00:00:07 Supply chain science and the creation of numerical recipes.

00:03:21 The difference between algorithms and numerical recipes.

00:05:21 Explanation of how numerical recipes are better suited for solving muddy problems in supply chains compared to algorithms.

00:06:00 Discussion of how algorithms are present in software companies and the danger of having a distorted vision of real-world problems.

00:07:48 Comparison of the optimization of a single screw in a machine to the big problem of supply chains.

00:08:02 Discussion of the importance of numerical recipes in solving supply chain problems.

00:08:54 Comparison of algorithms and numerical recipes in terms of objectivity.

00:09:44 Explanation of how the subjectivity of numerical recipes makes the expertise of a supply chain scientist crucial.

00:13:02 The importance of aligning the solution with the problem and minimizing the potential for errors.

00:15:52 Discussion of the need for processes and tooling to prevent mistakes and improve the quality of the solution.

00:17:16 Explanation of the problems that can arise with numerical recipes.

00:18:07 Discussion of how companies in the supply chain industry operate through numerical recipes.

00:20:01 Criticism of the tooling being inadequate to solve supply chain problems.

00:22:00 The importance of numerical recipes in being approximately right and agile in supply chain problems.

### Summary

In an interview, Joannes Vermorel, the founder of Lokad, discusses the concept of numerical recipes in supply chain optimization. He argues that algorithms and machine learning can give a false impression of objectivity and well-defined boundaries between problems and solutions, and numerical recipes are a better approach to handle the complex and changing nature of real-world supply chain problems. Vermorel emphasizes the importance of having alignment, correctness by design, and good tooling to prevent mistakes and ensure success in supply chain optimization. He believes that numerical recipes are essential for success in the unpredictable world of supply chains.

### Extended Summary

In this interview, Kieran Chandler and Joannes Vermorel, the founder of Lokad, discuss the concept of numerical recipes in supply chain optimization. Vermorel explains that he borrowed the term from a successful book called “Numerical Recipes” published in the 1980s, which offers a unique perspective on problem-solving.

He emphasizes that problem-solving in supply chain management is not as simple as having a clearly defined problem and solution. Instead, the type of solution used can shape the problem, with trade-offs and feedback loops existing between them. Vermorel believes that the term “numerical recipes” is a better descriptor for the approaches used in supply chain optimization because it acknowledges the inherent complexity and adaptability of these solutions.

Vermorel explains that algorithms, machine learning, and other terminology can give a false impression of objectivity and well-defined boundaries between problems and solutions. However, in practice, real-world supply chains present more complex, “muddy” situations. He contrasts the clarity of sorting algorithms, which have well-defined problem statements and mathematical properties, with the ambiguity of supply chain problems, which often involve negotiations, changing conditions, and other real-world factors.

For example, minimum order quantities (MOQs) in supply chains are not fixed like physical laws but rather are the result of negotiations with suppliers. If an MOQ proves to be problematic, a company might be able to negotiate a more favorable arrangement. A smart numerical recipe would capture these real-world options, making it a more suitable approach for addressing supply chain problems than traditional algorithms.

While Lokad does use many algorithms in its software stack, Vermorel argues that relying solely on algorithms can lead to a distorted understanding of real-world supply chain problems, especially for those with formal education in computer science or software engineering. This is because traditional algorithms are often better suited for clearly defined problems with definite outputs, whereas numerical recipes are more adaptable and better suited for the complex, changing nature of supply chains.

Vermorel believes that the concept of numerical recipes is a more appropriate way to describe the methods used in supply chain optimization due to their adaptability and ability to handle the complexity and ambiguity inherent in real-world supply chain problems. This approach recognizes the importance of trade-offs and feedback loops between problems and solutions and allows for a more nuanced understanding of supply chain management.

They discussed the challenges of supply chain optimization and the role of supply chain scientists in creating numerical recipes. Vermorel explains that despite decades of research, sorting algorithms for supply chain optimization still have pros and cons. He uses the metaphor of a complex machine, where even if a single component is optimized, it doesn’t guarantee the overall system’s efficiency.

Vermorel points out that real-world supply chain problems often require numerical recipes instead of strictly defined algorithms. These recipes are created by supply chain scientists, whose expertise plays a significant role in crafting solutions. Although algorithms are objective and rooted in mathematics, Vermorel acknowledges that subjectivity exists even in mathematics, with concepts like elegance influencing the perception of algorithms.

