of these get trapped. 00:27:48.380 --> 00:28:04.315 Sometimes we go back to them and we are like, but what did I say about this? There was an interesting bit that we should dig deeper into that we forgot completely. Right? Yeah. So a live recording keeps track of things and maybe ties into, I don't know, the latest research. 00:28:04.795 --> 00:28:47.650 Another project that my flatmate proposed this morning was related to the daily news about research papers. Nowadays, like X, Twitter, it's completely blowing up. Like every morning you wake up, there's five new papers out that are worth reading, and it's very difficult to keep the pace. So what we were thinking about is allowing agents to scrape all these new outcomes, all these new pieces of research, and going out to the kitchen in the morning, making a coffee, and get, you know, an audio play that is basically explaining, look, does this thing interesting because you guys work on this particular topic? It's worth reading, or something like that, like the daily news briefing about research. 00:28:47.650 --> 00:28:50.610 That would be one of the next projects, yeah. 00:28:50.610 --> 00:29:24.355 Oh, that's good, I appreciate that. Yeah, from my standpoint, the house management kind of things, the reminders, the recording that you know is private instead of going into one of these closed source, love all those, and then I'm an animal person. We have a bunch of dogs, and I live in a place with a lot of wildlife around us, and so I'm starting to, with some of the cameras that we have around outside, to prompt us and say, There's a raccoon climbing on the tree outside the window. You know, things like that. So little fun things. 00:29:24.755 --> 00:29:36.860 And I just appreciate you diving into that for a moment because so often we only focus on the work aspects of this, and these are capabilities that can affect your whole life in so many ways. 00:29:36.860 --> 00:29:37.660 Yeah. Yeah. 00:29:37.660 --> 00:29:53.115 Great little rabbit hole. Thanks for going there. Every AI expert seems to recommend a different tool. When I start a project, which one do I use? I'm Chris Benson, and this message is sponsored by Perplexity Computer. 00:29:53.595 --> 00:30:33.360 Computer breaks projects into tasks, automatically selects models, and coordinates agents I use Computer to research the systems that enable autonomous uncrewed platforms to work together as a swarm toward a shared goal. Its research insights help enhance my own technical creativity. For a limited time only, our listeners can try Perplexity Computer for free. Head to pplx.aipracticalai. That's pplx.ai/practicalai to try computer today. 00:30:33.680 --> 00:31:04.940 Most AI tools are still answering questions while Perplexity Computer is completing full projects. So Lorenzo, that was a good dive down into some of the personal aspects. As we come back in and get back to the work world, if you will, and you have folks out there listening. They've been learning about Federated Learning from you. They've been learning about Agentic's and their integration. 00:31:05.180 --> 00:31:34.910 Can you talk a little bit about how people should start thinking about Federated Learning and Agentics to pull this together in different contexts? So they're gonna be running some sort of business or organization out there. We don't necessarily know what it is, but do you have any way of framing kind of how to get started and the way you think about problem solving so that you can get productive quickly? Do you have any guidance around that? 00:31:34.910 --> 00:32:07.605 Yeah, Yeah, this is a very interesting question. I've recently got exposed much more to this kind of area because when I was doing only research for my PhD, we do have some interaction with industry, but it's not that deep. Now working with Flower Labs, there's a lot more going on and a lot more discussions and things to to learn from. And one thing that stand out to me was... That stood out to me was that now these models, especially the most capable ones, those that can do things, are getting smaller and smaller. 00:32:07.845 --> 00:32:22.420 Easier and easier to deploy, right? The amount of usage that this model have, these models have, is is even bigger and bigger, right? So the scale of usage is increasing, the scale of the models is decreasing. The costs are decreasing as well. Mhmm. 00:32:22.420 --> 00:32:43.740 Right? But one more thing that we suddenly realize is that in order to get these models better at doing something, and doing something is what you decide. Right? Like, I decide whether they're better doing what, right, because I use them. Or companies, enterprises, they decide what models are useful for because they use it. 00:32:44.700 --> 00:33:03.735 We realize that this requires deeper analysis. It requires collecting data. It requires identifying failure modes. And this cannot be done very easily if you don't control all these trajectories that we discussed earlier about. Right? 