00:00:03 Evolution of supply chain software UIs.
00:00:44 3D interfaces’ limited future in enterprise software.
00:02:19 Comparing future interfaces to silent anti-spam software.
00:04:02 Balancing productivity, employee engagement in software use.
00:05:48 Future of supply chain UIs and automation benefits.
00:08:00 Trust in AI forecasts and metric roles.
00:08:18 Perspectives: End-user and software company.
00:10:57 Cross-entropy in machine learning.
00:12:13 Complex metrics, benefits of outlier focus.
00:13:45 Automated supply chain systems and their challenges.
00:16:02 IDE concept in supply chain modeling.
00:16:38 “Correctness by design” importance, trial/error cost.
00:17:41 Envision: Programming language for supply chain modeling.
00:18:00 Envision’s features: Auto-complete, static code analysis.
00:19:06 Programming over visuals.


In today’s interview, Kieran Chandler and Joannes Vermorel are discussing the progression of user experience in supply chain software. Vermorel is explaining why 3D interfaces won’t be introduced, citing human anatomy and practical limitations. He is suggesting that the future should focus on practicality and invisibility, drawing parallels with anti-spam software. Chandler is questioning the trustworthiness of software that requires minimal interaction, but Vermorel is emphasizing the importance of efficiency and simplicity. They are critiquing the excessive interactivity found in software, proposing end-to-end automation in forecasting. Vermorel is putting a stress on outlier detection and consistency in forecasting engines. He is discussing the intricate nature of forecasting metrics and the significance of correctness by design. For improved productivity in supply chain management, Vermorel is envisioning the use of smart widgets.

Extended Summary

Kieran Chandler, the host, and Joannes Vermorel, the founder of Lokad, engage in a dialogue about the evolution of user experience, particularly user interfaces in supply chain software. Chandler introduces the topic by mentioning that unless one is a software engineer, the user interface is the main component of the software they interact with. He brings up the popular Hollywood depiction of future user interfaces, such as the Minority Report movie where Tom Cruise engages with a 3D environment. However, the fact that this 2002 vision hasn’t materialized prompts Chandler to interrogate Vermorel on the future of user interfaces.

In response, Vermorel clarifies that 3D user interfaces aren’t about to be introduced in Lokad or any other enterprise software environment in the near future. According to him, the reason is not a technological limitation but the human anatomy. He contends that humans predominantly perceive interfaces in two dimensions. Despite humans having two eyes and being able to perceive depth, Vermorel maintains that a third dimension doesn’t add substantial value to understanding the world. He also underlines the physical impracticalities of 3D interaction, pointing out how exhausting it would be to mimic Tom Cruise’s actions in Minority Report for a long period. Vermorel brings up the example of 3D mice which, despite being invented around 40 years ago, haven’t been able to gain traction because of the physical exertion required to use them.

Furthermore, he suggests that the future of user interfaces probably contradicts what most people expect, focusing more on practicality than spectacle. Vermorel uses anti-spam software as a metaphor. He values this software for its quiet, diligent operation, removing spam from inboxes without users even noticing it’s there. He views this as a model for the future of user interfaces, with unobtrusive and nearly invisible machine learning-driven software reducing user load and interruption.

However, Chandler challenges this approach from a company’s perspective, asking how companies can trust software that their employees seldom interact with. Vermorel recognizes the dilemma but reminds that companies pay for their employees’ time, and time spent interacting with software is ultimately an expense. He argues that a user interface that mimics social networks might be enjoyable and interactive, but it could also lead to more interruptions and less productivity. As such, Vermorel suggests that the future of user interfaces should favor efficiency, simplicity, and minimalism.

The conversation kicks off with a critique of the prevalent practice of providing employees with interactive software, suggesting it can be counterproductive. Vermorel argues that while breaks for coffee and brainstorming with colleagues are crucial for a balanced work environment, too much of this can be problematic, and companies need to trust their employees to get real work done. According to Vermorel, software requiring constant interaction can disrupt productivity, particularly in the context of supply chain management where ongoing interaction with software applications might not yield productive outcomes.

Vermorel then touches upon the nature of supply chain software interfaces, which tend to be rigid and dry. In current trends, software developers are striving to make them more engaging through interactive and collaborative features. However, this approach, Vermorel contends, often leads to wasted time as employees might end up dedicating entire days to tweaking forecasts for hundreds or thousands of products. This drains productivity significantly, and despite the apparent interactivity and engagement, it may not improve a company’s output.

