We are all familiar with AI in search engines (Google), voice assistants (Alexa, Siri) or social media (TikTok). Furthermore, AI has long since paved the way into other fields of application: bakeries and bakery chains use AI to minimize leftovers in the sales counter at closing time; Machine builders use AI for predictive maintenance and achieve savings of up to several million euros per year.
The introduction of AI is rarely plug-and-play, it’s no secret that the majority of AI projects still fail. No need to throw in the towel, this trial-and-error is part of the learning process. There are, however, some best practices, and the success factors for successful AI projects in companies can be named; Conversely, the pitfalls that need to be avoided are also known.
In this interview, I talk to Kevin Engelke (LinkedIn profile: click HERE), AI expert and head of the AI Competence Center at Beta Systems Software AG – actually, we are colleagues.
Note: The links to the various podcasts, blogs or further sources mentioned by Kevin were subsequently provided by Kevin and incorporated by me. For the sake of clarity, I have also formatted bulleted lists as a list in post-processing.
Further note: For reasons of better readability/comprehension, the simultaneous use of the language forms male, female and diverse (m/f/d) is not used. All personal designations apply equally to all genders.
Sebastian: Let’s start with a simple question. Where do you personally use AI in your private life? I use AI in various use cases; In most cases, I don’t even use it so proactively by training and applying an AI model myself; instead, I use applications that use AI in the background, so I’m simply a consumer. As an AI expert, however, I am aware much more often that a type of AI is being used than is perhaps the case with other users.
Online shopping, video platforms such as Netflix and YouTube and social media are examples of this – sometimes I act as a content creator, sometimes as a consumer.
My experience here is that the AI(s) work very well in the background, for example by using pattern matching for the selection of content for me, and I appreciate being tempted by the recommendations into clicking. For example, I like to click on the recommendations on YouTube’s homepage often, as they reflect my interests.
And then, of course, ChatGPT has become part of my daily routines. I don’t use it in the sense of “now I’m going to do my tasks with it”, but if I have a task – say, organizing and conducting an interview – then I would get 10 questions from ChatGPT as a suggestion, for example. It’s like asking a friend to contribute a few ideas to complement my existing ideas. In short: ChatGPT is a source of inspiration for me, sometimes a fail-check.
A few years ago, I installed AI in my smart home. I installed various sensors/actuators at home – for example, automating every light switch – and know when which lamp is on and which windows/doors are open. It’s totally nerdy, but it’s just fun for me.
And at one point I asked myself, “Hey, would an AI actually know when I want to do what? Which light should be on and when?” For this idea, I collected the data generated in the smart home and then evaluated it during a holiday over Christmas; In the beginning it was cool, but AFTER the holiday such funny things happened as: The lights come on at two o’clock in the morning. Why’s that? – Actually, you have to know that I tend to be nocturnal on vacation, unlike during my regular work weeks.
Well, how was the AI supposed to “know” that I was on vacation or in a work week at the time of creation, the pattern just didn’t fit. In short: That was a big funny fail – at least at the first attempt. To be honest, I haven’t had time for a second attempt ever since, but that time will come.
Sebastian: If companies want to use the potential of AI, be it to optimize internal workflows, to enrich existing products or be it to develop new digital business areas, what are the top three success factors? First, and that’s not even specific to AI projects: you have to clearly define:
You have to define the problem very clearly and “scope” it. For example, if you want to optimize a workflow, then that means that I ask myself the questions:
You should first select a small area from a larger range of topics and focus your solution on that first. And I also have to clearly define: When exactly have I solved my task?
Because very, very often it happens that the idea is placed: Client: “I would like to do something with AI.” Me: “Ok, what are you going to do?” Client: “I need the ‘egg-laying jack-of-all-trades'” – you know what I mean …
Technically speaking, companies often want to go from no use of AI to the full potential of AI technology with just a few steps, or even define requirements for a solution that would come close to “Artificial General Intelligence” (AGI). And you have to let go of that mentally. After years of trial-and-error with AI, we know one thing very well: AI works very, very well for specific use cases. So, it’s important to break down your problem to such a specific use case.