00:33:03.815 --> 00:33:43.180 Sure. Because that's the only way to really inspect how the model breaks, or where the model can be improved on. So one thing we see pretty frequently nowadays is that people start deploying these models, they start choosing them for a specific use case, but then they have difficulties in setting up applications or improving them or building on top of these things. But whenever they have control over the data and they can analyze what happened in the past, and what are the things that could break, or did break. There are many ways to fix that. 00:33:43.580 --> 00:34:05.200 And fixing that may require some additional training, which is sometimes expensive. Other times it's just a matter of the application side. Maybe you didn't think very carefully about that particular step, or that particular external tool that the model was using. But the only way to know is by looking at what happened. Like, going back, looking at the trajectory, and try to understand it properly. 00:34:05.280 --> 00:34:48.445 So this is why I think we will see an increasing effort, especially for big enterprises who, like, are very cautious about their data, their internal flows, because it's their actual IP, right? Like sometimes it's not just your product, it's about like what you do to make your product work, or what should make your product better. We'll see a lot of these enterprises looking into these trajectories and try to improve models by themselves without relying on external providers that often disclose, that often don't disclose the full traces or the full steps that the model took to get to a particular answer. Right? Mhmm. 00:34:49.325 --> 00:35:22.545 And on the other side, there's the interpretability kind of path, right? Like sometimes we, no, now we are trying to use models in very sensitive applications. I've seen models doing surgery for real, like commanding robots to do surgery, and this was super cool to see. But then I thought, do we know why they did that particular step? Or can we make sure that if they did that particular step, the model knew why this was the best choice? 00:35:23.130 --> 00:35:59.950 And it did evaluate all the potential things to do beforehand. Because there's plenty of sensitive applications where you would really like to have control over the motivations or the reasons behind specific workflows. And for this, there's the entire interpretability kind of area or path that is getting more and more important, and more and more interesting these days. And still, it requires a lot of data, it requires a lot of analysis, a lot of deeper analysis, and much more control over the model compared to what you can get out of a standard subscription from any big company product. 00:35:59.950 --> 00:37:05.515 Right. You raised some really interesting points I wanna ask you about there, and that is you talk about that models are getting more capable, but they're also getting both cheaper and easier to deploy as they get smaller, and we're also seeing more models out there toward a given application or a use case that one might have, instead of one model driving it, it may have multiple models, and they can be both cloud based, and they can be embodied in robotics or drones or other things that we're talking about that affect the physical world, and so it's kind of getting complicated. The architecture, not just the computing architecture, but the physical architecture of how these relate. Any thoughts about, with Agentics mixed in and the actions that that implies, any thoughts about how Federated Learning and Agentics and all come together when we are talking about real world? And people are, to your point, they are going to go into hospitals. 00:37:05.515 --> 00:37:49.200 They are going to have surgeries, and there may be a human surgeon present, but it may be that the actual surgery is conducted by a robot that's being driven by a model with a level of precision that the human probably can't safely do, which touches close to home. I have a stepdaughter who is a surgeon, I know her husband is involved in that. I'm very curious how you think about how all this comes together. As we're starting to wind up, you wanna lay out where things are at, and what you're thinking about it in terms of, at least today, for the moment, where we're at now, and what people can do? 00:37:49.760 --> 00:39:03.495 Yeah, so right now, we know that we can collect these traces, as I said, and these can be useful to improve applications, or create new ones as well, because whenever you put a model out of its comfort zone, so something that it's never tried before, it's gonna make a lot of mistakes, and you can only correct them if you know where this mistake have happened, right? So one thing where Federated Learning, or any collaborative learning technology, because Federated Learning is one of them, could help, is in making sure that we can pull sufficient episodes, sufficient use cases, sufficient situations, well, not environments, but maybe actions in environments that would be useful to improve the broader capabilities of these models. For example, in the surgeon kind of area, you can think about, okay, a surgeon does a lot of actions to perform a particular surgery technique, but everybody's different, right? Maybe you have one person that has particular features, another person has different features, and all the veins, they're different. So it's never gonna be the same at all times, right? 