The AI Technique Cutting Business Costs by 53%

The use of ML and OR allows you to go way further. So we did three projects where we had the best teams in the world that failed and really the best teams. We worked with Asin Toyota in Japan and they really tried for two years with the best teams in the world to solve a logistic problem. It's a logistic problem but again it can be anything

simple structures but it becomes far more complex when you have more and more variables coming in and it becomes quickly too difficult for a human. M so with O you are able to take decisions but at the same time with ML you're able to apply those decisions in the real world.

Is AI hype or help? Good day and welcome to the show.

I'm your host Pete Cooper and our topic is AI hype or help. Today's guest is Nikolai Vanm with his company Fun Art Tech. He has a slightly different angle on AI. He's working in AI and a complementaryary science called operations research. Let's find out what that's all about. Let's get into it. Nikolai, welcome. How are you doing today?

Thank you. I'm doing well

and thank you very much for inviting me.

I hope you are well too.

Yes, I'm well. Thanks. Thanks for asking. Um I'm I'm great. Um look, I had a look at your background. We we've met online.

You have a fabulous and deep background but just for the audience tell us uh what you do what is your business and and what do you do

so I consider myself as a scientist I have a business since 2017 uh I have a small startup in AI from Montreal

we are specialized in doing AI in a certain way

we have had some successes uh in big projects and basically we try to advocate for the combination of machine learning and operations research and actually other stuff too. But for sure machine learning, operations research, we combine them all the time to solve quite complex industrial problems.

Yes, operations research, we're going to get into that. And machine learning.

So, a lot of people are familiar with the idea of AI and they they open up their chatbot and so on. Um what's machine learning and how does that relate to AI?

So the dominant AI because we are talking about the AI actually there are several ways of doing AI.

The dominant way of doing AI is probably generative AI which is a part of machine learning

which is broader

but I would say that there are different ways of doing AI not only machine learning you have operations research I think it's AI some people would argue that it's not at the same time machine learning was considered 15 years ago as a sub field of o

so I find it funny that in some way you have the mother ship if I can say so operations research that is not AI for certain people but machine learning that comes from or and this was undisputed 15 years ago is AI actually you have problems you have domains uh science that can solve those problems call it AI or something else but

for me there are several ways of doing AI

operations research is one of them uh machine learning is the big trend today but actually they're really complimentary So I think of it as terms of AI as being the big big banner and then machine learning is part of AI and correct me if I'm wrong but I think of machine learning as a kind of a very very vast neural network and pattern recognition way of getting I guess outputs out of a machine in a different technique. The old the old way of software development was you would program something to do something like a mathematical qu question like 1 plus 1 equals in the box. You go go and it gives you a number out. So that's that's scripted.

But if you keep giving it lots and lots of mathematical um results, the machine learning will figure out what the the operator is, what the plus is and give you a result. So when you give it the next one, it's giving you an accurate result. How does that stand up with your understanding of machine learning?

So I have a version that is not really accepted. Um I would say that yes there are some differences and yes machine learning basically the paradigm is you have a model with parameters and then you are training training the model to get the right values of those parameters and then you can use that trained model to infer solutions.

M

like for instance you have a model to recognize cats pictures and recognize if on those pictures you have a cat or not.

So you have basically a model with parameters you are training that model so you find the right values for that model that correspond to the detection of a cat on pictures and some people say it's a completely different way of programming which is true and not true at the same time. Uh first of all we are using the same languages. We are still using C++, Python, uh Java, whatever. It's the same language. So it's still [clears throat] for me if then else.

Mhm.

Of course, they are some differences, but I would not look at the differences.

I would rather say, okay, it's a paradigm that works in this way. It's a bit different than the other paradigms, but I wouldn't say it's completely separated from the rest. M

you're talking about using a box and putting the plus in the model. This is what OR is doing.

Remember ML is coming from O. So yes, there are differences. They are not the same. It's not the same paradigm. But I would more look at the similarities

because there are lots and at the complimentarities because they're really complimentary. So,

so yes,

let's talk a little bit about operations research because that's a new term that um that you I don't hear this used very much.

I think of operations as a company has operations and it's generally okay the order comes in and somebody processes the order and then they ship something out the back. That's operations, right? It's it's the kind of mechanisms within a business or organization. Does that lead us into operations research at all?

So, it's probably the main target, but I would say it's a toolbox. It's a huge toolbox.

Mhm.

And some people consider ML as still belonging to that toolbox.

