Why No Models Today Will Be Used in a Year

Welcome

So twenty two years ago, I fell in love with a chatbot. Her name was. And really, really, I was, you know, summer night in Paris in my basement alone with my computer, and I I thought This could go in many directions. Up to you. I was mesmerized by this chatbot, you know, the fact that the machine is trying to understand you, language, and can generate some words. And I was really, really struck by this thing. And I decided, Okay, I will just spend my life building chatbots.

And so I founded a company doing customer service bots twenty two years ago, very early, and this is how I started this series of companies into domain.

When we started, I thought chatbots would become very, very intelligent after three, four years and get to AGI and replace humans in call centers. And the more I worked on the problem, the more I realized it was really, really difficult to do that. Actually, the first time we released the bot in Europe was in 2004 to the French railway company. And we were very, very happy to have this customer, and the deputy CEO tried it. And she asked, my name is wrong on the reservation. Our bot answered, hello, wrong on the reservation.

And then we realized, okay, there is still enough work. It’s not smart at all. And so the more I worked on these things, the more I realized it’s not ready yet. So suddenly, for the last year, two years, we reached a point where market timing might be finally finally good, you know, where the product is ready and people have evolved to. Maybe ten years after Siri, people have changed. So maybe the market timing is now.

I was sitting at Terrace at Facebook headquarters in Menlo Park. You know, we joined Facebook through the acquisition of my second startup. And sitting at this terrace with, like, cocktails looking at the sunset, suddenly, it reminded me of a scene in Silicon Valley. You know the show? Yeah. I love it. Where as they go on the roof Al Khuli, and there are these people doing barbecue and say, yep. We are vest and rest. And and I saw myself in this vest and rest situation. And after four years at Facebook, I suddenly realized, okay, it’s time to go out and back to the arena.

And we’ve learned so many things at Facebook AI Research. It’s time to try to push these things to the real world.

So first thing that amazed me when I arrived at Facebook is I thought all big companies were slow. And actually, when we arrived at Facebook in early twenty fifteen, there was maybe 6,000 employees, and it was going so fast. You know, it was a machine, very well oiled machine. Engineering was so efficient. I learned, well, it’s possible to be that big and very, very efficient. And when you are in this situation, there is nothing you cannot do. Just after I joined, I started to work on a project, and I wanted to hire 200 people to Human Concierge to train my bots in real time.

It was not in the budget, of course. But a week after, I had a meeting with Mark. And in about ten minutes, Zucker challenged a little bit my idea. And he said, Okay, let’s do that. And I could hire 200 people in Menlo Park. And when you have this speed and agility and unlimited resources of a company like Facebook at the time, I was really shocked by this at Facebook.

Many things. One thing I like, he spent not a lot of time outside the company. You know, he had, like, thirty minutes meetings back to back from morning to evening, and he prepared really well every meeting. So you have to send in advance, like, a short note about what decision you are expecting from him. If you fail to send this document twenty four hours, you know, by the minute before your meeting get canceled. And so in many cases, you arrive in front of him. He’s read your document.

He’s gathered all the information, the data he needs. So the the meeting is very fast. You know, he will challenge what you ask for maybe and make a decision. So this way to work, I think there is much to learn. I’ll give you an example of that. When I wanted to hire my 200 human concierge, my reasoning was we will train this AI. It will work really well, and we will steal business from Google in terms of search, like exactly what chatbots may do in the time to come.

I prepared answers to every possible objection. He could object about everything I was ready to say why my ID would work. And I arrive at the meeting, and first minute, he tells me, Alex, I agree with you. It will work. So instead of hiring 200 concierge, let’s hire 100,000, and we will kick wheel tomorrow. And then I completely reverse my course and said, no, no, no, Mark, actually I’m not sure if it will work. It will be crazy to hire 100,000 people. And I gave all my reasoning on the opposite sense, like we should be careful.

This thing might not work and so on. So very smart of him. I don’t know if it was on purpose, but he learned more about the truth of my project, attacking it from this direction, this angle, and it was very, very interesting.

It’s not getting easier. You learn some lessons. You don’t do the same mistakes again, hopefully, but get new dangers, know, new issues. The one I suffered from when I started Nabla is what I call the Kim Jong Un entourage trap. You know, you you see this video of Kim Jong Un, and he’s visiting something, and all the generals are around him with a notebook. And everything you say will be written down. If you fail to do that, I I guess you disappear the next day. And nobody, of course, will ever, ever say, disagree with you or challenge any of these decisions.

