Nicolas Pineau - Making Sense of AI

Welcome to AigoraCast. Conversations with industry experts on how new technologies are impacting sensory and consumer science.

Visit your favorite platform and remember to subscribe:

Apple Podcasts

Spotify

Please leave a positive review if you like what you hear!

Hi, I'm Dr. John Ennis, CEO at Aigora. In this episode I enjoyed speaking with Nicolas Pineau of Data Insights. Nicolas shared a lot about his experiences both applying and teaching AI within sensory and consumer science, and I heartily agreed with his perspective that we are just at the beginning of an exciting change in our field. I hope you enjoyed this conversation as much as I did. And remember to subscribe to Aigoracast to hear more conversations like this one in the future.

Short Bio Nicolas Pineau is a Data & AI strategist and Sensometrics expert dedicated to transforming food product innovation through advanced data science. As the founder of Data Insight, Nicolas specializes in R&D AI leadership and providing strategic data science consulting for global organizations.

With a deep academic and professional foundation from the University of Burgundy, Nicolas is a recognized leader in the field of sensometrics. His work focuses on bridging the gap between traditional sensory science and modern machine learning, helping companies navigate the complexities of AI integration in their research and development efforts. He is a frequent contributor to the sensometrics community and a leading voice on the practical application of AI for sensory and consumer insights.

Social Links & Resources

• LinkedIn: Nicolas Pineau
• Company: Data Insight (Founder)
• Expertise Areas:
  • Sensometrics & Advanced Statistics
  • R&D AI Strategy & Leadership
  • Food Product Innovation
  • Data Science Consulting

Episode Transcript

Welcome to Aigoracast; Conversations with industry experts on how new technologies are impacting sensory and consumer science.

John Ennis: Okay, welcome back everyone to another episode of Aigoracast. Today I'm very happy to have Nicolas Pineau on the show. Nicolas Pineau is a data and AI strategist and sensometrics expert dedicated to transforming food product innovation through advanced data science. As the founder of Data Insight, Nicolas specializes in R&D AI leadership and providing strategic data science consulting for global organizations. With a deep academic and professional background from the University of Burgundy, Nicolas is a recognized leader in the field of sensometrics. His work focuses on bridging the gap between traditional sensory science and modern machine learning, helping companies navigate the complexities of AI integration in their research and development efforts. He is a frequent contributor to the sensometrics community and a leading voice for the practical application of AI for sensory and consumer insights. So Nicolas, welcome to the show.

Nicolas Pineau: Thank you. Hi John. Very glad to be there with you today.

John Ennis: Oh, it's great. Yeah, actually reading your bio, I feel like you're a man after my own heart. So I'm very happy that you're here. I think we're going to have a lot to talk about today. So I think most people listening to the show probably have some idea who you are, but it might be good if you kind of go through your background and how did you get, you know, to be at Data Insight. Take us through your journey and how you got to get started here. Yeah.