Not many people, but again, ML comes from O and ML is using that toolbox or our algorithms. Um, so it's the main target operations and how you can optimize. But I would say it's more than that. It's a new paradigm.

With ML, you can solve lots of problems. With OR, you can solve the same problems. You will do this differently. And actually, both have advantages and weaknesses.

So maybe you could give us an example of operations research like something [snorts] that would be relatable to

Yes. So operations research is heavily used for logistics for instance.

So if you want to optimize let's say the routes. So you have some deliveries, you need to deliver packages.

Um, and you want to do this by reducing the cost or the pollution or both.

You can use operations research. And I would say if you want to optimize something probably you should look into what operations research can offer you.

Because basically it's the science of optimization.

Mhm. And the link between ML and O in broad terms is that ML is using those optimization tool to optimize the training or the infra.

Let's extend this example of logistics. Somebody

yes

has to go from the Amazon uh warehouse and he has to deliver a 100 packages uh that are in the back of his truck and they're all just dumped in the back of the truck. Uh he could go to the back of the truck and he could pull the first one out and drive there, right? And then go to the next one. But he's going to be all over town, right? He's going to be wasting a lot of time.

But if somebody's clever, they'll take all of those and they'll group them all together and they'll get a path that that's the most optimum. So that he literally does the one that's closest and then the next one and the next one. And oh, by the way, there's a traffic jam or there's a shortcut, then he can do two over here. I mean, there's quite a few variables in there, but in simple terms, he could do like one big loop and get them all done. Yes,

much more efficiently.

Much more efficiently. And the thing is with O you're looking at global solutions. So not an individual routes but really global solutions for the whole city,

the whole fleet,

all the deliveries.

Oh yes.

So it's a paradigm where you can have a look on global stuff while machine learning is more local.

There is a notion of locality for machine learning.

So taking that the operations research one step up. So, we talked about one driver with 100 packages. What about if you have a hundred drivers with a 100 packages and servicing like a whole city?

Now, you're getting into more complex situations. You need to be able to group them together and you need, you know, to make sure they're shuffled correctly between them. You've got some guy going east and some guy going north.

And, you know, uh in simple ter, you know, I can think of simple structures, but it becomes far more complex when you have more and more variables coming in. And it becomes quickly too difficult for a human, right? You couldn't say a human could do that every day. But with um with operations, research, uh machine learning and AI uh techniques all lumping all them all together, you could just basically press a button and gives you a result, right?

Hopefully you can do that. But for sure it can help. In practice, we see an optimization of 20 to 40%. So you are doing something 20 or 40% better. This means that for instance you can reduce the cost by 20% or 40%.

Which is huge

or you can deliver uh on time 20 to 40% more than if you are doing this manually. And the thing is you have some instances that are so big no human can solve that.

It's impossible.

You need a computer.

Yeah. I mean, Amazon's probably well known throughout the world now.

Yes.

They must have invested a lot into operations research.

Yes, indeed. And machine learning, too.

Yeah. And the result for that, if anybody's not if anybody's well, a lot of people are younger than me, but I can remember that you didn't really get much stuff posted to you. It was too expensive. Yes. Something special from overseas from a family member. That might happen.

Most of the time everything was done by going to the shops, to the supermarkets, to the various stores. Now, I mean, I'm don't think I'm uncommon. I'm a a significant percentage of the things that I want, I'm ordering and they're they're available to me the very next day. I ordered some things yesterday and they're already here. Sometimes even ordering in the morning and it's there in the afternoon.

Yes.

I mean, there's an element of prepositioning too in this which we haven't introduced like certain products. Uh, I'm in the Lehigh Valley, so there's a big um there's a big hub here, so that makes it a little easier um to get stuff quickly.

But wow, what a lot of uh what a lot of variables as you say. I mean, even a team of people couldn't couldn't possibly do it.

Well, we're talking about millions of variables.

Millions of variables.

Yes.

Millions of variables.

And then you can combine them in all different kind of ways. Mhm.

So basically you have a search space that correspond to that which has more solution or parts of solution than the number of atoms in the universe.

Really?

Oh yes. We're talking about huge instances.

It's really really big.

More possibilities.

Yes. Than the number of atoms

than atoms in the universe.

Yes.

That's a big number.

Yes. That's a big number. Yes.

[gasps]

Yeah, that that boggles the mind. I mean, you think how could you have a computer that could even manage that? I mean, because if you if that's atoms in the universe,

well, the idea is not to use brute force because if you use brute force, then you cannot

okay,

visit them all. So you need to be clever and this is where AI comes in because you are able to search and to say okay that's a part of my search space that I don't need to search because I already know because I have some ideas about it that it's not worth it or I just leave it I go somewhere else and maybe I will come back.