And starting my third company, after two exits, I felt like my investors were always agreed. My team always agreed, people around us. And so I wasn’t challenged enough, and then we made some early mistakes that probably were not challenged enough. What were some early mistakes? So when we decided to start in health care, we were an engineering team, you know, very strong in AI and engineering, but not so much in health care. And we made actually a good decision to start as a b to c health care company to learn.

But then we lost our way a little bit, and we went too far before realizing that if I thought the right path and that we should go back to a B2B business model, we probably took too much time to make these decisions. Just an example.

The solution first is, I think, is with external people. You know? Even my investors were too close to me, but if you go to people farther away, they have less who are very successful, and they have no incentive to please you, and you get more information, more data from them. And internally, it’s a matter of just telling your team, okay. You should change our decisions, and we did that a lot at Nabla, and sometimes too much, you know, they change everything after that. Also, after you make one or two mistakes, the team will more naturally come forward, and this is before you reach the right level of challenge.

For the general public, it looks like a very big step function with huge advancements every ten years. I think from the inside, it’s much more continuous. So for instance, chatb GPT, you know, is based on GPT-four, which is based on GPT-three, was released three years ago. GPT-three is a large language model. It was invented before that. Transformers, where the paper was released in 2016. The progress is continuous, but the public perception of it is very, very discontinuous. Can I ask you, what do you make of the

From our standpoint, it’s it’s really ridiculous. I’ve been building AI companies for twenty two years, so I I’ve seen this cycle several times. So after a while, you’re not surprised anymore. The only thing you need to know is when to add or remove AI from your deck, and it would change about every three or four years.

I I don’t think it’s fair. When the C language came out around 1972, some people said, hey. Now it’s so easy to build software. Every software is just a thin layer on top of C. And, of course, it’s not true, and we know it. The same when databases came, you know, in the early eighties. So it’s not fair to think that any AI application on top of LLM is just a thin layer. LLM is a new kind of resource. It’s like a new kind of infrastructure that everybody can access, sure, but then it’s completely novel.

It’s very hard to control. You have hallucinations. It’s it’s non deterministic. It’s very hard to configure. You have several knobs you can change between fine tuning or how you prompt it and many other. It’s very hard to control. And also, changes every week. You have new LLMs coming out that claims to be better than the ones from the week before. How do you decide when to change your LLM? How do you do it without breaking your existing users and products? There is a new infrastructure. There is a new game, new rules.

The companies who are the faster to understand these rules and play along these rules will win and and build the best products.

I think I have because I know how it works internally. So I know the limitations. When something is wrong, I know where I should poke to find a solution. I also know when there are many, many and more and more LLM’s available, which one will probably be the better suited to my problem. And I know what kind of machine learning I have to build around my LLM to balance for the weakness of it.

It’s how it will work, of course. You you have in every two weeks, incredible things coming out of research. And if you just assume it’s a static in a year or from now, your products will be very dumb compared to your competitors’ products or more expensive to run or very slow.

I mean, by construction, the LLM has to output something. And so if there is nothing to say, it’ll make something up that looks natural. So this is probably what Emad meant. By design, he cannot not output something. And so this would be an hallucination.

Absolutely agree with that. Will you drive the car you drive today in ten years? I don’t think so. There is so much progress so fast that I don’t see why we wouldn’t use the same in one year.

I think this is always one train late. It was so having a lot of proprietary data was very, very important for the last cycle, five years ago. Maybe it’s less and less true. You need some to bootstrap your model. So for instance, at Nabla, we built a data set of 30,000 medical consultations with patient concerns. We have to pay doctors. It’s very hard to build this because we need this to bootstrap our product. But then with the new pre trained models, fine tuning is very efficient with a little bit of data.

There is a I don’t know if you’ve heard about LLMi model that came out three weeks ago.

So large language model is trained in two phases. First phase is unsupervised pretraining. So it’s just fed with huge amount of text, and it’s trained to predict the next word. And so this is what we call unsupervised pre training. Unsupervised because you don’t need any human intervention. The text by itself is enough because if we try to predict the next word, and on and on and on. And then the second phase is fine tuning, when you use different machine learning techniques. So for instance, reinforcement learning for chat GPT, where you train your model to follow instructions.