Nicolas Pineau: So yes, before doing my PhD even in in Burgundy, I was actually I did two master degrees in food science and food technology. So my background is really deep into food, since it's a passion for years to just cook as well at home with kids. And after that then I came to to Dijon in the lab of Pascal Schlich first for the the end of my master degree. And this is where they were starting to work with the TDS, Temporal Dominance of Sensations method. So this was actually originated by an idea from F. Coster, who wanted to have a kind of harmonium or kind of piano of sensations where the panelist could play the the song or symphony of of taste. And they started this way and and Pascal embraced this this topic and started with the first experiments with that. But it was a bit too complex for the people to play with 10 strokes and and to play this kind of piano with all the the keys at once. So it came back to a pianist playing just one stroke at a time, uh, still with the intensity. And when I arrived in the in the project it was to to to have a look at this data, see how we can analyze them and and propose these TDS curves. So this was one of my first paper and I guess still today the the most cited ones. And then after that, I evolved in the domain of sensometrics, working a lot on panel performance, um, meta-analysis of many data altogether. And my PhD was really about panel performance and the longitudinal study of of panelists over time. And then I started my professional career at Nestle, where I spent most of my uh professional life a bit more than 15 years, not 50 yet. Uh and there I was starting as a statistician, then uh evolving more as a data scientist, still in the sensometrics world, so really uh diving into um different kinds of problems to uh innovate better uh with our products, being doing some some modeling, some preference mapping, again meta-analysis and spending a lot of time on the data side of it, because as soon as you want to model stuff, you need to have your data ready. And it makes that we spend a lot of time building databases, uh connecting the data, making sure we can be benefit of what we're collecting every day. And uh I uh also was leading a a network of statistician um uh toward the end of about 50 guys all around Nestle to try to dive into new methods. Before moving to the US, where you are John, and here I spent two years working for ADM, Archer Daniels Midland, it's a not so well-known food company as well, they're selling food ingredients or or grains or flours or oils, these kind of things. And here I was director R&D data and AI. And this was in 2022 at the moment where ChatGPT 3 uh started to come. So it was a great opportunity for me to start playing with the these kind of new models, see how we can use these LLMs in a professional environment, see how we can use it to access our own data and and and the have a better way to summarize all of the knowledge you can have in a company. I don't know if it's a famous quote in in the company I was or it's a general one but being both at Nestle or ADM, I was used to hear "if only Nestle knew what Nestle knows" or "ADM knew what ADM knows." And this is always a big challenge and I feel like with this LLMs, there is really a a step forward here to gather data together and being able to to use them. So that was one of the big challenges there to to use these models, as well as some models to generate images like uh what is called GAN, this generative adversarial networks or VAEs or or diffusion models. And this can be used of course to generate images, they were built for that, but we could also use it in another way to to generate recipes or to generate processes. And it makes very good uh uh things results here as well. So it was a very nice playground, also playing with a molecular dynamic simulations and and other stuff like this. And then back to back to Europe, back to Switzerland, uh then I wanted to yes to start on my own with uh founding this Data Insight company, more as a consultant to help companies in their AI journey and see in particularly in the sensometrics domain how to how to grow in this domain.

John Ennis: Mm. Yeah, that's great. It's really interesting. You know, our careers are kind of similar and, you know, similar paths. Your coming of course from a food science background, I'm coming from a math background. But we've ended up in the same place and I've drawn a lot of the same conclusions. So, you know, in our preliminary call, one of the nature things you said to me that I thought was really interesting is that AI, you know, we're talking about what what do people get into trouble with with AI. And you said to me that you don't think AI can generate new information. And I thought that was a really good way to look at things. So maybe you could tell our audience about this lesson and what do you mean when you say AI can't generate new information?

Nicolas Pineau: Yeah, and I thought a bit more about that since we talked together. Maybe information is not the exact word in the end. But um, I go in this direction that for me, even if we take back um, science. Science for me is a kind of back and forth between observations, and we see something, and then we try to model it in a way or another. And this is a theory, and we try to to confront this theory to observations to see whether it works or not. Typically if I take back uh, Newton, when he was seeing this apple going down to to the ground, was proposing the theory of of of gravity. And this holds true for many, many years, uh until uh they came to some issues where gravity theory was not working perfectly well and they refined it and it makes that Einstein proposed the relativity theory which is another one. But it came again from the back and forth between observations, so really something we measure, and a theory, a model, that makes that it seems to work. And it's the same for us, it's the same in what we do in sensometrics, we do a lot of modeling of uh sensory characteristics according to recipe and process parameters. In a given way, large language models are also doing about the same. They also general- they're told generative algorithms but they also predict, huh, an LLM is a predictive model in a way, it's a probabilistic approach that is predicting what is the most likely next word. And it works extremely well. But it's a prediction. And when we hear sometimes that there are some issues that uh um these models may poison themselves because they will learn from what they predict. This is exactly the point, is that it's not anymore an observation that is used to train them, it's a prediction that was made at some time that is assimilated to a new observation. And to me in the exact same context, when we talk typically in our domain more about the synthetic data. And we say that we can generate new consumers. I feel like sometime we we don't keep in mind that these additional consumers that are generated, they are kind of predictions as well. They are they are things that we end up with from our models. But they are not new observations. And if we mix it with real additional observations, then we may be lost, because we feel like we created observation and there is no way that a model is creating observations. It will create a prediction, very good one, very useful one, but they are not observations, they are something a bit different. And I feel like the difference is still important for a good science uh environment so to say.