Yes. Yes.

Well, I I'm familiar with this uh very much from engineering because I studied engineering and they would often the situation would become so complex that the mathematics would speak you're trying to solve a ma like engineering problem the mathematics would be so vast you couldn't possibly solve it.

So what they would often do is they'd say well that's negligible and they would just with a stroke of the chalk on the chalkboard they'd say that's very small like it's going to zero strike that off

and you can often do that and get a very very good engineering estimation very simply by striking off all of the things that have very little effect

and there are two different things because you have a mathematical problem and then you have reality. So you model reality but you know that your model is not a perfect copy of reality.

So it doesn't make sense to have the best mathematical solution.

What you [clears throat] want is to construct a solution that you can use in practice in reality.

So you don't need to find an optimal solution. You find a very good solution that corresponds to the problem you try to solve and that can help you.

Uh that's the goal.

So back to our example of these guys driving around their trucks. Maybe the optimal solution means they'll they'll get the 100 packages done in in seven hours and 50 59 minutes, but then if I if you just make it a little easier, they get it in eight hours.

Yes. And also you might have some unknowns. Maybe there's an accident.

You probably didn't plan that in your model.

You just said, "Okay, everybody is driving

and there's no problem. " Maybe there is congestion that you didn't foresee.

Maybe one of the driver fell ill.

So what you really are looking for is a solution that you can use in practice.

Mhm. Mhm.

That's the important stuff.

Yeah.

So you don't care too much about the mathematical complexity.

Okay. So got an understanding of what operations research is and machine learning and AI. We covered those. So how do you help organizations with that? But I mean are you are your clients logistics companies or is there other areas that we need to we need operations research?

So first of all we are not doing only operations research. Actually we never do only operations research.

We are always combining operations research with ML and then depending on the project with other science.

Mhm.

It's not only for logistics. So logistics is really our forte because we did several huge project in logistics where we have successfully found great solutions

but actually you can apply this to everything.

We could construct a chatbot or we had a project where we try to insert emotions between robots and humans.

Are you trying Yes.

In such a way that the discussion between the two would be easier

or we did some computer vision projects and actually we did a computer vision project where from time to time you had the cameras that went off.

Mhm.

You didn't have images.

Mhm.

And most of the time the images were blurry so they were of very bad quality because if you're using ML you need very good quality data. You need lots of data and you need a lot of good quality data.

When you're combining ML and OR, basically you're combining with ML data and with OR knowledge.

So how Okay, so this was interesting because I've had a long history in machine vision and cameras and things. Yes.

Sometimes, especially outdoor cameras, they can have a bit of water on the lens.

Yes.

Uh and that'll blur the image. So, how does the work you do um help with that problem? Like like here's an example of an outdoor camera which uh has got a bit of occlusion because of maybe the sun or the rain is on the lens and it's supposed to still be detecting people coming in and out. How can you help with that problem?

Well, you have some knowledge about the fact that you are monitoring people. So, very stupid. You know that most people have two legs, a head, a torso. Uh they're moving in a certain way.

So if you have occlusion,

you know that if someone is coming behind someone and is moving, probably you'll find it back.

Mhm.

It's very stupid. But I mean what I'm saying holds for every problem and every complexity you can imagine because you can have the same uh way of thinking about problems. M

so knowing that you have two person going this way you know that the person will not disappear. So you know in advance that probably you'll find the other person

after I don't know a certain amount of seconds. This is knowledge and with that knowledge you are able to construct better solutions.

Yeah.

And not only in terms of quality of solutions but also and this is something for which I'm really a strong advocate. You need less data and you need less power. So for instance, we did a computer vision uh project where basically you didn't have access to the cloud.

Mhm.

So there was no use of using deep learning with huge models.

You had an edge computer. Basically you have a very small computer nano.

Mhm.

Which is powerful but it's not that powerful. Mhm.

And the thing is we knew about the objects we were monitoring that they were behaving in a certain way. So we use that knowledge. Actually I can tell you more. It's about sausages and detecting sausages on the grill.

Oh, right.

And it's very very basic, but basically you know that sausages cannot fly. So if you detect that the sausage is not at its place where it was before, for instance, because the camera went off because of the heat.

Mhm. Well, you know that someone took the sausage, so that's knowledge.

Something else.

I was going to say, you know, there has been times when I and I big for one for the barbecue where I've dropped a sausage and and I guess as it's coming through the air, it's flying through the air to the ground.