Like, for chat GPT is when I ask a question, you should tell me something that looks like a right answer or looks like a reasonable answer to what looks like this question. And so the fine tuning is the second phase where you give more precise instruction on how it should behave. To do that, yeah, there are several techniques. So everybody is still learning. You know, it’s like huge bag boxes, these things, and everybody is still discovering new things. But three weeks ago, there was a paper about LIMA.

This LIMA paper shows that with only 1,000 question and answer examples, so very, very small data sets, they get something for use for fine tuning, so the second stage. They get something that performs better than GPT-three and almost at the level of GPT-four with only 1,000 QA for fine tuning. So the thing with LLM is it looks like a huge pretraining with text. It’s running mostly through pretraining with huge amount of text. And then with a small amount of data, with high quality data, you can get a lot.

You can get what you need and train your model to do what you expect with very small data sets.

We probably have most of the existing companies. Many of you will will be started. There is a lot of hype, and it’s public, so I can spill the the bin about Mistral, a new one created by three engineers, two of them from from Facebook. And I know them, and you can still build this financial model.

To make a successful financial model company, you need incredible scientists and a huge amount of money. And if you check these two boxes, I think you can still do it. And you need focus. Google and Meta could have done what OpenAI did, you know, easily, and they didn’t because it was not the priority, and there is only legal people around. But I think a new startup with enough talent and enough money, and when I say enough money, I’m talking about millions, just to be clear, then I’m sure you can do at least as well as OpenAI.

So incumbents have a huge advantage through distribution. That’s the one thing that is hard for us, start ups. But incumbents have many disadvantages. I think first, they are very slow. You mentioned Adobe. I was sure you would because this is the one example that everybody has. But then who else is is was as fast as Adobe? So I think most of the incumbents won’t move fast. And in in many cases

That’s fair. You know, some of them can move fast, and it’s not fair to say they are slow, and that’s it. I think the incumbents suffer from problems because in many cases, they will do AI enhanced features. And this is what all the one you mentioned did, the Notion. You still have a document. You have an edit. You have a cursor. You write, oh, by the way, you can call chat GPT. And to to to summarize a paragraph, it’s what I call spreading a little bit of AI dust on the magic dust on your existing product.

Who knows how differently you can think about building knowledge for your company? Probably something will come and destroy Notion and Google Docs and all of them with a totally new paradigm that is made possible by AI. Certainly, these incumbents won’t won’t do it. So I think the best incumbents can benefit from AI to be competitive in their existing markets, but I don’t think they will invent totally disruptive things that will kill them. Kodak invented the digital camera, and of course, they never released it because it would kill their main revenue, which was the films.

I think the reason is simpler than that. The reason is nobody could predict that LLM would be so useful and powerful before you train one at this scale. And who in the googleorg chart had the incentive to invest $500,000,000 and just to see this without any business benefit for the company? If you add to that the legal department who is not that happy that you are, like, releasing random chatbots who can say anything, then nothing happens. And this is what happened at Google and probably also at Meta, where nobody has the incentive to do that.

And so OpenAI, as a private company, they need to show something to build products, and they have this curiosity and the financial power, and they did it. When

So, obviously, I think the foundational model that will win will be open. But that being said, we should be careful because even an open model trained with open data, to me, is not that open because when you have like 300,000,000,000 parameters and it’s a huge black box, don’t understand why the output is what it is, is it really open? And so I’m just putting a little bit of salt on what I hear. And, like, open models are you understand everything. It’s explainable. It’s predictable. It’s not true.

An open model can be as hard to predict than that’s a closed model.

I’m not sure about national datasets. If you want to sell something to a nation, it’s good to say that. But more per industry, per vertical, that I said, of course, will improve the quality of the model. So if you train an LLM only on medical data, you get a better LLM for medical applications. But even if you feed the LLM with curated data, it’s not guaranteed that the output will be good, will be perfect, that it can trust the output. So feeding your LLM with trusted data doesn’t make the output trustable because of how LLM works.