00:39:04.375 --> 00:39:47.230 And of course, you cannot scale up the amount of trials that you do unless you have a very good way to simulate this world, but we're not just there yet. So currently, the only way to increase the amount of improvements, like to increase pace of improvement for this model, is getting more and more data, and make sure that you can train on this data in order to improve the model. And one clear way to get more and more data is sourcing it from many different places at once. Many different hospitals, or many different silos, many different areas. And collaborative learning technologies, Federated Learning Technologies, can help a lot with that, and they can help right now. 00:39:47.230 --> 00:40:01.665 Like, for the kind of collection pipelines that we have currently, for the current training methods that we have currently, these collaborative learning techniques can help out of the box. 00:40:01.665 --> 00:40:21.170 Fantastic. I'll have to, and that gives me a lot to think about. As we come to a close here, I would like to finish by asking you to think a little bit creatively about the future. We like to finish up episodes with 00:40:21.490 --> 00:40:21.890 a little bit 00:40:21.890 --> 00:40:56.565 of insight into what your dreams are about, what you're thinking about the future. We often will ask, know, it's kinda like the shower before you go into bed, you're laying down, or you've had a glass of wine, or whatever you do to just chill out and relax, and as your mind flows, like where do you think things are going? What are you hoping to see, and what would you like to be part of in the future? Wax a little poetic, whether it happens or not, just the way you say, I'd love to get an idea of where you think things might go. 00:40:56.565 --> 00:41:48.110 Yeah. So my strong hope is that the work I am doing right now and the work I did in the past few years is gonna serve as a massive enabler for people. Like, as models are reduced in size, as time goes, as compute gets better, I hope that everybody will be able to run the same pipelines I've developed and I'm developing on their own laptop, on their own infrastructure, in their own houses, and to enable them to do whatever they want. Like, whatever they want, whatever they want to improve their house. As you were saying earlier, one can, you know, apply this to all the cameras to inspect the wildlife going around the house, as well as doing something very useful, but also for work, or for your personal life, or for optimizing some of the processes that you feel they are very, very annoying. 00:41:48.110 --> 00:42:11.360 Like, most cases, the best innovations come from, that's my personal opinion, come from two factors. One is the constraints. So when you're under constraint, usually you develop best solutions to to go around them or to work around them. And whenever you find something that is very annoying to you. Whenever I find something that is very annoying to me, I really want to get it done. 00:42:11.360 --> 00:42:14.880 Like, this shouldn't be annoying anymore to me. What can we do to make it better? 00:42:15.125 --> 00:42:23.525 I like that. That's... I I like that a lot. And and I like that perspective that's kind of holistic to you... To life there and towards... 00:42:23.525 --> 00:42:45.435 Yes. You're doing things for work and your job and things like that, but you're also enriching your life, in that way. It's a great perspective. Lorenzo, thank you so much for coming on the show. I've really enjoyed the I learned a lot from you, and I hope that as you proceed with your research going forward, that you'll come back on the show and some 00:42:45.435 --> 00:42:45.595 of the 00:42:45.595 --> 00:42:47.115 things that you've learned in the future. 00:42:47.115 --> 00:42:49.515 Totally. Thanks a lot for having me. It's been a pleasure. 00:42:49.515 --> 00:42:51.115 Absolutely. Thank you very much. 00:42:51.355 --> 00:42:52.155 Thank you. 00:42:56.955 --> 00:43:17.315 All right. That's our show for this week. If you haven't checked out our website, head to practicalai.fm and be sure to connect with us on LinkedIn, X, or Blue Sky. You'll see us posting insights related to the latest AI developments, and we would love for you to join the conversation. Thanks to our partner Prediction Guard for providing operational support for the show. 00:43:17.555 --> 00:43:26.880 Check them out at predictionguard.com. Also, thanks to Breakmaster Cylinder for the beats and to you for listening. That's all for now, but you'll hear from us again next week.