He proposes an alternate viewpoint where end-to-end automation is seen as the desired goal, rather than interactive forecasting. This suggests that forecasting in supply chain management should be fully automated, allowing people to contribute more productively to the solution instead of being stuck in repetitive tasks.

Chandler then asks how companies can trust a completely automated system. Vermorel responds by suggesting that companies should concentrate on identifying and dealing with outliers, rather than becoming overly engrossed in the specifics of the forecasting software’s metrics. From a company’s perspective, the critical aspect is to look for aberrant behavior in the forecasts, much like checking for incorrectly classified emails in a spam filter.

In terms of the software company’s perspective, Vermorel elaborates that the focus should be on enhancing the forecasting engine’s consistency across multiple datasets from different companies and periods. Back-testing is also highlighted as a valuable method for refining the forecasting process.

Vermorel starts by discussing the complexity of forecasting metrics, which, in his view, constitute the heart of their supply chain software, accounting for approximately 50% of its complexity. The software comprises hundreds of these metrics. However, Vermorel explains that revealing the full extent of these metrics to users might result in confusion rather than clarity due to their sheer volume and complexity. Therefore, he recommends that users should concentrate more on the output decisions generated by the system, especially on the outliers, which are the decisions that seem obviously wrong. These outliers warrant immediate attention as they are the most likely to be expensive from a supply chain perspective.

The conversation then moves to the future of software and whether it could operate in a ‘full auto-pilot’ mode, akin to anti-spam software. Vermorel asserts that supply chains are inherently more complex than spam filters, comprising various human, machine, and software components. Therefore, a one-size-fits-all automated software solution is unlikely to be effective. He believes the automation of complex supply chain management might be possible when artificial intelligence reaches or surpasses the level of human intelligence, but concedes that such a situation is still a long way off.

In this context, Vermorel reveals that the process is ’non-Hollywood style,’ alluding to its lack of glamour. Crafting of code is crucial in supply chain management. However, achieving ‘correctness by design’ is important because trial and error can be costly in this field. Vermorel proposes a programming environment that fosters correctness by design and shares that Lokad has developed a domain-specific programming language called Envision, which includes features designed to achieve correctness by design through static code analysis.

Vermorel envisions the future of user interfaces to involve creating smart widgets to boost the productivity and efficiency of supply chain scientists, who are both scarce and expensive resources. He contrasts this vision with the Hollywood-style 3D user interfaces which prioritize visual appeal and showiness over practical usability and functionality.

Full Transcript

Kieran Chandler: In today’s episode, we’re going to be discussing the evolution of the user experience, and in particular, how user interfaces have changed when it comes to supply chain software. Unless you’re a software engineer yourself, the chances are the user interface is the only bit of the software that you actually deal with. A lot of the time when people are asked about the future of user interfaces, they refer to films like Minority Report where Tom Cruise can be seen gesticulating in a rather cool-looking 3D environment. However, this film was released back in 2002 and it seems we’re no longer any closer to reaching that vision. So, Joannes, last time I checked, Lokad is still very much a 2D environment. So when’s all that going to start changing?

Joannes Vermorel: Let’s be clear on one thing - three-dimensional user interfaces are not coming anytime soon in Lokad, and pretty much in any enterprise software environment either. The core reason has nothing to do with technology, it’s a pure matter of human anatomy. Your perception of user interfaces is mostly two-dimensional. Yes, you have two eyes and you can see some depth, but it’s mostly a two-dimensional perception. A third dimension doesn’t add much to understanding the world. When it comes to replicating what Tom Cruise does in Minority Report, standing with your arms up for ten minutes, it’s just too tiring. That’s why 3D mouses, which were invented something like 40 years ago, never took off. You have to be an athlete to use that. The future of user interfaces is pretty much the opposite of what people expect.

Kieran Chandler: If you’re going to crush my dreams on what these future user interfaces are going to look like, perhaps you could share your vision of what these user interfaces will actually be like in the future. Maybe you’ve got an example that you could share with us here?