So there are

So an LLM is a huge probabilistic machine. It’s like an autocomplete. I think it’s the best image. And it will complete the beginning of your sentence at any cost. It will always complete it. And you may have like 10 sentences, let’s say, in the input data set that are true. But maybe the LLM will start with the beginning of the first sentence and switch to the end of the second another sentence just because it looks good. But logically, factually, this output will be totally wrong. And I simplified a lot, but this is exactly how it works.

And the form is perfect. It always looks like sounds like very good answer, but the the reasoning, the the the facts may be totally off invented or wrong. And even if the input dataset is wrong so when people say chatbpt gave me a wrong output because it was trained on Reddit and Reddit is a lot of noise, Yes. But only partially. Even if you remove Reddit from the training dataset, it doesn’t mean you won’t have, like, random answers.

I fully agree with Yann. And I think if you really understand how machine learning works and you’re not looking for free publicity, I don’t see why you would say something like that, that getting more intelligence will make them want to kill humanity. I don’t even understand the path leading to this thinking. Why does someone

I’m not in the head of Geoff. I I don’t know. I I don’t know. And I I know Yann is also surprised and Yeah. Do you think Elon

It’s weird. At the same time, said that I know he was trying to build a team to compete with OpenAI. He hired somebody from DeepMind, like the same day where he announced we should pause for six months. And it looks to me that the people who are proponents of this pause or to more regulation are the one who feel they are in advance. And so it’s like a way to say, guys, let let us. We are in front. I don’t want more people to start the race.

Can I ask you?

Finally, it’s here. When I see people using chatbots and learning to do their job differently with the help of chatbots, I really feel we are on the verge of finally having a huge impact with chatbots. For the first ten years, I tried to convince people that chatbots were ready and to buy my chatbots. And they say, no, it’s not. And then for the ten following years, people told me, oh, chatbots are incredible. We can replace human with chatbots. And I said, no. No. No. No. Don’t do that.

It’s not really. And finally, now I feel we we are at a point where many things will be impacted by chatbots, which doesn’t mean, you know, replacing people with them. It’s a great tool, but it doesn’t mean it should because of the chat form, it’s tempting to think, okay. Let’s just replace this guy, this worker with a bot. In many cases, I don’t think it’s ready for that, but it still have a huge impact of how work is done.

AI will not replace doctors, but doctors who use AI would replace doctors who don’t. Definitely. AI can have a huge impact on doctors the way they work, and those who will embrace this change will strive.

The first thing I learned when I started to operate a health care company is that for the first fifty years of computerization, computers have been a bad news for doctors. Before computers, they see patients, they care. There is a little bit of admin work that is done by somebody else and not so much paperwork. And they’re very, very, very happy. And suddenly, people like us come to them and say, Okay, now there is an electronic patient record. You need to document. You need to fill this form.

You need to do this and that. And we have trouble to get the money from the insurance company, so we need to document more make sure the claims go through and so on and so on. So the other day, I visited a small clinic in Paris. And all the doctors looked very happy. And I said, what happened? It’s not long. No, no. The computer system is down today. We are just using papers and pen. And they were so happy, all of them. And so you have to realize that the state today is a very bad state where computers were bad, bad news for them because they are drowning under this administrative work, documentation, for many reasons.

Doctors spend on average 49% of their time doing this kind of administrative task as opposed to caring for patients. It’s bad for everyone. It’s bad for them. Three out of four doctors suffer from symptoms of burnout. It’s bad for patients. Actually, three out of four suffer from burnout. Yeah. Three out of four doctors have suffered from burnout symptoms in the last year. And the main reason of this is they all mentioned the pressure of all this admin work, basically. What is

The biggest work they have to do is clinical documentation. So they have to document everything they hear, their decisions. And in big systems that we call EHR, you know, electronic health records, which are dinosaur systems. And the reason they need to do this documentation more and more, first, is the financial reason. To get reimbursed by the insurance companies, especially in The US, you need to have like a very solid file with all that. Otherwise, the insurance will take the first opportunity, the first pretext not to pay your claim.

The second reason is for legal protection because medical malpractice trials, you know, it’s so expensive. Your best protection is to document this into the right way. And so this is why mainly, I mean, work started to grow bigger and bigger. And these EHRs are very, very old and clunky systems. I checked the other day. There was one specific clinical action that took two twenty seven clicks, mouse clicks, to be performed in the system. And I posted this on LinkedIn, and many doctors answered and said, no, actually, it’s more than 300 clicks.

