I graduated in 1992 with an engineering degree entitled “Artificial Intelligence and Advanced Programming”. At that time, from a class of 300 people, we were four that chose this specialty. It was at a time in history when AI was in “winter” mode. People had heard so much about it but none of the fancy promises of the 1960s had been delivered.
It all started with a Pacman clone
My first job was to develop (advanced programming) a Pacman clone, Conny, with its inevitable “intelligent” (AI) monsters chasing him. I then continued for more than 20 years studying knowledge representation and how to easily represent and execute business logic defined by business people using domain specific languages, which is also, technically, a part of AI (expert systems, rule engine with forward chaining, operational decision management). It was approximately 2018 when I decided to get into machine learning, which had become the AI topic. I wanted to understand why people were so crazy about it and why people were thinking that, finally, “general” AI was about to be within reach.
I joined Rapidminer, which is one of the leading platforms for data scientists, so that I could be at the heart of the revolution. What I then realized is that not much had been invented since my machine learning course of 1991: basically, the idea is still to gather a lot of data, try to identify patterns in this data and use these patterns to make predictions about other data and/or the future. Why this mattered was that enormous amounts of data could be accumulated and that the genetic and neural network algorithms that were super computing intensive (but which were invented decades ago) could finally be run within reasonable timescales. When I began to search for anything that looked like general AI, all I could find was Watson (that could win Jeopardy), Siri (that could understand “tell me a joke”) and DeepMind (that could beat almost any human at Starcraft) – all of which were useful AI applications in their own right but would never qualify for a general JARVIS/SkyNet-like AI.
So I was reassured that general AI would not take over humanity any time soon and was also quite surprised by the capabilities of what could be done with machine learning. Basically, most of the business applications that I had spent 20 years modelling using knowledge representation and explicit business rules to generate decision trees, lookup and decision tables could be replaced by a smart machine learning model that only needed enough data history to provide acceptable decisions or predictions. But what was also clear was that the rest of the problems remained:
- How to map the representation of information within the computer systems to which business people manipulate in their everyday life
- How to integrate decision systems into business and IT processes while still providing the possibility of supervision and auditability by humans
- How to do more than just predict the future without trying to influence it
Solving these problems was the key to making AI/data science projects successful. However there was also a new class of challenges brought by the machine learning way of making AI-based decisions: the need for data science skills to carry the correct model training, the need for data engineering to extract, transform and manage the huge amounts of data and, finally, the fact that most of the time it is difficult to collect the right kind of data.
I was seeing a lot of machine learning projects failing because they could not address these problems. Most of the time, companies thought that machine learning was a kind of magic: just accumulate as much data as possible, feed an AutoML system with all the data and “Voilà!” all business problems solved. Sometimes they managed to solve some of these problems but the one or two remaining were enough to make the whole project stumble.
I then went and spent a year trying to solve Environment, Health and Safety problems without AI or advanced programming. Yet my brain was still intrigued by machine learning and all the new fancy advanced programming techniques that had come to life in the past 30 years. It was clear that I wanted to get back to AI and advanced programming and make it work for businesses.
That was the time when I met Fabian Rang and Thorsten Heilig who had developed a concept that would solve these very problems to finally enable AI-based decision-making. We discussed it and my first feeling was, “That could work”. I joined them shortly after and helped them build paretos. Today, just over a year later, I can say that deciding to join paretos at the time was the correct decision – I was right to trust my first feeling!