When it comes to numerical recipes, Vermorel argues that some aspects of reality are too complex to fit into a mathematical framework. Although advanced statistical methods can extract patterns from data, there are instances where judgment calls are necessary. For example, supply chain scientists must make decisions based on unique situations, which might not have prior examples in sales history. Vermorel likens this to the culinary arts, where chefs of varying skill levels create dishes that might be highly subjective yet still considered excellent or subpar.

Discussing the challenge of maintaining quality across different clients and industries, Vermorel acknowledges that there are multiple angles to consider. One key aspect is to ensure that the engineers do not betray the business, as they may be tempted to create formulas that appear sophisticated but do not address the underlying problem.

Vermorel discusses the importance of having alignment between the problem being solved and the quantitative modeling being applied, as well as a tooling that minimizes the amount of daily foot gunning. He stresses that having correctness by design is crucial to preventing terminal mistakes and ensuring that even when people are too tired to be smart, they can still make smart decisions. Vermorel also mentions that half of Lokad’s success comes from knowing how to roll out a quantitative supply chain initiative.

Vermorel highlights that companies in the supply chain industry operate through numerical recipes, but many are still stuck using classical algorithm-based approaches. He notes that while spreadsheets are the embodiment of the understanding of how to model the supply chain, they are not suitable for dealing with uncertainty or multi-echelon supply chains. Vermorel criticizes the tooling, saying that it is inadequate and highly subjective, and includes a lot of narrow numerical recipes. He believes that companies need to engineer plenty of processes to prevent numerical stability problems that can shut down a factory or warehouse.

Overall, Vermorel emphasizes the importance of having alignment, correctness by design, and good tooling to prevent mistakes and ensure success in supply chain optimization. He also highlights the limitations of spreadsheets and the need for better tooling to deal with uncertainty and multi-echelon supply chains.

He argues that present-day companies operate through numerical recipes but frequently face inadequate tooling and processes, such as silos. Vermorel believes that numerical recipes are here to stay and are the right mindset to have when supply chain problems are involved. He explains that numerical recipes are formulas that don’t have purity and are not like electromagnetic equations, which are incredibly pure and tight. Supply chains are complex and require hundreds of semi-accidental conditions and factors to make sense. Vermorel emphasizes the importance of having something that can be versatile like a recipe and can cope with changing conditions. He compares this to top chefs who can improvise and adapt to missing ingredients, short timelines, and changing constraints, but there is always a method to their madness. Vermorel explains that at Lokad, they cultivate a method to cope with the mess of supply chains. The main conclusion of the episode is that numerical recipes are essential because they embody the thinking that it’s better to be approximately right than exactly wrong, which is crucial in the unpredictable world of supply chains. In conclusion, Vermorel argues that having a versatile numerical recipe that copes with changing conditions and constraints is key to success in the supply chain industry.

### Full Transcript

**Kieran Chandler**: Hey, much like a top Michelin star chef, a supply chain scientist has to create recipes that adapt and evolve to every scenario. As such, today we’re going to investigate what it takes to create these recipes, and in particular, what characterizes the ones we use in our supply chains. So, Joannes, we’ve used this term “numerical recipes” a few times before. Why did you think it was important to revisit it?

**Joannes Vermorel**: This term, I stole it from people in the ’80s who wrote an incredibly successful book called “Numerical Recipes”. It emphasized a certain way to look at the problem. You see, there is this idea that usually you have a problem and you have a solution, but the reality is not as simple. The type of solution you have literally shapes the problem, and there is a back and forth between them. You have a trade-off in the way you want to approach your problem, depending on the way you approach your solution.

The key idea is that we want to deliver numerical results for companies who run actual supply chains. The problem with other terminology, like saying we use algorithms or machine learning, is that it emphasizes something completely objective and well-defined, where you have the problem and the solution, and then for the same problem, you can have competing solutions. But the reality is that when you want to deliver results for an actual supply chain, the whole thing is a lot more muddy. It’s a very accidental process with many bumps along the road. What you get in the end is a numerical recipe, which describes the chain of numerical computation to get the results.

**Kieran Chandler**: Why is it that something like an algorithm wouldn’t be appropriate to describe that? I mean, what is it that an algorithm misses out on?

