Mike Vizard: Hello and welcome to Techstrong.ai. I’m your host Mike Vizard and today we’re talking with Etienne Bernard, who’s the CEO for NuMind, and we’re talking about how large language models will be used to drive NLP models and how symbiotic all this is going to turn out being. Etienne, welcome to the show.
Etienne Bernard: Hi, thank you. Thank you very much. And thank you for inviting me here.
Mike Vizard: So you guys have a lot of experience with natural language processing, otherwise known as NLP. How will these large language models that are used for generative AI platforms kind of come together and the two maybe drive each other?
Etienne Bernard: Yeah, I mean, it’s a big revolution that is happening. I mean, I think everybody has realized that, since GPT, and they can be used, this language model can be used for so many thing you mentioned, to generate things, but also to help us interact with a computer in a different way through chatbots, through the sort of digital assistants that I think are gonna spawn everywhere. But one, one other thing that they can do is just understand text. And there are many important applications for businesses that rely on that, such as analysts, analyzing the feedback of your customers, for example, understanding they’re happy or not, what they’re talking about, analyzing topics, all these sorts of more classic text understanding that is now unlocked; thanks to these models that understand language much better than before. And so kind of the idea of NuMind is to use them to tackle these sort of tasks.
Mike Vizard: So is an example of this, is maybe people will respond or engage through text or natural language, or voice for that matter, and the computer will be able to understand their intent and summarize that and kind of make that a little more actionable for somebody on the receiving end of all that?
Etienne Bernard: Oh, for sure, for sure. I think that’s a big application of these language models to unlock these conversational interfaces. And with NuMind, we’re going to get there as well. But we’re starting the path of like, only understanding that for example, classifying documents, extracting information from documents, these sort of things. So you know, legal documents, medical documents, or understanding job offers. So we have one customer, for example, what they do is they try to figure out which is the best job offer for a given resume. And so you need to understand the text on both hands on the resume and on job offer to be able to do this match. We have another customer, which, for employees to ask questions in the company, they try to understand the question and then understand the abilities of the other employees and figure out which of these other employees can answer the question best, that’s another application. Then I think we’re going to move to a more generative task and more to create conversational agents, because I believe there’s a way to create these chatbots in a much simpler and better way than is currently done. Thanks for these large language models.
Mike Vizard: Do you think people appreciate what’s really happening here? I mean, if I look back in time, we spent 40 years creating all kinds of abstraction layers to talk to machines. And now we can talk to machines as if they were another human being and it kind of changes the whole interface. And I’m not sure everybody kind of just suddenly gets that.
Etienne Bernard: Oh, definitely. I agree. It’s going to change software, because until until now, it was either through graphic interfaces, or through code that you could interact with a computer like the way we use a software. And now with this ability to talk in a natural way with natural language it changes a lot. So I think software is going to be changed a lot. It’s hard to say exactly what it will become. Some people say there will still be graphic interface but they will kind of generating on demand when it makes sense to create this graphic interface. And yeah, and all the assistance of course, I think we’re going to have, you know, digital assistants in your phone that can help you to do so many things; of course, you know, book train tickets, but also you know, help you to understand something like a tutor, personal tutor for basically every kid and later on medicine to have some sort of medical assistance, and even later on when they are good enough to have a medical doctor in your phone. And yeah, this application of course – I think it will have big importance, the matchmaking ability we just had to deal with customers, you know matchmaking between a CV and job offers. And we can imagine that when we have that kind of understand us, and understand everybody else. And if we managed to make them talk, so to speak, in a privacy compliant way, then we can imagine that they make much more, you know, finding somebody to work with finding customers much more easily. I think the matchmaking business is going to explode. And it’s going to fluidify you know, the system and economy in a way. And I think it’s going to lead to an economic boom. I’m pretty, pretty convinced about that. And to come back to your original question, do people realize? Yeah, I think yes. Now I think they realized something is happening, for sure. For sure the other extreme is that we believe that we’re going to have super intelligence in like, in like two years, and it might destroy humanity, which I do not believe, at all. But at least the people that are in the domain, and in tech in general, I think they now see that something’s happening. Yeah, for sure.