John Ennis: No, I completely agree with that. It's almost like the information content or the amount of surprise if you think about entropy, you know, information theory, that if you're using your data to generate these new synthetic data and I haven't really thought through the statistics of this, maybe you have, this would be an interesting topic for you. But, you know, you think about independent observations, right, for doing statistics. And the synthetic data are not truly independent, right? Because they're being if you're using the day if you're gonna combine synthetic data with your actual data to try to improve your power, you know, how much how dependent or independent are these observations? Obviously they're not fully independent because they're based on the real data. And so the statistics, you know, the power calculations and whatnot might be wrong if you're if you are, you know, it's a little bit like the beta-binomial where if you got the same people evaluating samples again and again, how much new information is there in those additional measurements, right? You know, in the extreme there's no there's no new information in in those additional measurements, right? If people just always detect the difference or they always don't, then, you know, you're not getting any or a preference test, right? I ask you, you always have the same preference, it doesn't matter how many times I ask you, there's no new information. Um, so yeah this this is quite interesting. I tend to think of synthetic data as, you were talking about, you know, if Nestle knows what Nestle knows, I tend to think of synthetic data as a good way to explore your data. That you might have uh, and that's the way we use it like in Thea or in Forethought, is the idea is if you have a certain amount of information, you want to explore it, maybe you'd like to apply, I I like unfolding, right? And you need a certain type of data set to unfold and you don't actually have that type of data and so you're going to say, okay, based on what we know, what would we expect if we were going to go ahead and generate a, you know, a full, um, you know, suppose that we had what we really want is like sequential monadic data to unfold it. We have a a statistics, you know, we have some approach we want to run and we can't run it because we don't have the data in the right shape. Well, what would we expect? What would be a reasonable guess for what we would get if we were to run this experiment given what we know? And so all you're actually doing there is you are hypothesizing what might happen and you're using AI to help you to do that. But I still think that should only be taken as provisional. That it's a it's it's really exploratory data analysis at that point. It's it's not and you should still run real research, you know, to go So those are my thoughts. But what what are maybe you want to um share your perspective along these lines.

Nicolas Pineau: Yeah. I I fully agree with that. For me as long as we keep in mind the fact that what we have in front of us is generated data from what we collected at first and it's a view of what it could be, I feel like it's a very good approach and actually, uh generative models are very good at generating. They are very good at proposing proposing profiles or consumer answers that look like something that is credible. So it's a very good way to to have what if scenarios. And in this case it can be a lot of things uh very good approach to explore the space, generate new ideas, think about the same problem in a different view angle. As long as we keep in mind the fact that these are generated data that are not observations and that potentially at some stage there will still be a need for a check to be sure that it fits the the proposition. I I feel like it's a tremendous tool and it's a I would encourage people to use it because it it really opens the the the yeah the flow to to new ways that you may not think of. It may also open the, it might be a very good way to focus on a subpopulation that you may not have handled in the same way otherwise. Is it that different at the end of the day than really looking at this subpopulation and just looking at at 10 guys instead of augmenting it to 50? Potentially the answers might not be that that different. But in any case it's a it's a very good way to explore. And I really feel like that when I use these LLMs more for generating ideas, it's it's amazing because it opens the the the yeah, the flow to to ideas you may not have think think of and maybe half of them are completely crazy and you won't go further with them. But for the other half or even if there is just one that is really something that rings a bell for you and makes that it's a it's a new project potentially, it's a new opportunity, then it's a it's a tremendous help. So so yes, in terms of generation and exploring it's a it's an amazing tool definitely.

John Ennis: Right. Now I definitely agree with that. When I go on say LinkedIn and I read people's complaints about AI, I find that it's almost always because they're misusing it. That they're trying to expect the AI to be something it's not. Either they're trying to treat it like a search engine when it's not a search engine, right? Or they're expecting that it is um has the same primacy as a real observation. They're they're losing the perspective that this is either a recombination of information that's been put in the model or it's an extension, it's a best guess. It should only really be seen as a guess or a draft, right? That I mean Google and OpenAI really do know a lot. I mean Anthropic knows a lot about the world, right? They really do have a lot of information. And if you have some sub-sample and you've only collected a few real observations from that group, you might be able to extrapolate that into what's believably plausibly correct. But you shouldn't trust that. You should say, okay, you know, you do your simulation and you say, all right this is interesting. Where did we get this from? You trace it back to the original sources, which in my opinion you have to be able to do. If you can't ground your answers in original sources then I think you really have to be very careful. Um, but that should only be taken as provisional. Just a signal that something's worth exploring. That's a yeah, that's that would be my view. What what advice do you give your clients now when you've got clients, they're getting involved with LLMs, they're getting involved with generative models. What are the general kind of big picture bits of advice that you give them?