Yeah, you have edge cases where you can say, okay, the sausage was dropped and then someone put it back, of course.

But I mean basically you know and it was in the US and so you know the length of the grill and you know that in the US people don't like to touch each other and you know also that basically they have a bag in one hand and they're picking the sausage with the the other hand.

So basically you know that at any time most likely you only have two hands.

The other thing is you know that you want to monitor which sausages were taken off but you know that you need a human hand. So instead of looking at the sausage, maybe you can look and this is what we did at the hands.

So you could start by detecting the hands. And actually you don't need to detect the hands, you need to detect movement.

But detecting movement, not the fly. A hand

doesn't need much data or much energy.

You can have very basic algorithms for that.

So what I mean what were you trying to do with that um algorithm? Were you trying to detect the number of sausages on the grill?

Uh, it's a little bit more complex than that. So, we needed Well, first of all, the real deal, the real project was to be able to train the whole stuff in less hours than a known team. Um, I mean, the project was handled from another team. They failed and their training was taking long hours.

So our mandate was not to change their algorithms but it was to change the training time

because they discovered that they you had to retrain each time you were implementing

uh the whole project into a supermarket.

So that was the goal. But

you're in a supermarket, there's sausages on the grill, people are coming to buy them and you want to count how many get

get purchased.

So the idea is to detect if the sausage is well cooked or not. That's okay.

The other thing is to detect what are the best selling sausages during certain days, certain hours, things like that,

right?

And to recognize the sausages because there are 80 different types of sausages.

Okay?

So rather than a ma a person having to manually note down that or type it in or use a scanner or something, you're going to use machine vision to solve that problem.

You want to automate this.

Yeah. Okay. Interesting. Well, uh, my background is also in, uh, edge compute. Um, we have a problem, well, we had a problem with our product of detecting vehicles coming up behind people on bicycles, um, because it's a hazard, right? And and sometimes you might not see them. Um, and we were getting the the edge compute to run at 15 frames per second. And it was pretty much brute force, I think, in your example, because we would just take one frame.

We would detect all the vehicles, put bounding boxes around them, and then put that onto the display, and then do it again and again, like 15 times a second. There was nothing in there to overlay that with any logic or any physics or any assumptions about the real world which could have been really handy as as you say like if the sun happens to be really low, you got a glimpse of that and then it disappears and then it comes back again. Well, cars don't disappear so you can fill in the blanks a little bit. Um, we would do a little bit of averaging, but we never did anything more sophisticated than that.

So if we lost the frame and all the frame was out over here, it would sort of average it out um and it would so it would be a smooth representation of what was coming up behind you. But that's interesting. I mean it's it's I've spoke to a few other people about you know the frontiers of of AI and one of the big limitations at the moment is that it's it's just a huge pattern recognition um uh LLMs. I'm talking about a huge pattern recognition uh model

and uh it doesn't really know what the real world is and doesn't and it can't in any way think for itself.

So it can make some pretty wild assumptions about what it's actually seeing and doing. Um as you said sausages don't fly.

So you know you won't have one just coming randomly in from nowhere. It'll have to be held. It'll have to be on a surface. If it's not falling it'll have to be held. Right.

Yes. Okay.

So, it's melding this real world physics and the real world I mean you know we do this don't we as humans right we we sort of have this I don't know how it works maybe you can tell me but I understand how a neural network works and that's essentially how you know AI is at the moment and LLM's the pattern recognition but we also seem to have this ability to understand physics and the real world which is much more than these models are doing

well there's some combination of knowledge data with neural networks,

right?

But the idea is really to say, okay, I have an open mind and I know what tools exist and I know them them on a fundamental level.

I know what they're capable of or not and I know how I could combine them together.

Uh see, if you use operations research, you're able to model concepts, but actually you can do this with ML2. So you really need to understand what kind of concepts you want to model.

But one thing that I really like for instance is the fact that with you can unlearn very quickly.

So this is a open problem in ML because people don't know exactly how you can unlearn

because basically learning or unlearning would entail learning more but learning in such a way that you have unlearned what you already have learned. I tell you what, I know all about unlearning. My golf is is terrible.

Um I'll openly admit that. And every time I go and have a lesson, I have to unlearn all the bad habits that I've accured acrewed over the years and try to put a new habit in there and half of the battle if not 80% of the battle is unlearning bad habits.