**Joannes Vermorel**: I’m using the term “recipe” precisely to say this is not an algorithm. For those with a background in computer science or machine learning, you would have learned about algorithms in your textbooks and courses. Let’s take the archetype of the algorithm, the sorting algorithm. You have a collection of objects with an order relationship, and you can sort them using a well-defined series of steps. At the end, the collection is sorted, and your algorithm has properties such as memory consumption and complexity.

There is a variety of sorting algorithms with different properties. Some are deterministic, some are stochastic, and some are very good if the data is already partially sorted. But the thing is, when it comes to supply chain optimization, we need something more adaptable and flexible, like a numerical recipe, rather than a rigid algorithm.

**Kieran Chandler**: The sorting algorithm is a super clear-cut situation where you have a problem statement that is completely non-ambiguous. So you want to sort a collection of elements, given an order relationship, it has mathematical clarity to it. On the contrary, when you think about the sort of problems that you need to solve in actual supply chain situations, it’s very muddy. I mean, you have MOQs, but MOQs are not like the laws of physics; they are more like the result of a negotiation with your suppliers. So if numerically an MOQ proved to be really a problem, maybe you can actually make a phone call with a supplier and arrange something that is in between. So you see, it’s maybe a smart numerical recipe will capture this sort of option that exists in the real world, but suddenly it doesn’t have this kind of crystal-like purity to it.

**Joannes Vermorel**: Exactly. I mean, at Lokad, make no mistake, we use tons of algorithms, like every serious or semi-serious software company out there. The Lokad stack is literally a very long series of algorithms. Because the way we have engineered Lokad around a domain-specific programming language called Envision, our compiler is like an endless series of algorithms that transform the script itself into abstract representations, down to the series of execution for the compiled program that needs to be executed, etc. So, algorithms are all over the place.

The danger here is that, just like naïve reductionism, it’s not a danger to the uneducated audience. If you had the privilege of having never completed a master in computer science, or you’re not a software engineer professional, this is probably not the sort of problem that you will face. But the problem is that if you happen to be very educated in these things, what you’ve been taught in classes and what you’re reading in most computer science books give you a very distorted vision of what problems really look like for real supply chains.

Algorithms are very useful, and it’s good that Lokad can rely on a collection of sorting algorithms that have pros and cons, which are completely well understood thanks to decades of research that have laid out a comprehensive mapping of all the various dimensions for this tiny problem. But it’s just that, it’s like you have a very complex machine, and you achieve perfection for a tiny cog. So yes, if you look at one screw and you would say, “What is the optimal metal for the screw?” And because you have a problem that is so well defined, so narrow, you might have an answer which is, you need to use exactly this type of steel for this screw because it’s completely optimal with regard to all the constraints.

**Kieran Chandler**: So, Joannes, let’s talk about supply chain optimization. Is it really possible to find the optimal solution for a supply chain?

**Joannes Vermorel**: That’s a good question, Kieran. You can optimize some parts of your supply chain, but having the screw at the right place in the machine, for example, is not enough to solve the big problem. You need to consider every detail of your massive setup, and when you put everything together, it really makes sense. When you go into the real world to solve supply chain problems, you end up with numerical recipes rather than algorithms. The emphasis and attitude of the person crafting the thing are not the same.

**Kieran Chandler**: I see. Let’s talk a little bit more about the actual person that’s creating these numerical recipes. How reliant are you on their skills and expertise?

**Joannes Vermorel**: Quite a lot, actually. That’s something that should not be neglected. When you look at an algorithm, you would say it’s completely objective, a mathematical framework with proof and well-defined. Algorithms are a branch of mathematics, a pinnacle of objectivity. But subjectivity exists very much, even in mathematics. If we go to numerical recipes, the idea is to objectivize everything, but I believe it’s another bad case of naive rationalism. The reality is too complex to fit into any kind of mathematical framework we know of.

**Kieran Chandler**: I see what you mean. So, are there situations where you have to make judgment calls?

**Joannes Vermorel**: Yes, there are plenty of situations where you have to make judgment calls. For example, how do you deal with a discovered situation from a supply chain perspective when you don’t have prior examples in your sales history? At some point, you need to make a decision that takes into account this bizarre situation. There is no alternative but to have a smart supply chain scientist who has a good grasp of what is actually happening in the supply chain and make these judgment calls.