Second point is: collect and prepare data. That’s why I usually go into the data very quickly in the analysis/evaluation of AI projects. They must meet the following criteria:
If you don’t keep an eye on that, an AI project can get out of hand. We remember headlines where, for example, an AI developed racist traits. This can often be attributed to the fact that certain characteristics are underrepresented in the data.
For this reason, in the so-called “data exploration phase” it is so important to evaluate the data – with regard to quantity, distribution and quality from a wide variety of points of view. Sometimes it may be required to correct or supplement these data sets. The same applies to the quality assurance phase.
And here’s my third point: ethics and morality. This is something that we humans, the project participants, have to think about, because AI analyzes existing patterns in existing data in a very objective and statistical way. We have to be aware of the impact on people, on society, on the environment, especially when AI solutions fully or partially automate certain processes.
Besides ethics, the aspect of legality also counts: What am I allowed by law to do at all? Just think of GDPR. I don’t want to make a judgement on those regulations; they are simply the starting point and we have to abide by these rules. Full stop. And in doing so, we should remember that even major players in the field of AI are demanding a certain level of regulation, Sam Altman for example. And that shows: People want it. You just have to be careful not to fall into over-regulation.
Sebastian: So AI usually requires a lot of data. So, for a company to be successful, does it need a new data-centric culture? How can something like this be developed? Let’s start with a look back at the recent history of AI: I start with huge amounts of data, train an AI and it then recognizes patterns in the data; and for each special case, I used an existing architecture of a known model as a basis, but I trained it from scratch – which was an incredible effort.
In the meantime, there are well-functioning approaches such as
The principle behind few-shot learning and zero-shot learning is the adaptation training (re-training) of an existing neural network, which was actually developed for a different problem. For example, it was found that neural networks for image processing – so-called Convolutional Neural Networks – are super good at recognizing shapes in the hidden layers that have become increasingly complex:
And it is precisely these fully trained neural networks that can be “re-trained” very well. Let’s say, I have a neural network that is optimized to classify an image into the categories “dogs” or “cats”; this neural network has not actually learned what a dog or cat is, but it has learned: What are typical representations of shapes in images in order to be able to make the classification. And based on this, it is often enough to retrain only a small part of the neural network. This works even if the neural network was previously designed for a different task.
The astonishing thing about this is that in order to solve a different task, e.g. a different classification, only a fraction of the amount of data is needed compared to the original training and thus significantly less computing capacity. Why? Because the recognition of lines, shapes and structures has already been learned by the neural network and can often remain unchanged.
This is how the following example projects manage to achieve remarkable results with just a few data sets:
So today we often need significantly less data than just a few years ago, and that leads to the next important aspect: The culture of dealing with data: So how do you create a culture in which it is recognized that you collect data, and what you do with it in the first place? I believe that the best way to create acceptance is through understanding.
Trust is fundamental here, we have to make it transparent and understandable to those involved
The General Data Protection Regulation (GDPR) and the Federal Data Protection Act (in Germany: “Bundesdatenschutzgesetz”) also help and oblige us to do so, because data may only ever be processed for a specific purpose and because the processing must be made transparent and understandable for those affected.
When dealing with data, Google’s former guiding principle always comes to mind: “Don’t do evil” – I very much support this basic idea.
The technical possibilities to store documents and data, to set up data warehouses with structured or non-structured data already exist today. You can use document databases, graph databases, classic relational databases, etcetera: In short, there are numerous technological possibilities to collect and store large amounts of data. I don’t want to say that the whole thing is easy or doesn’t require any effort, of course it does. But it has probably never been as technically feasible as it is today, and it will only get easier in the future.
I therefore see the main focus on ensuring that those affected receive full transparency about the planned project and that it is also explained in a comprehensible way. It must become clear that the project makes sense and creates added value.
In addition, as DataScienstists / DataEngineers, I see it as our moral duty to understand the impact and functionality of the emerging AI solutions (keyword: Explainable AI) in detail and to explain it to those affected in order to be able to continue to evaluate the topic for themselves.