Joannes Vermorel: The interesting thing about the future is that it’s already here, it’s just not evenly distributed. If you want to have a look at the future, you should look at your anti-spam software. This sort of software filters all those interesting propositions coming from weird countries that you’ve never visited, and that offer you dreams to get rich. The interesting thing is how this software does it silently and diligently. If it’s very good, you never even realize it’s there. A very good piece of anti-spam is the sort of thing that just does its job for you so that your inbox stays clean, but you barely notice that it exists. That’s the future of most machine learning driven software. It will be something that is ambient and nearly invisible. It’s probably the opposite of the very cool three-dimensional user interface that you can see in Hollywood movies because there is nothing to see, so it’s not very visual.

Kieran Chandler: I certainly like the sound of anti-spam, it would definitely go a long way to reduce the amount of time wasted reading those really interesting business proposals from Nigerian princes and princesses. But that’s really my end user perspective. How about companies? How can they trust this piece of software if they’re never really interacting with it?

Joannes Vermorel: That’s an interesting dilemma. As a company, you have to pay for your employees and so any time they spend doing something is time that you ultimately pay for. So what do you want for your software, for the software that your employees are using? You could look for something that is a bit like Facebook - it’s social, it’s interactive, and people enjoy it a lot, but it’s also full of interruptions. So, it’s very funny because if you implement something that is very much like a social network, people would enjoy it a lot.

Kieran Chandler: If you were paying people to spend even more time in front of the coffee machine, you’d expect a few breaks in the day for them to relax, reorganize, and brainstorm with colleagues. But if this is constant, how does work get done? Companies are paying employees for real work to be done. So, can they trust software that demands constant interaction from employees? Because, to me, that seems like the opposite of productivity.

Joannes Vermorel: Indeed, there is a bit of a dilemma here for the company. They shouldn’t trust a piece of software that requires constant interaction too much. It’s counterproductive. Let’s look at this from a supply chain perspective now. Many of these apps weren’t interesting in the first place, especially in regards to demand forecasting software. The user interfaces for supply chain software can be quite dry, and while there is a trend to make them more engaging with collaborative forecasts, it’s not as effective as it seems.

Kieran Chandler: Can you elaborate on that? What can we expect from these user interfaces in the future?

Joannes Vermorel: While making demand forecasting more interactive and collaborative sounds appealing, it becomes a massive productivity drag. Imagine having hundreds or thousands of products and everyone in the company spending the entire day looking at curves, time series, and tweaking them. Even if it appears interactive and social, it’s not going to improve your company’s performance. For any company of any scale, the goal should be end-to-end automation with zero edge cases in the forecast, not collaborative forecast.

Kieran Chandler: Are you saying that people won’t actually be included in these forecasts?

Joannes Vermorel: Exactly. We want to achieve end-to-end automation and remove all productivity drag. If people have to do something, it should be something that contributes to the solution, not repetitive tasks.

Kieran Chandler: But how can we trust the results if people aren’t included in these forecasts? We would still need some metrics to assess the software and someone to check these metrics. How would this work in practice?

Joannes Vermorel: That’s a good question. We have two perspectives here: the end-user perspective and the perspective of the software developer or the company writing the code. From the end user’s perspective, you want to look at outliers or aberrant behaviors. Just like for your anti-spam, you don’t compile statistics of how many emails are correctly or incorrectly filtered. You occasionally check your spam folder for any misclassified emails. Similarly, in demand forecasting, you look for outliers, forecasts that are excessively large or too small. Those are what you need to keep an eye on.

Kieran Chandler: You don’t need to compile statistics, you just want to get rid of all the outliers. Now, from a software company perspective, when you want to improve a forecast, you don’t want to take the data set of one company at one point in time and see how you can improve one metric. Instead, you want to collect all the data set that you have. For example, we have helped over 100 companies to optimize their supply chains, so we have well beyond 100 data sets to optimize. You want to make sure that your forecasting engine is consistent and improves consistently over this entire mass of data set, not just one. Also, you don’t want the massive data sets alone, you want to do a full back test, going back in the past one week, two weeks, and so on. That’s how you approach this game of optimizing the world forecasting process.

Joannes Vermorel: That’s correct. However, even if we could share these metrics with companies, I am not certain it would greatly help them trust the software more. The problem is that the most relevant metrics nowadays, such as cross-entropy used in deep learning (and which has been used for more than a decade for anti-spam), are quite complex. These metrics apply to probabilistic forecasts and are vastly superior to classic metrics like mean absolute error or mean absolute percentage error, which are dysfunctional and yet still the standard practice in supply chains.