But you see with operations research this is something that you can do quite easily. I mean nothing is easy but there are some tools that you can use

so that you can unlearn and you can unlearn very quickly

and this is extremely important. If I take your example of looking at the occlusions of the cyclist and the cars,

if you don't do an inference in real time, if for instance you say, okay, I have one hour of video.

What you could do is okay, you can try to infer what happens. But the more you try, maybe you'll have more information and that more information tells you more about what happened before.

But if you only use ML,

it's over. You have learned

and that's it.

With a certain type of combinations of ML and R, you can say, "Oh, I'm going to backtrack. "

Now I know more because maybe you didn't see the car or the bicycle or whatever. And then you see it again later. Then you can say, "Oh, now that I see that piece of information,

I [clears throat] know that I should have seen it before.

Let's backtrack and start all over again.

" But

if you need to retrain heavily,

nobody is doing that. But with certain combination of or ML, you can do this very quickly.

Right? So what's the sort of schema for this type of thing? I mean the way we implemented um our solutions, it's like a black box. As I said, we we keep training then the model and then we just deploy it and then we'd have the cameras just pushing 15 frames a second in and then we'd be getting our bounding box data out. What do we how how would we integrate the O concepts into that?

Well, it really depends on what you want to achieve. What is your goal?

Um what I really like with the combination of LM ML and R is that you get some robust solution but really robust solutions

in the sense that if you want a system that tells you okay this is a solution and I know 100% that this is a solution

and this is a context I don't know what to do with it

so I cannot give you a solution. This is something very difficult to do with ML only. ML will give you a solution no matter what. Because in ML,

let's say I'm trying to solve the problem of the the sun, the sun coming in and oluding the camera so that I'm not always seeing the car clearly because, you know, for maybe fractions of a second that it's just too bad.

Well, how do I implement O to fix that problem? Well, you could say that in order you have constraints which is extremely powerful because those constraints should be satisfied or satisfied up to a point. I mean you have lots of possibilities.

But you could say that okay I detected a car. It means I need to find the car back in the other frames.

Mhm.

Again if you can go back.

Mhm.

So you could then add a constraint saying that car must be there.

Okay. So this kind of retrospective thing.

So

not only it goes way further than that.

Yeah. But I just I'm just trying to solve this particular problem so we can get a real handle on it.

Um the data is constantly streaming one way through the black box. Reconstructing it retrospectively doesn't really help the rider because they are they're just seeing it on in real time. Uh there would be other applications perhaps if I wanted to track a car and then send it to the cloud to say that was what happened. Then I can see the retrospective side of things working. Yeah.

Okay. Gotcha.

Well, for instance, when you are trying to go frame by frame, you're doing a kind of matching

between the frames. You say, "Okay, I had one car or two cars and then the next frame I should find them back. " So, it's a kind of matching.

I don't know if you know but there is a matching theory in O

that you can use for that for instance.

Oh okay. So not necessarily retrospectively but if we were seeing the car coming behind us on the bike the car ML's doing its job and then the sun gets really bad. We can say well the car can't disappear so it should be there right? So we still show the icon it doesn't disappear and then you know maybe it appears again later. Okay. So, it fills in the blank in real time, too. Okay.

So, for me, it's difficult to answer your question because I don't know exactly what your goal is and I don't know exactly what the context is of your project.

If you want, we can talk about this uh later.

Yeah, we can talk about something else that's closer to your uh experience level if you like. I mean, I don't mind.

Sure.

The use of ML and O allows you to go way further. So we did three projects where we had the best teams in the world that failed and really the best teams.

So we worked with Asin Toyota in Japan and they really tried for two years with the best teams in the world to solve a logistic problem. It's a logistic problem but again it can be anything.

Mhm.

Um the main problem was the size of the instance. The problem was huge and so you couldn't handle this with

what was the problem you were trying to solve.

Now the problem it's a logistic problem where you are transporting automobile pieces. So they have factories.

Yeah.

And they need to transport those automobile pieces between the factories and it's not a simple pizza delivery in the sense that you have some pieces that needs to be at one factory at the same time and things like that.

Mhm.

But again the real difficulty is not the problem in itself that is quite known. It's really the size the number of automobile pieces they are moving every day. Mhm.

It's huge.

Mhm.

And so you need a way to be able to handle the whole supply chain at once.

Mhm.

And this we could do it because we had that hybridization of ML and O because O only wouldn't work. ML only I don't see how we could do that. So is this related to this the the just in time uh manufacturing approach that that's very that came out of Japan? Um is that why you have this this need? It was not just in time for this case but I mean it can be applied for just in time. No here they are producing lots of car and they know that their machine constructing cars is not going to change and they are not doing just in time for that because they know in advance that they're producing thousands and thousands of cars.