Sebastian: What is your master plan for the AI transformation that you have in mind as head of the AI Competence Centers? What are the key milestones and measures? In fact, I have a master plan. However, it is certainly not a master plan in the sense of: “Works for all companies”. It is a master plan tailor-made for Beta Systems, because the starting conditions are different in all companies. What’s the starting point here, at Beta Systems? – We are a long-standing software company, with a large number of products, we operate on a wide variety of platforms (zOS, Linux, Windows) and we have an immense variety of application areas and partly highly specialized software solutions in different business areas, that is: data center intelligence and identity access management.
Having said that, my master plan includes:
Firstly: Knowledge, understanding, involvement. At Beta Systems, I am the first to establish the topic of AI as a separate department in the company. Although there was already a discussion of the topic area before, also with partners, the current establishment of a “Competence Center” will clearly accelerate this process and take it to the next level.
My goal now is to deepen the knowledge and understanding of AI throughout the company, from trainees to board members. We have to involve people so that we can use an exchange and thus the potential in the minds of all employees. In the last three months alone, and also in my private projects, I have realized: the deeper the understanding of the technology, the more it promotes an exchange and fantastic ideas can be developed – across all departments and experience levels with AI.
Secondly, the whole project will take place “step by step” and grow successively: I have already mentioned it in the first part of our conversation: We will identify problems/ideas/projects, compare them with each other, prioritize them and strictly scope them.
We will then implement these projects with AI-based solutions; and the following applies: an AI solution can build on or use the previously created solution. I don’t build the AI that can then solve EVERYTHING. In practice, this means that we make a (small) model, use it to solve an identified problem, take the problem or solution further and build a larger model or the next AI solution based on it. Later, we might link the two models etcetera – like in classic software development: one problem solution at a time, and then we see what we can create on the basis of the two. And we are constantly developing and designing more functions, modules and solutions.
Whether the resulting AI model will be a classic approach from machine learning, a small neural network, a transformer or a deep neural network with billions of neurons, or even a mixture of them, will of course depend on the task at hand.
Thirdly. We can run some parallel projects, which is often better than “putting everything on one horse”. But we must not lose focus. We have already identified a lot of projects in different business areas. These are evaluated and prioritized internally: What are the priorities? Then we commit to it and work on the topic.
And here’s a crucial point: Fourth, we have to see, how long is it worth working on this? In other words, evaluate continuously. Since my time as a software developer and team leader, I have been strongly oriented towards the methodology of agile software development. We have to evaluate continuously: Are we on the right path, am I going in the right direction? Is the project still worth it? Have the requirements perhaps changed since the start of the project? This approach is known, for example, from agile software development, it is not AI-specific … But it also applies there.
And finally, Fifth, proceeding according to methodological-didactic principles, which I have already mentioned in part in the course of the conversation:
Sebastian: There are already some AI projects at Beta Systems; What are your principles in project selection and project management? When would you finish a project, and how do you find promising AI projects? I am currently in the process of developing an in-house evaluation form. In doing so, I was guided by the preparatory work of well-known large companies and open source communities. A large number of such guidelines are available online, as well as evaluation forms and checklists for AI projects, for example at government levels, but also from large tech companies.
What is it supposed to be about? On the one hand, it’s about ethics and morals. On the other hand: technical evaluation. You should approach it like a user story from agile software development. Who am I, what do I want, why do I want this, what do I want to achieve? And if it turns out that you want to use AI to evaluate or compare the performance of employees or the like: then don’t do it.
At this point, I would also like to make it clear once again that data protection is a very important topic for me. This may seem counterintuitive, as it is often said that AI companies or AI experts do not attach so much importance to data protection. But I attach great importance to it … after all, my own data is also affected. In addition, I would like to stress once again the importance of acceptance through understanding.