The challenge we face is two-fold. Firstly, we have to communicate numbers that are very alien to most companies. Secondly, when you want to build a forecasting engine like we did at Lokad, having metrics forms about 50% of the technological effort. They are not just a small element at the end of the work of designing a forecasting engine; the metrics come at the very core and they represent literally 50 percent of the complexity.

It means that we don’t just have a few metrics, we have literally hundreds of them. In practice, it doesn’t prove very helpful to share this wealth of metrics because it would take literally a book to explain what all those numbers mean. In the end, it generates even more confusion than being helpful. That’s why we typically suggest that instead of trying to understand all those metrics, companies should focus on the outliers.

Don’t try to reverse engineer the metrics in the software; it’s very complicated and not necessarily helpful. Instead, focus on the decisions that are generated as a final output of the system and focus on the outliers, the decisions that are obviously wrong. Those are the ones that need your most immediate attention because they are the ones that are going to cost you a lot of money from a supply chain perspective.

Kieran Chandler: We keep mentioning this term “anti-spam”. If the future of software is going to be like anti-spam, the difficulty lies in the fact that supply chains are inherently a lot more complex than just filtering out a bit of spam. Would this actually work in practice? Can we actually have supply chains working on full autopilot?

Joannes Vermorel: Yes and no. Indeed, a supply chain is a very complex system with lots of humans, machines, and software involved. So, there’s no hope that a piece of software with default settings can do everything. Anti-spam works silently without any need to set it up because all email boxes are pretty much the same, so you can have an automated setup for anti-spam as well. However, when it comes to the optimization of a supply chain for a given company, you need to understand the company’s strategy, financial incentives, client pain points, and a ton of other things. These aspects cannot be discovered by the software itself. Maybe in a century when we have human-level AI, this might be possible.

Kieran Chandler: Artificial intelligence, something that’s as smart as a very smart human or maybe even smarter, could possibly have a completely automated set up for complex supply chains. However, we’re currently quite far from such a scenario. That’s why at Lokad, we have these supply chain scientists. The supply chain scientist’s job is to model a company’s supply chain in a way that is both accurate and efficient. This poses a challenge in terms of user interface, because to do this effectively, it’s almost like needing an integrated development environment. Is this correct?

Joannes Vermorel: Absolutely, it’s a complex situation and it doesn’t have that Hollywood glamour we touched on earlier. The reality is, writing code is a craft. With good tools, you can do it faster and better. Correctness by design is very important in supply chain. Trial and error can be theoretically appealing, but in the real supply chain world, it’s extremely costly. You don’t want to make thousands of mistakes in purchasing just to eventually get it right. That would cost millions. That’s why you need a programming environment that aids in achieving this correctness by design.

Kieran Chandler: So, it’s not as glamorous as it sounds but there’s more to it than just writing code, right?

Joannes Vermorel: Exactly, it’s not Hollywood style. It’s like writing code. We aren’t trying to solve a general programming problem. We just want to solve the problem of quantitative modeling of supply chains. That’s why we have our own domain-specific programming language called Envision. Envision comes with features that are designed to offer a degree of correctness by design as you write the code. You can have productivity with features like autocomplete and you can achieve many degrees of correctness by design through static code analysis. For example, it can detect if a variable that you’ve introduced in your script has no impact on any supply chain decision. This could mean you’ve forgotten to plug an economic factor into your model, or you’ve simply discarded some dead code.

Kieran Chandler: So, even though it’s not a 3D user interface like in Minority Report, there is a future for user interfaces in this programming environment?

Joannes Vermorel: Yes, indeed. The future of user interfaces at Lokad is geared towards making supply chain scientists, who are valuable and scarce, more productive and efficient. The focus is not on a 3D user interface where you can slide things and draw charts visually, but on smart widgets in the programming environment.

Kieran Chandler: Well, that’s enough to wrap things up. Thanks for taking the time out to discuss with us about user interfaces of the future. It’s definitely been an interesting conversation. It’s really interesting to compare the Hollywood vision of reality with what will actually happen in the future. Thanks very much for tuning in to today’s episode. As always, we’re here to help if you’ve got any further questions on user interfaces, and we’ll be back very soon with another episode. Until then, see you very soon.