M

but you could apply the combination of ML and R to anything just in time or not. But it was not just in time

in that project.

It was really how can we take the whole process at once. Remember I told you that with

I'm trying to get to

Yes.

what their business was trying to achieve with using O. What was the outcome? So basically they didn't care about what we used. So they tried for two years of the use.

Yeah. We're going back to the use. What is what is it as a business? Were they trying to get more production efficiency? Were they trying to lower the

the cost of storage? Um no. What were

the idea is to lower the cost.

So I can say this publicly because it was a public contest. So it cost them 500 US millions per year.

Mhm. And the idea is how can we reduce the cost?

Mhm.

So reduce the cost by 53%.

This is something like 250 uh US million dollars.

So reduce the cost of moving parts between factories. Okay.

Yes. Because you're able to decide what type of truck you will use, when you will use it, on what road, with uh which driver, how will you put the pieces inside the trucks and things like that.

And actually, it goes even further than that because you're able to say, "Okay, I have my um factories.

Maybe I can add some intermediate depots where I can get my trucks and they will unload and reload and then you can reload even better. So you even reduce more the cost.

But basically the goal is how can I reduce my cost.

So just so that people understand because not everyone knows how cars are built. There's different factories perhaps around multiple cities that are producing bits and pieces for Toyota's cars. Um, and they may or may not be owned by Toyota. They may be outsourced. [snorts] Now, I would imagine that being a Japanese company, just in time is pretty important.

And if anyone doesn't know what that means, that means that you you only have the stock that you need at hand for the manufacturing you're doing at that time, just in time. So, you don't have a big storage facility. That that's the thing because storage is very expensive. You you you got to pay for it, you got to hold it, you've got to house it, you got to pay your vendors, all this. It's much better just to have it and use it straight away. And so this to me what you're saying is the O supports just in time.

Yes.

But makes it work much better than um than than I mean just in time was in the 70s I think that they

that they come out with this.

But basically it's not O.

It's really the combination of O and ML. M

so with O you are able to take decisions but at the same time with ML you're able to apply those decisions in the real world.

Yeah.

And basically there is a difference between the predictive and the prescriptive paradigm. So the predictive is the one almost everybody is using. You're trying to predict.

But let's say you are predicting three scenarios A, B, and C. A has a probability of 95%. So you're almost sure that this is going to happen.

But then you have B that is happening maybe with a 4% chance and then you have C 1%.

So if you are in the predictive world you could say okay 90% I mean 95% of the time something like that uh I'll have scenario A done except that maybe scenario C is a catastrophe. You really really don't want this to happen. M

you cannot just look at your predictions.

You need something more. You need to be able to take a decision

because you will not act the same way if you know that C cannot happen. Let's say C is the end of the world.

So that C that only happens 1% of the time becomes maybe more important than scenario A that happens 95% of the time. M

so you need some tools to take decisions and this is where O can help a lot.

You have tools in O that deals with unknowns unknown unknowns and that helps you take the best decisions with what you know or don't know.

No interesting

Mel can help you too but I mean there are different tools and actually the idea is to combine them.

Okay. Oh, that sounds uh and I think you said to me that you were quite successful with that project. Is that right?

Uh the one with Toyota? Well, we got 52% of cost reduction.

Um and actually until now we did 13 projects

and we have a 100% success rate scientifically meaning that we did scientifically we managed to really find a solution that was working for the customer.

M

I'm saying this because there's a difference between the science and doing the algorithmic stuff and then the industrial uh business um I would say side of stuff which is way more complex because as a small team when you are in such big projects you have politics and so on and there it's very difficult for us to be able to have our say. So basically what we do is to open the door. We are called to say okay what would you do and then we say okay this is one way of doing it and we show the way and then I would consider that as a scientific success.

And then you have the uh the other side that is a little bit more difficult for us as a small team.

So how is this all changing with with the boom in AI at the moment? I mean, it sounds like operations research isn't isn't something that new. Tell me if I'm wrong. It sounds like it's something that's been around for a while, but now AI has far more capabilities. Um, and is that changing operations research?

Oh, yes. So, operations research basically exists since the Second World War. M

basically you had questions like how can you optimize the way to kill people or not to get killed

and this is basically how operations research started.

In the 80s operations research was really big.

Mhm.

And then since 2000 or even the '90s ML is becoming what it is now.