And so this is the stage today. So when we look at

Yeah, so at first, the solution, I think, how AI will impact doctors, it will bring an AI assistant to every doctor, to every clinician. This AI assistant will be aware of what’s happening, what the patient is saying, what data we know about this patient, what documentation, what work needs to be done.

We don’t record the meeting, but we capture the audio. We don’t store it. We just capture it and then drop it. And so, yeah, it gets a lot of context from the audio environment, what we call ambient system.

So this AI assistant is connected to the EHR. It So will get the data from the existing source of truth, which is these patients’ records.

So in theory, there is a new law in The US. It should be In theory. That, you know, Epic spent billions to millions to try to block, but it finally was after, like, I think twenty years of battle was passed. And so, yeah, there should be interoperable. Like, the new model in EHRs are like CRM systems today. They have an API. It’s easy to integrate.

It’s not a problem. The World Health Organization says that we are missing 18,000,000 clinicians by 2030. So the state is not it will put people out of the jobs. The reality is we are desperately looking for clinicians in every country in the world. So if you can make their lives easier, it’s a big win for everyone.

There are many barriers. I think the mistake that many AI startups are doing in health care and what we did initially is we focused directly on the patients. It’s too much, too fast. Cannot jump. Health care system is so complicated. You cannot regulation, for many, many reasons, cannot do that. Technically, we’re not ready to do it. We don’t have the data. We should first focus on clinicians, way things are done today, change their life with AI, and then together we’ll go to the next step. So I think, to answer your question, that one of the main problems is that we are trying to go too fast to a patient facing perfect system.

I’m spending a lot of time in hospitals. I’m spending nights in the emergency services call center in Paris listening to everyone. It’s How is that, and what do you see? So it’s interesting. You know, I met a doctor in a cafe in Paris. I made a demo of Nabla on my phone, and he told me, Alex, you are going to come to the emergency services call center next week all night with me. It turns out this guy was he’s like the chief emergency physician in France, Patrick Perloo, invited me to spend nights over there.

What I learned spending nights over there is it’s a little bit like a call center. You have the first team is taking calls. Overall, in France, it’s about 40,000,000 calls a year. And the first responder, they don’t make medical decisions. They are not doctors. They’re not even nurses. Their goal is to just to understand what is the address and what is the situation, what is the emergency. It’s very hard to get that because people are panicked and you have lots of people calling for very small things, really close things, and you have real emergencies hidden.

So their job is to sort through these very, very complicated situations when people are calling. And they actually write a note, structured note, with this information. They click submit. And then you have regulation doctors who make the decision, should we send the real emergency services? Should we send the firefighters? Or should we just have a doctor talk to this person over the line? And so it’s a two step decision. And a lot is done around documenting the situation. And these things, you know, it would be so easy to do with what we do at Nabla and, you know, what we know how to do with AI today.

The goal would be to keep the person who is answering the phone because their know how in terms of empathy, calming people down, understanding the context, very complicated situation again. Typically, you have several people. They’re screaming. Nobody knows how many victims, what’s happening. Very, very complicated situations. And these people, I spent hours and hours listening with headphones next to them. They are experts cutting through this darkness and understanding what’s happening. So we want to keep these people, but they are typing with two fingers on the keyboard and missing some information.

It takes a lot of time to get and so on. So having an AI assistant work alongside them listening to the call and documenting. I tried it. You know, it works already perfectly with what we have. It would be a perfect team. So this is an example.

What I saw is a few nights, you know, maybe one in five results in services being sent. And in these cases, three in four are firefighters who have medical doctors and so on. But they don’t handle the most critical medical agencies. They are handled by what’s called in France, and it’s one in four.

So we’ve seen a change in that in the recent years. And up to three years ago, you know, almost every doctor would tell me, okay, you don’t understand our problems. Go away. I’m already fighting against my EHR every day. I don’t need more stuff. And it really changed recently. The core reason has nothing to do with AI is that the health systems are collapsing everywhere. Know, I know the situation with NHS. Here, it’s worse in France. In The US, all the major hospital groups are losing money.

Why are they worse in France? So in France, we have a huge shortage of doctors. And there are like zones in France where you have to drive 200 kilometers and wait for three weeks just to see a GP. So this is obviously a big problem. Globally, the problem is the same everywhere. It’s access to health care. In France, you can get good doctor quickly if you know somebody in the system. In The US, you can get it if you have a lot of money. It looks different, but eventually, it’s a problem of not enough doctors.