Mike Vizard: Do you think as we go along – will I have assistants for every little function? Or will I have this kind of one super assistant that does multiple functions for me? Or, I mean, am I gonna have my financial services and my medical assistant? Or will there just be one assistant that has domain expertise in everything I need?
Etienne Bernard: That’s a good question. I’m not sure actually. I think, I mean, it’s going to, I think there is going to be space for both. And, we see that right now, right? ChatGPT is a general one, but then you might want to customize it; you may adapt it to your task. So there’s also the fact that, in order to be cheaper to run, it’s better to kind of separate it into smaller ones that can run on your phone later, instead of being everything on the cloud. But that’s – I don’t have an exact answer. I wouldn’t be surprised if at some point it is just one assistant, and at least for a personal assistant. Yeah, for a personal assistant, it might be one in your phone, eventually.
Mike Vizard: We’ve been investing in it for the better part of four or five decades now. And productivity numbers never really kind of jumped, they kind of steadily increase. But do you think maybe we’re on the cusp of some sort of massive jump in productivity, and we will see some interesting changes in our economies as a result?
Etienne Bernard: Oh, yeah, I believe so. I believe so. I mean, matchmaking was an example that I was saying. I think will fortify the economy a lot. And then the ability to do stuff will increase, obviously, I mean, we can see now things like copilot and so on, that can write code for you. And I think it’s going to improve, and later on, we all will kind of, at least in the IT world, in the tech world, kind of be product managers in the sense that we’ll discuss the systems and it’s going to be a two way discussion, by the way, which is kind of missing right now, even though ChatGPT went clearly in this direction. And ChatGPT4 as well. Still I think they’re pretty passive in what they ask you; it’s mostly you telling them to do something and then giving an answer, but later on they might ask you well, why are you asking me these and ask a question. Okay, but what what should I do in that case? And much more by directional type of interaction and this will allow people to create stuff, of course, much faster than what they do now. So yes, I think it’s going to lead to an economic boom.
Mike Vizard: We hear a lot about these large language models. And everybody kind of assumes that maybe there’s one big one that runs ChatGPT. But as far as I can tell, there’s going to be a lot of these different language models running around out there and a lot of NLP engines to access them. So the whole thing sounds like it’s gonna get pretty dynamic. How many LLPs do you think we’ll be engaging with over time?
Etienne Bernard: This, okay, this is interesting. It depends on the rate of progress in a way. We’re starting to reach the limit in terms of data. Before we believed that it was more the size that matters, so we made it massive. Google did 500 billion, some time ago already. Now we realize that actually that is quite important. And we’re kind of reaching the limit. So if we don’t improve, like right now, the same architecture is this transformer architecture, which is basically the same as when invented in 2017. And we just scaled that, just, you know, it’s just the same thing with much more data and much more compute; a bit smarter way in the way we select the data. But I believe that we’re going to, you know, GPT4 is amazing. I believe that it’s not going to dramatically improve anymore. And if that’s the case, then we can start seeing another movement, where people manage to make good large language models smaller. And then because compute improves, the graphic card improves, and so in that case, it as you say, it won’t be 1-2-3. Organic companies that own this model are going to be much more democratized. It’s going to be, it’s gonna be much cheaper to train them. And therefore is going to be 1000s. But it could also be that we find, you know, a modification to the current system. And so it keeps getting much better. And then the kind of leader like OpenAI, you know, and a few others will kind of keep this lead of maybe one year ahead of everybody else.
Mike Vizard: If the models are getting smaller, and I need less data, I could probably spend more time making sure that the data that goes into that model is of the highest quality. Unlike, say, OpenAI, which kind of tried to, you know, Hoover, the entire universe. Would that result in those models being a little more trustworthy and more accurate? Because there was more controls applied?