Nicolas Pineau: Yeah. It depends at what stage of the of the process. Let's let's say they're in the in the story, but um, more or less for me the the first one is to really give it a try. It's not just to listen before. So it's not just to send a question, take the the answer for granted and and then go doing something else. Um, what I love to do with the the people I can have in in training session is to to test different options, to to to test Gemini, to test Claude, to test ChatGPT and see how are they different. To feel like there is a kind of different tone at the back of it, related to the way these algorithms were trained, with different people, different approaches. Also if you just redo the exact same prompt twice with the same algorithm, of course you have a slightly different answer because it's a a probabilistic approach with some temperature somewhere that makes that you don't have always the exact same answer. So it's a kind of a you need to get used to what's the kind of answer you can you can get from it and you need to parameterize it a bit so that it it comes to to your taste. But to keep in mind that it's a tool. For me it's a great extension of the toolbox that we have in data science. It makes that now we have a a new toy that is amazing, that is doing a lot of very interesting things, but it's a tool, it's not a it's not an intelligent in the same way as as humans are, but it brings a lot of value. It's so helpful to summarize information, to um find some info somewhere even though as you were saying you have to double check it because it may also create completely fancy references. And even I don't know if you tried with your own references, but if you try to list down your your your last papers, let's say two-third of them will be exactly right but then you discover new papers you never know about.

John Ennis: Might be good ideas! Might be a good thing to go do.

Nicolas Pineau: So that's really crazy and it's very good at proposing tests that looks very uh very possible. So this is yeah, you need to practice, I feel like. You need to yeah, to to jump in for a bit, to give it a try, try to make it crash, come back and then have some rules that makes that you know how to to proceed a bit with that. And then when you have specific use cases then you can order it a bit, or if we talk about more maybe even AI agents, where you kind of already have a a let's say part of an agent is most of the time a kind of prompt process where you have some steps where you can define a bit how it should go. And then you have tools that make it makes it that it's not anymore only about large language models and generating text, it might be about making a real calculation or it might be about many other other things. And then it starts to be really an interaction between a new tool that is extremely good at understanding text and syntax and how um sentences are built and your true data that are aside. So for me there is really a kind of a interface to make between between these tools that to make it really useful for for companies. Beginning with connection with the API or MCP or whatever, but there is really something to to appropriate the the the these new tools and and make that they are yeah, they can do whatever you would like to have and not the not the reverse. I'm used to say in in the training that uh LLMs are very powerful assistants but they are often very poor masters. So we just have to be careful that they are assistants, we can ask a lot of things, we have to stay very critical, we have to to to to keep in a yeah, to keep in lead mode so to say.

John Ennis: Yeah. No, I totally agree. In fact I was teaching this morning and I said like we have to remember like the AI works for us and not the other way around, right? And we are managing the AI and I actually think people who have managerial experience have the inside track. It's much easier for someone who has experience as a manager to get good at AI than someone who's an excellent individual worker. If somebody is used to just being excellent at their job and they haven't managed other people, their excellent developer or whatever, I find that very often they'll ask the AI to do something, it doesn't do it as well as they would do it and they say oh it doesn't work. Whereas the manager's mindset is more like you give some work to somebody, they come back to you, you know, parts of it are good, parts of it aren't, and they need feedback and then they do it again and you have this iterative process. Um, does that agree with what you've seen, you know, you've been teaching, your core, I actually would love to talk to you about your courses and how they've been going. I know you've been teaching um some courses over there. Yeah. How how have your courses been going and what have you been seeing you know, as you've been teaching people?