Both. Since it takes ten to fifteen years to train doctors, decisions you made twenty years ago have a big impact today. So first, there are less and less doctors. Many of them are aging and will be retired soon. And so the number is going to drop. And in terms of demand, everywhere in the world, have a huge rise of chronic disease like diabetes that takes a lot of medical resources. Chronic disease, we always see as engineers, young, healthy people, see that easy things with medicine where, oh, I’m hurt, I go, there is a quick diagnostic decision.

Maybe I spend two days in the hospital and then it’s over. But what takes resources in the health care systems is not that, it’s our aging people and chronic disease and often combined.

So this is the biggest difficulty when you are trying to disrupt health care, actually, especially for a start up. You have the payer, insurance or public payer. You have the providers who are providing care, and you have the patients. And it’s a very complex relationship with three parties. And typically, the one who benefits from a product is not the one who is paying for it. So typically, the patient will benefit, for instance, from your better health care app, but it has to be paid by the payer, by the NHS.

And if the provider disagrees, they can block everything. So you have to solve these three mass problem from day one, which is very, very, very complex. And so you have to find a balance between, as a start up, if you optimize a very small process, it’s easy to deploy because it follows the lines of the existing system. So for instance, if you do a better way to take appointments, online appointments instead of calling, you don’t change the health care system with that. Even that is hard to push, but it’s easier because you don’t change the fact that there is an appointment, there is a consultation, there is a synchronous visit.

So you fit into the system. If you do something too ambitious, it will never fit in the system. It’s I think it’s impossible to go to market with something too ambitious. So I think as a catalyst, we have to find the right level of ambition where it fits enough in the system to be adopted, but it’s not too much disruptive from day one. Otherwise, it’s impossible to go to market. And who pays you? So learn very hard lessons. If you are thinking about trading a health care startup, the very, very first question I want to hear is who is paying for those products?

Who is paying? What is the business model? Then you can, in the startup, in YC, in all start up schools in the world, they tell you, don’t start with a solution, start with a problem. But in health care, even starting with a problem is not enough. Start with who is paying and then how you do is frame the problem for this person to pay, and then what is the solution for this problem eventually.

So in emergency services, the only potential payer is the government. And so not a good good go to market. You know? Yeah. These are first go to markets. And that’s why I didn’t start with that. Now, you know, we’ve deployed our product in The US. Thousands of doctors are using it every day. We are well funded. Now if the emergency services in France really want my product and the sponsor is, I think, like the god of emergency in France and the minister of health looked at me in the eyes and said, we really want it.

It may be the right timing, but it would be a very, very bad idea to start with that as a start up. And maybe if you’re lucky, you do a pilot, and then you wait, you wait, and you it’s what I call desperate pilot.

We started with The US because the problem we are solving now of clinical documentation is huge in The US, bigger than in Europe. And so the willingness to pay for the existing providers, and again, you need to an entry point in the system is very high. We found this go to market. We also made it possible for physicians to use our product without any approval from anyone. Bottom up, they can go to nabla.com. It’s a web application. It’s a Chrome extension. They use it tomorrow, then they love it.

And then they will talk to their boss and say, we’d like to use this in the full clinic.

I think it used to be true, but it’s not true anymore. When I built my second company, wit.ai, I happened to stumble upon Adam Sheer, the the creator of Siri, at a parking lot in the valley after an event. And he helped me a lot to think about wit.ai. He gave me the confidence to follow what I had in mind. And owing to this confidence, I was able to refuse referrals from Google and others who wanted to buy us too early. And so it has a huge impact on my company.

If I were in Paris, no adam share, nobody has done this before, maybe I would have sold it to the first 10,000,000. Okay, let’s sell it. It’s good. And so it used to be true, I think, that you needed to be in the valet with having these kind of people around you. Now I think it’s less true because it’s everything is distributed. The talent is very, very distributed. It will get more distributed. And I think you should get close to your customers.