Etienne Bernard: Absolutely, absolutely. Actually, that’s kind of the business we are in to create custom models. And, clearly, but when you start with this base, or within a model training with everything, it’s smart enough that it can adapt pretty well with a small amount of data. And, nowadays, clearly, it’s not about volume anymore, at least to customize, okay, to create this foundation model. Volume is a big, big part of it. But to adapt it to your case, it’s definitely about the quality, the quality of the data you give it; it’s about defining your task, which is not easy. So it’s all about training. The thing already knows how to speak. Now you just need to define your task in the best way, which we know is not trivial, because when you define your task to another human, it’s not you know, it’s not straightforward. You need to define a task and there’s always some misunderstanding. And so you need to discuss and this other human lesson – your students need to perform the task, and then you do some review together. And that’s how you define your task. In the step by step interaction. Which by the way, we call that interactive AI development, to really stress this idea that you need to, to have this tight, high bandwidth interaction between you and the machine to define your task in the most efficient way. But for sure, it’s not about giving it a huge amount of data anymore. This is over. Except for the foundation models.
Mike Vizard: We have seen everything from calls for moratoriums to the Italian government trying to bend GPT. When they figure that out to the other extreme, where people are trying to use it for every silly thing they can imagine, what’s kind of the reasonable middle course here?
Etienne Bernard: I think for now, I don’t see any threat whatsoever. So actually, I think I am great getting crazy, because then we might start seeing some threats. I mean, we need to learn, right? So new technology, we need to learn, and probably better to do mistakes now and figure out what wrongs could happen in order to, you know, potentially regulate how this technology can be used and deployed and so on. So right now, I’m on the side – you know, I’ve seen some open source where it’s like GPT can even modify its own goals. And you know, these include some memory they tried to do; one is called Auto GPT4 or something like that. And then I think there’s others. I think it’s actually good. I think we’ll see what it does. I don’t think it’s gonna give something amazing at the moment. And this is the theoretical idea that these things could improve themselves. And I think it is, to some extent, it’s true, but I don’t think they will prove themselves tenfold. I think we see that it’s actually very exciting. All the exploration, you know, it made me think of a related thing where a language model perform a task, and then you ask them to criticize the task, criticize the answer, and iterate over it. And they give something better. So yeah, this is, you know, regressivity sort of work. When you combine with our software model could indeed improve performance. So yeah, no, I’m not worried at the moment. And I do think that AI will transform the world so much, and also gives power to individuals so much, that they’re concerns on the long term. That’s, that’s for me, it’s clear when everybody can synthesize virus in the home. That’s a problem, of course. So I think there will be a problem. But the thing I’m not worried about is the pace of improvement be so fast that we won’t be able to see this problem coming? And that’s a big question. By the way – the whole doomer’s argument is the thing is going to improve itself. It’s not really clear how, because often you need hardware, so you’d need to add – the thing is going to need better graphic cards to, you know, to run faster to train better. But anyway, this idea that this thing is going to be smarter than us and therefore will create something even smarter than itself, which creates something smarter than itself. And then it’s a complete loop and can go really, really fast. That’s an argument. This I really don’t see at the moment. If in five years, we start seeing these loops happening, then I will say we will need to be more careful. But at the moment. You know, this idea of stopping big training for six months, I think it’s going to create more harm than good. I mean, I don’t think – I mean, I’m pretty, pretty sure. Because all these you know, nice digital assistants, like I want to have an educator for every child on earth, and I want to have a doctor for everyone. I want when everyone has a medical doctor in their phones, I want these medical doctors to kind of talk to each other – privacy problems, of course – to figure out a bunch of things on health. And, you know, I think the amount of good that will be created in the next years with that is so big that we should not slow down at the moment. Unless we start seeing this iteration at some point, which we don’t see at all.
Mike Vizard: All right, folks, we’re at the ultimate teachable moment where the machines are learning from us. And when you’re learning from the machines, just make sure you know where the plug is, just in case things go haywire. Thanks for being on the show.
Etienne Bernard: Yeah, thank you very much.
Mike Vizard: All right. And thank you all for watching Techstrong.ai. You’ll find this episode and others on the website. We invite you to check them all out. And once again, we’ll see you next time.