**Kieran Chandler**: So, I think there’s a judgment call about how those things should be reflected numerically in the system. And that’s just like, you know, this chef metaphor, at some point, you know, it’s not because the choice of the way you do your exact, you know, recipe is super, super highly subjective that at the end of the day, you don’t end up with, you know, a crappy chef on one side and an incredible, you know, other, I would say, chef of incredible talents on the other side. You know, it, even if you cannot define, you know, clear-cut rules that let you sort out which are the good ones, the bad ones, clearly the extremes exist nonetheless. And people that are, you know, educated to some extent, you know, they can make a judgment call about who is, you know, a great chef and who is a crappy chef. And the extremes are fairly obvious. And if you want to have all the nuance in between, you will probably need to have more skills yourself and be versed into, you know, the culinary art and cooking. But, you see, that’s, this is not, this is fairly rational to proceed like this. Okay, let’s stick with kind of the low-carb kitchen then.

**Joannes Vermorel**: Um, so I mean there are so many angles to the discussion. And first, you need to make sure that you don’t betray the business. The biggest danger when you put, you know, a smart engineer in front of a problem is that the engineer, you know, by training will always come up with a formula that looks like a very profound and very scientific. So, and again, you know, I believe that there is a saying that says that there is like a free path to ruins. The most enjoyable path is women, the fastest way to ruin is actually gambling, but the surest way, the surest way to ruin is to hire more engineers. So, first, you need to make sure that you have alignment in terms of vision between the problem being solved and all the sophistication, you know, in the quantitative modeling that is being applied. That’s the first thing. And, by the way, this is also why at Lokad we cultivate having a lot of materials on our website, on YouTube, in many places, is that we need to cultivate this understanding of the problems themselves. So, that’s the first thing, is alignment, you know, between the technique and the business. The second thing is you need to have a tooling that minimizes the amount of daily foot-gunning taking place. You know, foot-gunning is just you have a gun in your hand and you shoot your foot. And, literally, those things tend to happen over and over and over, especially when you start dealing with, I would say, fancy numerical recipes. What do I qualify as fancy? I mean, there are plenty of companies that say, “Oh, we use TensorFlow.” Yes, excellent. So now, you’ve just acquired 100 more ways to shoot a bullet in your feet.

**Kieran Chandler**: Okay, I’m gonna jump in on that, Joannes, because that’s a very good point that you bring up there. How do you minimize the amount of daily foot-gunning, because it seems that a lot of companies are out there buying a lot of guns to shoot themselves in the feet with?

**Joannes Vermorel**: Yeah, absolutely. And, you know, the thing is, I think there are different kinds of tools to approach this problem. But, one thing that is very important is to

**Kieran Chandler**: Some of those ways can be exceedingly creative and have a lot of surprises going on. So, business alignment first, and then you need to have tooling that, by design, gives you a high degree of correctness. Correctness by design is something that is very prevalent in terms of thinking at Lokad.

**Joannes Vermorel**: Although I’m a big believer in education, I believe that it’s best when, by design, people can be allowed to make mistakes. We hire smart people, but even smart people have bad days, or once in a while, they didn’t sleep very well. So, you want to have tooling that prevents you from making super dumb terminal mistakes, so it supports you to be smarter even when you’re too tired to be smart.

**Kieran Chandler**: And then maybe the third idea is that you need to engineer plenty of processes.

**Joannes Vermorel**: Yes, for example, at Lokad, I would say half of it is really about the know-how of how to roll out a quantitative supply chain initiative. When you say “rollout a quantitative supply chain initiative,” it means, for example, how do you end up with numerical recipes that don’t have terminal problems? When I say terminal, I mean something that would just kill the initiative because the problem is so big that people decide, rightfully, that actually killing off this initiative is the best way forward.

**Kieran Chandler**: So, what sort of problems could you have?

**Joannes Vermorel**: The numerical recipes can be bad in so many ways. It can be bad in terms of variance of compute time, where it’s way too erratic. So sometimes you run the thing, and it takes one hour, sometimes eight, and people are not exactly sure why. That’s a big problem. It can also be bad because it’s pretty opaque. This black box effect is very difficult to have something that is both numerically smart and not an immediate black box, including for the supply scientist themselves. You can also have numerical stability problems where, on average, your recipe is excellent, but in 0.1% of the situations, it’s downright insane. That creates a lot of operational problems for the companies because supply chain costs tend to be concentrated on the extreme. When you’re roughly right, it’s just fine, but if you’re downright insane, you can literally have a big operational problem that shuts down a factory or warehouse.