Yes. French startups. So in France, we have very good AI engineers, machine learning engineers, because the education system is free. It’s very focused on mathematics. So we produce lots of good engineers, but we suck at growing companies. Really, we are not bad at that. I don’t know why. Maybe we lack discipline. Maybe we

Exactly, Stable. It’s enough. And so if you are normal, it’s hard to refuse a $1,000,000,000 offer if you haven’t done, you know, 10,000,000 before. But this is what Zach did, what Google founders and Sergei and Larry did. So maybe we’re not focused enough or disciplined enough, and we don’t have examples around us to grow this company. Hopefully, it’s changing. We have very good scale ups in France now, but it it takes some time too.

So Europe is probably ten years late compared to US and China. And with a new regulation, we are probably gearing to take fifty years more. Talk to me about that. Why do

So the new regulation is a disaster. If you look at it, the people who wrote this, they they went too fast. You know? Good regulation is about good timing and involving the right people who are actually doing stuff in this domain. I think Europe being late felt like, okay, let’s be at least first in regulation.

Some aspects say, for instance, that you should be accountable. I’m simply saying that you should be accountable of the all the data you use to train your model. For instance, you should make sure that you have a proper license, like explicit consent from everywhere it comes from. And I see why it’s a beautiful idea. Who would disagree with that? But in practice, it means 100 of the LMs that were trained these last three years would be illegal in Europe. It’s not connected to the reality. The intent is good, but the limited puts on how to train a model and how to operate a model makes it in practice compared to what we do today makes everything illegal.

What do European startups do then? We move to The UK. Maybe for one reason, Brexit. Move to The UK? Yeah. Finally, Brexit is maybe was good. You know, I’m half joking. If the regulation if it’s really like this and nobody can challenge it, startups like us may have to move. Or maybe we can keep the team in France, but we have to physically train the models elsewhere. I don’t know. It’s a big, huge tension.

First, I’d wait a little bit because it’s too early. We are still in the demo phase of AI. You know, LLM’s, it’s not really used a lot. It’s beginning to be used. And as I said, regulation timing is key. You shouldn’t be too late. But if you’re too early and nobody knows what actually are the risks, no, not the risks you think because you read some science fiction books, but it’s a real risk. Nobody knows yet. And so I think first, I would wait maybe a year or two to learn from the field what should be regulated and why.

And then I would involve people who build these models, of course, but also people who are using it, you know, the users, the general public, which was not done so far. So they have, like, experts who are smart people, but there is no reality of LLM yet.

They have a key advantage that there is no GDPR or very few regulation internally. And so the data qualities they have and the amounts, both the quality and quantity of data is is incredible in China. And so for instance, for health care data, they probably have everything about every individual in China. And if you are supported by the government, I’m sure you have access for for research to all this. They are lucky for that.

They all have issues, their issues. So we talked about the advantages of China, but they are very close. They are getting more and more close. They have maybe the wrong incentives at the research level. It’s hard to predict what will come out of China. We talked about Europe, where regulation is probably the biggest problem that will keep Europe behind. And we are already behind. That’s my concern. The US, you know, is always in a very good position. Immigration laws may be something that will be a problem eventually in The US because talent is so distributed.

I we can learn deep learning very easily from everywhere in the world on a tablet today. It will be more and more like this. And if it’s so hard to go to The US to work for US company, eventually, will have an impact, I think.

I disagree. AI will enable a new generation of players in every industry that will kill the incumbents eventually. What do you think is that time scale? Five years. Existing services companies, like consulting companies, are embracing AI. They will eventually remove the big data from their website and put AI, and then they will put LLM to replace the world. And they will they will take this business, I think. There was a company called Element AI a few years ago in Canada, a very ambitious company who tried to do services like that.

And as they eventually failed, they were acquired by ServiceNow for sort of months arise. What was the

Make it more diverse. I’m really tired of talking to people like just like me. It’s a cliche to say that, but it would be a lot of value if we get more diverse. I have another quick anecdote. When I was at Facebook, somebody discovered that the model that was supposed to tag images that was 100% successful to finding tennis balls in pictures. Actually, if you remove the tennis ball from the picture, it was still finding the tennis ball. So it didn’t learn to find the tennis ball.

It learned the context of a tennis ball, like a racket. In the second phase, it changed the color of the people on the picture. And if you add more black pixels in the hair, suddenly the model thinks, oh, it’s a ping pong ball. It was very early days when we discovered that some models have bias. And this is bias in this case is it didn’t learn to look at the ball. It learned to look at the race of the people around, and it’s actually making this decision tennis versus ping pong based on the race of the people around.