And actually there is a kind of I I don't know if I can say war between the two but basically there is some kind of competition between the two and ML is clearly winning and unfortunately people in ML say okay forget about the war we can replace them

maybe one day they will be able to do that

but today I don't think so.

Mhm. Okay.

And as I said they're really complimentary. So the idea is not to replace one with the other doesn't make any sense. They are both beautiful science with advantages and weaknesses.

So then your business does actually combine the two, right?

Yes. All the time. We never ever do a project only with o or only with ML.

For us, it doesn't make any sense.

Yeah. Yeah. Okay. So I guess you think that perhaps AI could overtake O and and and do without it. So this is only a temporary solution.

I don't believe in this but you never know.

What I know is that what people claim when you have people telling you, okay, we are basically doing everything with ML and we can do everything. This is

Maybe one day, you never know,

maybe one day it will be the case. I don't believe in this.

Uh and certainly not for today. Today it's not true. There are certain problem you can solve with ML but you get very poor solutions

with O you're able to optimize.

You are able to take some decisions and you cannot do that with ML.

The other thing is one of the reason why we created the company is that we advocate for a frugal AI. So for us frugality is extremely important. So basically in ML you need it's a kind of brute force. You need lots of data, lots of data of good quality.

The process of training is most of the time very costly.

Mhm.

When you combine both, you can reduce that cost. You need less data. There are lots of problems where you don't have data. You have very few data or the data is of very bad quality.

If you only do an ML approach, you're stuck. It's over.

M

so the combination of both lets you still do something meaningful

even if you don't have data and when I'm talking about frugality when I'm talking about the project we did with AC Toyota and they were really surprised we're talking about the whole supply chain for the transportation for all Toyota it's huge guessing how many minutes and on what type of computer we could solve that problem

four minutes on a laptop

4 But it's on a laptop, right?

Not a data center, not a huge computer, a laptop. Four minutes.

Mhm.

For the whole supply chain.

And And that was using But that was still using ML.

Oh, yes. No, no, we needed ML. Without ML, we couldn't have done it.

Yeah.

No, no, we need ML.

You got a model running on the laptop?

Yes. The whole thing runs on the laptop in four minutes for the whole supply chain.

Cool. Yeah. Yeah. So I think there's a sort of uh the average person would say that AI is cheap. In other words, I'm I'm getting a ChatGPT a free version and I'm getting my you know I can generate images and I can do research. Most people I think would say that's cheap.

And even if you get a subscription though, you know, it's still not a lot of money. So when you say it's expensive, relate that to to us. How is it expensive?

It's extremely expensive.

It's cheap for the moment but I don't see how we can keep it that way

because the real cost is not passed on the consumers

for the moment there is a kind of arm race in AI because AI is extremely important uh even for the domination of the world

y

and so you don't pay the exact cost so there is a cost in terms of data

and these people are a little bit aware of this because you have all those lawsuits uh because basically they stole everything that was accessible

because they needed lots of data

but you also need a lot of energy. So the LLM when you are training them it cost millions millions and even when you do the inference it still cost millions.

You need to pay for the electricity. Electricity is the worst part because if you have access to cheap electricity it's still okay. So, North America, Asia, it's located Europe. It's not at all. Electricity cost a lot in Europe.

It's a huge deal.

So, this is interesting. So maybe we don't have any numbers we could put to the table right now, but it sounds to me like that the investment as the big I mean there's half a dozen really big players out there now trying to um get their share of the trying to get market share and they're essentially buying their way in by racing ahead with the best models.

It's a this arms race as you call it but that's essentially investors money going into that. Yes.

It's not the customers.

Public money. Lots of public money.

Lots of public money, but also like it's not me who's using ChatGPT or Claude. I'm not paying that. I'm only playing a small piece of that total cost,

right? So, we I mean, I don't know how long that'll last. I mean, it's an AI boom. Um, and I I would advise people to, you know, use it now whilst it's cheap. Um,

yes. But it will change. I'm aware of that as well. Um I I what I think will happen and and tell me if you think I'm wrong.

I don't think it'll get more expensive, but I think we'll have other things like advertising and other other things that get in the way of us accessing it um to pay for it. That [snorts] that's what I think will happen. I don't know what you you think.

It's already happening in a way. And there are some techno optimists that thinks that we'll find a way to get access to cheap energy like fusion or something like that might happen. But if not,

it's almost certain that we cannot continue like that.

Yeah,

that was interesting what you said about North America and and Asia has cheap energy, but Europe does not. I've not heard.

Well, it depends where in Asia because Asia is big.