Nicolas Pineau: Yeah, no it's a it's a very good experience actually and yes it for me every every teaching lesson is a bit different because the discussion will go in a way or another based on these on these LLMs again. Um and and yes for me it's very interesting every time to to see how they they expect to to handle it. And um what is interesting is also the feedback they may have and the the way they try to handle it. And as you say there are different ways to do it. Whether either you are really an expert in a domain and you want to have the exact answer that you would expect, and most of the time it doesn't work, because it would be something that is relatively not average but uh it does not go into all of the details of an expert could could have for instance, so it's not as useful as you could imagine. But when you ask questions about a domain that you don't master or something that is not your core expertise, then it start to be very interesting because you learn average plus things that are very interesting and that are put in your own language because you ask the question. So it makes that the discussion is very yeah, it's very it's much more fluid in this case. And actually in the in the training I don't go very far into this this part of the LLM. It's more to to have the people able to to handle the one they have in front of them and to have a couple of references and questions and how to prompt in a relatively efficient way so that they they they avoid hallucinations and this kind of things. And at the end of the day, most of the discussion is not only about the LLMs, but it's also coming back to the data. Cause to me what will make the difference within a company or another, it's not really the LLM because every every other week you have a new uh version of the latest LLM that beats uh all the other ones. So we will never be able to catch up with that. But at a company level what will really make a tremendous difference is the data from this company. And the fact that they do have information that are proprietary, that are coming from their consumers, their products, and this is what will make the difference when they ask an LLM to to go further with this information. So the part of building this data, making sure that the information they have, the observations they they they collected over the over the past years are possible to access in a secured manner, this is really also very very important before going back to the LLM because everybody can go to Gemini, Claude or whatever and ask the same question. Um, but if you input your own data sources then it makes a much much bigger difference to the at the end of the day.

John Ennis: That's right. Right. There's no real no real competitive advantage to just using Gemini by itself because everyone can use Gemini. So I totally agree with that. Okay. Let's talk a little bit about machine learning because I know that's another area that we have um kind of common uh interest in. How have you seen, you know, you think about tabular machine learning and how, you know, we've found for example that the whole area of connecting text and quantitative numerical data is really just exploding right now. That, you know, predicting for example something we're doing a fair amount of now is a client might have um many rows of data where they've got a sensory profile and then an expert description. Sensory profile, expert description. And we have found that um some of these models like the the Google, I don't want to get too technical on the show but T5 Gemma 2 is a very good model, it's a encoder-decoder model from um Google. It can be fine-tuned to predict with very good accuracy the descriptions within the universe that the you know client is collected a lot of data, what would be the text description of this profile? Or if you have a text description, what would be the best sensory profile. So what sort of things do you see in your own work kind of like now when you go back to the more tabular machine learning world? What sort of things do you see kind of opening up and you know is it more about preparing data so you can do the tabular machine learning? Is it more about going further with your interpretations? Like how is it affecting the more traditional machine learning?

Nicolas Pineau: Mhm. Yeah, definitely I feel like what you say is a yeah, yeah, completely true. The fact that there is a whole field of of potential here between text data and and tabular data and trying to make the relation between between the two. Um there are some even some packages now in R where you do have already some people try to interface both and make that somehow on a classical procedural approach you transform the results of your ANOVA into some text information, so kind of uh this product is significantly higher than this one and so on, and then you feed the LLM with this information to go a bit further into the interpretation of the data. So it can be a for me a way where you benefit from the help of an assistant to interpret a bit further your data or at least to give you directions of interpretation that then you can challenge on your own. So this is definitely a way that might be of interest. I'm thinking about the the package nlr for instance that was developed a couple of years ago I feel like. That might be very interesting there. Um and there might be also the other way around is that if as a company or as an expert you're able to crystallize a bit what's your knowledge of a domain and feed it to an LLM, then you have an assistant that suddenly uh is able to help a lot of people in the in the company based on that. And if part of this knowledge is based on tabular data, then then of course there is a very interesting things that can be done there. Something I didn't investigated that much so far but maybe you did is this TabPFN uh...

John Ennis: Oh I've worked yeah I've had a lot with TabPFN. We have a machine that runs TabPFN in the other room. We have two Mac Minis that are set up to do TabPFN!

Nicolas Pineau: Okay!