The purpose of the evaluation form is to make projects comparable in order to have a uniform basis for decision-making in the selection of projects. As I said, I’m working on the first draft, which is being created in close collaboration with other stakeholders at Beta Systems. In addition to ethics and data protection, important assessment dimensions are, of course, feasibility, data availability, implementation effort, risk for the company and the like. We also need KPIs that allow us to evaluate project success during implementation and after completion. And we also need to make it clear who would be involved in a project and whether they are available at all.
You have to have a medium-term perspective when making such decisions about AI projects: The decision on the implementation or feasibility of a potential AI project is not an eternally valid “yes” or “no”; sometimes there can be good reasons not to implement an AI project idea immediately, for example because the data basis is missing. In this case, a decision can also mean that implementation is postponed, but a data collection project is started.
What I want to avoid is that we do AI just because it’s a hype topic. And that is precisely why we need to carry out a proper evaluation before deciding on implementation.
And what I also want to avoid is reinventing the wheel. There are a variety of solutions and algorithms. In the vast majority of cases, people have done this before. There’s a whole bunch of minds out there who are damn smart; and here I like to use the term swarm intelligence, which is from my favorite forum Reddit. Connecting, exchanging ideas, supporting each other and creating great things together.
Sebastian: What haven’t I asked yet, what would you like to add? I think one point that we need to push hard overall is the acceptance that AI has potential at its core and will trigger things in the world that we can’t imagine it will work for.
To give you an example, there is the project to build fusion reactors, one of which has recently made big headlines because for the first time more energy has been generated than used. According to my layman’s understanding, the challenge is to keep the ultra-hot plasma with millions of degrees Celsius in a “stable form” in space, which is probably done with super strong magnets, which makes the plasma float in space. However, in order to maintain this stable shape so that the plasma does not decay, these magnets have to be controlled at very short intervals, I think it was fractions of milliseconds, and a system has to be trained for this purpose, which takes over this task – i.e. the control of the magnets.
Now, here’s my point: According to my hearsay, the proposal to use AI to solve this challenge initially caused people to shake their heads, because: “How can AI be used sensibly here?” What kind of patterns should be relevant for the analysis of which AI could be used? Well, you guessed it, AI has actually made a decisive contribution. And that’s what I mean… Involve AI in finding solutions with an open mind.
By the way, I also recommend the podcast by Lex Fridman on this topic, which deals with these questions in particular: What do we trust AI to do, how does the use of AI influence us, and so on.
Bottom line: We humans should not become the limiting factor of AI, because that would limit the potential of AI. And let’s be honest: I couldn’t have imagined ChatGPT a year ago, well, in the complex form that is feasible today, GPT-4. But here, everything is impossible … Until someone does.
Sebastian: Kevin, you’ve already mentioned the right keyword. Namely, podcast and other reading tips. Finally, could you give us some reading tips for two or three good blogs about artificial intelligence? For all those who are inspired by your interview to delve deeper into the topic… Of course, of course. You have to know: I’m a child of the internet, I grew up a lot with blog posts and YouTube videos. And in the meantime, it is true for me that blogs often reach the quantity and quality of a book … if not already exceed.
And what I really appreciate about it is the fact that it’s very interactive. It’s not a one-way street, you get to know what other people think about it in the form of comments; and it also links to other blogs. Overall, it’s also very dynamic.
First of all, I’d like to mention Josh Stormer; his videos take some getting used to in the first few seconds, because he usually “sings” the topic for the first X seconds in his videos – just take a look, you’ll know what I mean. But he really manages to convey his content according to the ELI5 principle, also with great pictures and funny characters, etcetera: Josh Stormer on YouTube
I’d also like to name Lex Fridman. He discusses AI, such as the one with the fusion reactor, as well as other topics, in his podcast, but also in his LinkedIn posts.Click here for the podcast: Lex Fridman Podcast
I also spend a lot of time on medium.com/tag/artificial-intelligence . In my opinion, Medium is now highly commercialized, but there is still very good content there.
As a community , I recommend www.huggingface.co and www.kaggle.com. And my all-time favorite is www.reddit.com, where I’m active in the subreddits around machine learning, data science and data visualization