Yann said all

Yeah. More or less, but all of them, and many models have huge bias. The guy, you know, very strong researcher who found this thing in at Facebook very, very early before everybody talked about bias is Mustafa Sissay. He’s from Senegal. I don’t think it’s just luck that somebody with a different angle than we have, different culture, found out a huge hole in the models that we hadn’t found before.

I don’t think AI will kill investigative journalism, finding the right sources, the right information. So this part of media, I don’t think will be replaced by AI. Now the generation part at the output probably will be disrupted a lot by AI.

That fundamentally, people don’t really change. What makes you say that? I mean, you can help people to grow. It’s true for yourself. So you can correct some weaknesses. You can be stronger at what what you are strong at already. But I think you have to accept that some of the fundamental characteristics of yourself and the people around you won’t change. And if you expect them to change, it leads to big failure and mistakes.

It’s that they think it’s conscious. They are influenced by the form, like, oh, perfect answer. A perfect form, and it’s a chatbot. It answered my question. And influenced by the form, they make conclusion on the deep inside, you know, of the system and say, oh, it’s it’s conscious. And for me, the biggest proof of that is that in 1966, when the the mother of old chatbots was released at MIT, you know, ELISA, the very first bot, for a few months, people around the world thought that AI was solved, that there was a human, you know, sentience AI in ELISA, which is a very, very simple pattern matching system with regular expressions.

But because of this, the way it was presented, people fell in the trap, and I think we keep falling in this trap again and again and again for sixty years. What’s the strongest

It’s very hard. Oh, yeah. Yeah. When we change belief, we tend to forget that we have the contrary belief before.

Yeah. It’s it’s a very good one. And we had exactly the same situation at the beginning of Nabla when I thought, if patients love your health care product enough and you prove the health benefits of your product, then payers will pay for it. And it’s now that I’m saying it so wrong, it was obvious, but probably a big belief we had that was painful to and expensive to learn. Who’s been the most helpful angel? Probably Yann LeCun, especially when we doubt about something or he’s very confident that we are on the right path and it it means a lot.

What makes the best VC? So I worked with Andre and Arvits with my previous startup. It was incredible because they never call. They never ask for anything. We we didn’t even have board meetings. And every time I need something, in twenty minutes, I get huge help. Maybe it’s to find a real estate problem because I don’t have an office. Maybe it’s organizing a huge event, I don’t want to hire a marketing team. I want to use them for one week. Or I have a strategic decision to make, and they give me access to the best people in less than a day.

And so the fact that you are in control and you ask for help and they deliver it, we could grow with dot ai to a very nice company with only 13 employees because we are relying on analysts and analysts for all these things. It was incredible. But for my first startups, I did some strategic mistakes, go to market mistakes, and then I wish my VCs back then would have been more involved with me and coach me. And it really depend on the founder stage. Neutral VC who is not here is much better than a bad VC who can have a net very negative impact on funders.

I’m helping funders every day.

Again, it changed a lot. So ten years ago, VCs in Europe were all from the financial industry. There is nothing about financial industry, finance in being a VC, a little bit, of course, but it’s mostly about entrepreneurship. So ten years ago, all the VCs I was working with in The U. S. Were former entrepreneurs, and all the VCs I knew in Europe were former bankers. In most situations, you don’t need your VC to be a banker, but you need them to be an ex entrepreneur.

Hopefully, ten years’ time, you know, every physician has their AI assistant doing a lot of stuff for them and helping them to be 10 times more efficient, seeing more patients, spending more time with every patient, making better decisions. I also think that at the higher level, decisions that require a big picture of the data, a large view will be taken by AI, not by people. So for instance, in an emergency services regulation center, when you have thousands of people needing help and you have some resources, what is the state of your resources, what is the demand, how you make these decisions at the CT scale.

Today, done by humans, but it’s a miracle that it doesn’t collapse. And sometimes it’s collapsed when it’s a big event. 2033, where are you then? Where’s

So my dream is to build a health care system from scratch without any of all these limitations and constraints we mentioned. So controlling full stack patient experience, provider, hospital, data driven prevention. So my dream is to be able to build that. And to do that, I still need to learn a lot in health care, and I need a few billion dollars.