John Ennis: So I feel like yes this is definitely a big one. But I feel like you have more experience in this one. Yeah, I still need to to. Okay, well you and I should we'll after the show I'll I'll tell we'll go through some of that because that's really exciting that um yeah, okay. I could talk about this forever so let's get back to you. Um. Okay. Yeah, well this is all very good. And I think communication, that's another thing that you kind of mentioned this a little bit with the interpretation. I think that in sensory one of the things that's held us back as a field inside companies is we do really good research but then it's hard to share it out and hard to make it real for management, you know? And I think that um any tool that can help us not just, you know, interpret our results but then connect them to the larger business goals is also that's some that's a big area. I mean is that something you're like I'm not really sure what's a typical training for you? Like when you're I've, you know, some of my clients have told me about your trainings and that they want excited to go to them and all this. But I don't really I haven't really kept an eye on exactly so what is it what are what would be a typical training for that you would run for a company?

Nicolas Pineau: Yeah. So so most of the time what I'm asked to to to go for is let's say it's in in four parts. So one is more about AI in general, so about the vocabulary of AI, about how do we use LLMs, algorithms to generate images and and videos and how to handle all of them together using tools that you're very used to like NotebookLM and all of these. This is already just how to use existing uh LLM tools uh inside. Then I'm going more into how to use it for on on the consumer side of it. So and here it's more related to the analysis of uh open comments, uh the the way to even before that, how to generate ideas, to generate personas, to generate product concepts that might be of interest that then you can challenge from no idea at all to a presentation that you can sell or you can uh provide to your your your manager or this kind of thing. So how to make the consumer story from from start to to an end. And then the third part is more about the sensory side and for me on the sensory side it is more really about how to have this kind of AI assistant that can drive you uh in the in the maze of all of the statistical analysis that might be of interest and helps you to know which one to go to. Because most of the time you are used to your software where you have the collection of data and then you have some of the options for the analysis but maybe not the best for you, and here the the LLMs or any agent can help to drive you to some interesting uh new ways to look at your data typically. And the last part is not really AI again, it's data. It's the fact that if people want to go back to their company and start with their own AI agenda, they need to be clear about what are the data they have in front of them that they can use to make the best use of out of it and make that it will really differentiate from what another competitor can do. So this is mostly about these these four four pillars I would say and then it depends from time to time I can focus more on one or or the other.

John Ennis: I see. And have you had to deal very much with uh AI skeptics? Because I've been fortunate that, you know, I even teaching my course and pretty much the people who show up are people who want to learn AI and they're bought in. But have you have you been in a situation where you've been you're teaching a course at a company and you have people there who really don't want to believe that it's going to work or they really are against it, or you know, have you had to overcome these, you know, kind of pushback? Fortunately it's not something I've had to deal with very much in my um...

Nicolas Pineau: To some extent a bit in the sense that I'm usually asking about the people what what does it mean for them AI and what they what do they think about when when talking about artificial intelligence. And very often I would say two-third of the population are are very positive about it, they see all of the opportunities, that they can do their job more efficiently uh and they can be faster and explore other opportunities. But there are also always, maybe because this is the the the French classical touch to to criticize everything, but there is something about um I may lose my job and uh I will uh I won't have any opportunity to do my job anymore because AI will do it instead of me. And this is a topic that we most often discuss quite extensively also in the training to say that for me, as any new technology it's a it's an opportunity that you can take or not, and the thing is that it might be a very very good assistant and it can help a lot to be more efficient to go further to that what people are doing now and it can be an opportunity to spend more time on something that you prefer to do and you like better to do and spend less time on admin tasks and and these other other approaches. But I understand the concern because to some extent yes, uh AI seems to be able to do things uh faster than humans. Still we have to keep in mind that it's just a tool and it has no right for decision, meaning it will never be judged in court. Uh the user will. Uh because at the end of the day if somebody is using the results of an AI to take a decision, the decision is the responsibility of the human. And it's not the responsibility of the AI. Maybe it will change and in the future there will be some uh laws to judge AI algorithms but today it's not. So I understand the concern and at the end of the day for me the point is it's it's an opportunity as a tool to grow and I feel like it's way better to try to embrace it than to refuse it because it might be much more harmful.