But I mean, there are some places where energy doesn't cost much,

especially China,

and in Europe, the bill is huge. But it's not only a question of money, it's also a question of pollution

because producing that energy is producing a lot of pollution. M yeah.

So now we we estimate that four or 5% of the energy consumation is for the data center for AI worldwide.

Wow. I didn't know it was that high.

That's huge.

Yeah. Four to 5% of the CO2 emissions is for is is near.

Not the CO2 emission, but I mean the the quantity of energy, the way it's it's spent. Yes.

Mhm. And it's likely to become bigger and bigger.

Oh yeah. I I I can't see how we can slow this down.

I I mean I I I'm concerned as well. What this seems a little bit like out of control, but you know what can we do? I mean, can we live without AI now? It's here. You know, the genie's out of the bottle. I mean, it's happening.

Yes. But

and it's going to consume our energy. And to stay relevant, we have to adopt AI, you know. So,

yeah, but there are some physical limits that we will hit. And whether we like it or not, we have a physical world that is a closed world and it's finite.

So, we cannot I mean now we are talking about constructing data center with nuclear powerpoint. Okay. We're talking about sending those data center out into space.

Y

um are these really good solutions

even more when you know that there is another way of doing AI

that is way less uh demanding for the energy or the data.

Yeah. And that's the the operations research.

Not only there are other I [clears throat] mean there are several ways of doing AI. I mean ML is one of them but there are several ways of doing things

depending on the problems. Uh recently there's an article that came out about quantum computing for certain type of problems and basically they didn't need much energy at all but again it's for one very precise type of problems.

Again people in quantum computing tell you that they will solve everything.

Again this is

Like people in or will tell you we can solve everything. No that's No one can do that. M

today there is not a single approach that is better than the all the others. We really should combine them and then we should take some decision about what we want to do or not as a humanity because AI can help or can also be against us.

That's also another problem. And basically you have two camps. You have the techno optimistic and then you have the technopessimistics.

Uh I'm in between. And in a way both are right and wrong

because one thing that people should understand is that AI is more than a tool. It's becoming something that will be more clever than us.

If you think about the caterpillar, there's no problem. Everybody understands and accepts that it's stronger than any human being. The strength of a caterpillar is way more than whatever human being. But now we are constricting something, a tool for the moment. That

You mean the the insect, the caterpillar?

No, I'm talking about the the huge crane.

Oh, a crane.

I'm sorry. Yes.

The one that can uh take huge rocks that weigh tons.

Um, and now we are constructing something that actually will become more clever than us. So, some people think it's impossible. I think it will come and it will come very soon.

And what do you do when you have a tool that can manipulate you because it's cle more more clever than you?

Oh yeah.

So there are some decisions we should take but probably we will never take them because there are few people around the world taking those decisions

and it's not humanity in its entirety.

No it's not.

No it's not

Nikolai. We we're we're getting to the uh to the end of our time. Do you have any questions for me before I give you our wrap-up question?

I do have some questions, but maybe not in front of the camera.

No, I'll stick around. We'll talk afterwards. All right. Well, I'll give you the final question, Nikolai. Is is AI hype or help?

It's both. And I would even add is it hindrance?

It can be of tremendous help for humanity, for nature, human beings, society, companies, but it's really a question of choice. It can also destroy us literally. I know that people again are in two and they don't talk to each other and they don't see that there is a middle ground and that that middle ground can be terrific or catastrophic.

Everything is possible. We don't know where we're going. What we know is we're going there fast and we're going there without thinking.

That's a huge problem.

Yeah. Yeah. Well, the AI is doing the thinking for us, isn't it? All right, Nick. Nicola, thank you. Thank you so much.

I I'll stick around after the show. We'll have a chat. Um, but thanks thanks for coming on.

Well, thanks for inviting me.

Okay, bye-bye.

Bye.

Well, that was an interesting deep dive into operations research, which in simple terms is the science behind logistics, how to make uh you know the movement of trucks and ships and packages very very efficient. And that um is complimentary to machine learning. So that's part of it. But also there's a decision side of things, a logistical side or a logic side of things.

We talked also about how that can be very efficient and that can help with one of the big problems we're facing right now which is the huge energy usage of data centers and AI. Up to 5% of the total energy is being used uh in this field of AI and operations research can significantly reduce the energy consumptions. So very interesting field. If you want to know more about Nikolai's business, please u follow the links in the description. If you want to know more about Skillion and AI labs and what we do, also do the same. I hope that helps and thanks for watching.