**Kieran Chandler**: Let’s talk a little bit more about the supply chain industry itself. How much have you seen companies in that industry implementing numerical recipes themselves, or would you say that the majority of people and companies are still stuck in that classical algorithm-based approach?

**Joannes Vermorel**: The funny thing is that the vast majority of companies operate, I mean, literally all of them, through recipes. This algorithmic thinking is a recipe for some kind of data science disaster, so actually, there is a lot of hype, but there is basically nothing in production. So, everybody operates in practice through numerical recipes, and 90% plus of the market share is just Excel, but people look down on those.

**Kieran Chandler**: Excel sheets, saying it’s just Excel, no, it’s not just Excel. It’s the embodiment of the understanding on how you should actually model quantitatively your supply chain. So, those Excel spreadsheets, they are literally the numerical recipes, and they are the refined version of those recipes. In this respect, it’s quite good. Where it’s not so good is that the spreadsheets, in general, don’t matter if it’s a spreadsheet on a desktop or on a web app, offline or online, a spreadsheet or the tabular thinking is not exactly suitable to solve supply chain problems.

**Joannes Vermorel**: My big criticism is that the tooling is inadequate. You can’t deal with uncertainty, you can’t deal with cannibalization, you can’t deal with a multi-echelon supply chain. There are so many problems that literally don’t fit in a spreadsheet, no matter how you package the spreadsheet. My criticism is not that the problem with spreadsheets is that they are numerical recipes that are highly subjective and include a lot of narrowness. This is not part of the problem; this is literally part of the solution to the problem. My criticism is that this tooling is usually inadequate. Present-day companies operate through numerical recipes, but they don’t acknowledge that this is a good thing, and this thing is not going to go away. This is literally a very reasonable proposition to address supply chain problems. But the problem they face is inadequate tooling and frequently inadequate processes, such as, for example, the divide and conquer problem that we discussed with silos, where people can be trying to deal with pricing on one side and planning on the other side while it’s literally the two sides of the same coin, as discussed in the last episode. Numerical recipes are here to stay, and my position is that it’s literally the right mindset whenever supply chain problems are involved.

**Kieran Chandler**: We’ll start wrapping things up then. What is the main conclusion of today’s episode? Why is it that numerical recipes are so important, and why is it so important to change that mindset?

**Joannes Vermorel**: I believe it’s because numerical recipes are the embodiment of this other line of thinking, which is, “It’s better to be approximately right than exactly wrong.” You will end up with formulas that don’t have purity. They are not like electromagnetic equations, where you have those super neatly defined equations that can define everything that happens in terms of electromagnetism. It’s incredibly pure and tight, but supply chains are not like that. Supply chain numerical recipes are going to be hundreds of semi-accidental conditions, factors, and twists so that the whole thing makes sense, so that the whole thing is approximately correct and doesn’t do anything completely insane. It should be highly predictable, so you don’t have too much surprise, ideally very little amount of surprise in the numerical outputs of your recipes. And it should also be versatile, just like a great chef’s recipe.

**Kieran Chandler**: Um, you want to do a dessert, you know what, I’m not going to allow you to use sugar today.

**Joannes Vermorel**: Oh crap, I want to do a dessert. How am I going to do a dessert without sugar? That’s, you know, the sort of thing you need to be super agile for. So that if there is something missing, just because you had weird conditions going on, like a pandemic, you’re not stuck. You have a way forward. And by the way, that’s very interesting because those top chef shows give you challenges where either you’re short on time, you know, if you have only 30 minutes to prepare something that normally would take like four hours, or you’re short on ingredients, or you’re short on tools, or you’re just short in general. And yet, you need to find a way forward. That’s, again, I believe what those recipes look like. I mean, you have weird constraints that are changing over time. It’s a situation that comes with a degree of surprise.

The real chef is the one that can literally improvise. But if you watch closely, those shows reveal that there is a method to it, and that’s really what differentiates a great chef. The great chef is not somebody who is going to do random things when facing a missing ingredient or a very short timeline. You can really see that there is literally a decade of experience in how to cope with this mess. There is a method to it, and that’s exactly the sort of things that we cultivate at Lokad.

**Kieran Chandler**: Okay, we’ll have to wrap it up there, but I think that analogy of a top chef is one that’s really kind of powerful and definitely one we can relate to in this office. We’ve got many fans here. So that’s everything for this week. Thanks very much for tuning in, and we’ll see you again in the next episode. Thanks for watching.