John Ennis: No, I completely agree with that. In fact I was in a starting an automation project for a client and we were working with the people who are doing a lot of the what you might call mental there's really is something I would call mental manual labor, right? Where it sounds like a funny thing to say but it's true. These repetitive relatively boring tasks, you know, that involve reshaping matrices and moving this here and this thing from Excel goes over there. Or, you know, let's read all these um documents and extract whatever, and it's uh sometimes quite boring work. But it's somebody's job to do it, right? And, you know, I was talking to the people and I was saying look, the the problem the world is not going to be problem free. There will always be problems. And actually in my experience, I'm busier than ever. You probably have this experience too. That I have more to do than I've ever had to do before because suddenly AI has made it possible for me to do all these things I always wanted to do. And now I can do them. Now I can do things that I didn't even know were gonna be... I can work on problems that I never would have even imagined I could work on. Um, so I think it just keeps on unlocking new problems to work on. And as long as humans have the mindset that they care about their job and that they care about, you know, if you work at a food company providing good food that people love at reasonable prices, that's what you care about, there's always going to be new stuff to do. And maybe, you know, the the job will change, but it's not going to be the case that there's nothing to do. There'll be stuff to do. So yeah. And then to the last bit about the ownership of the decision, I also find that machine intelligence is pretty orthogonal to human intelligence in the sense that um very, very good at uh knowledge synthesis, very, very good at um, you know, automating repetitive mental tasks. But not very good when it comes to assessing how important things are. There's a kind of a lack of common sense. And you know, somebody has even if a machine can optimize to a metric, but somebody still has to decide what metric. Because there's it's a big multi-objective optimization problem. And how do you, you know, trade do these trade how do you manage these tradeoffs? Only a human can really say what the right tradeoffs are. And so um I don't I'm six months ago I was very worried. I'm not worried anymore. I feel like I I really don't, I feel quite optimistic that things are gonna be okay. All right, well we got to wrap up here. But um what advice do you have then for people, you know, you've already given a lot of good advice about try just try it, just whatever. But you know, as we go into this new way of working, what are some tips you would give to the, you know, to the audience on how to continue to succeed with as these tools become more common?

Nicolas Pineau: Yeah. Um, I guess I'll gonna repeat myself. For me it's a it's an opportunity to learn faster than ever for us as human because we have a tool in front of us that can really drive a discussion and it makes that spending time with it to really handle it and have our own opinion about it for me it's the the very first step. So um, I I love to learn new things every day and for me it's the best opportunity ever because I I don't need to search for the information. I can get it. Of course I need to check for it, but at least I have a very good first hint and this is this is uh this is very useful. And and for me um, yes it's to keep this kind of awareness about what's coming next and uh as we discussed also uh before. Uh for me it's just the beginning of the story. LLMs are a very good step and I feel like we are close to um a level where in terms of language uh we already had made tremendous progress and there might not be much much bigger advances. But there are other domains, talking about robotics or vision or these kind of things. If you think about these world models typically where they try now new approaches not to be based on on text but to be based really on trying to understand uh let's say the the the latent variables behind the scene in images to be able to to predict phenomena like uh like uh yes, bouncing of of a ball related to gravity again. So there will be many other things to to come. Uh they are people at MIT working also on other architectures of these neural networks to make that we don't need again hundreds of hours of videos to train a or thousands of hours of videos to train a car, but a couple of hours like as a human could do when he is uh learning how to drive. So there might be many other things to to come uh typically with these liquid neural networks, I'm I'm keeping an eye on that, I feel like potentially something important will will come from there. So I'm sure uh if we would have this talk in one year time, there would be so many new things on the table.

John Ennis: Oh I agree. Yeah. No, it's great. All right, well Nicolas this has really been a pleasure. It's been a nice to find another, you know, traveler on this journey that you're out there, you know, it's really, really nice to talk to you. Well thanks a lot for being on. How can people get in touch with you? What's the best way to reach you if they want to connect after the show?

Nicolas Pineau: Yeah, sure. I think it's relatively easy to find my website and typically on LinkedIn, it's also easy to to find me.

John Ennis: Okay, sounds great. All right, well thanks a lot Nicolas.

Nicolas Pineau: Thank you.

John Ennis: Bye.

Okay, that's it. Hope you enjoyed this conversation. If you did, please help us grow our audience by telling a friend about Aigoracast and leaving us a positive review on iTunes. And if you'd like to learn more about Aigora, please visit us at www.aigora.com. Thanks.