Unexpected Side Effects

In the podcast Die Maschine: Kontrolle ist gut, KI ist besser (in German language, by the radio station Deutschlandfunk) a scary fictional story is told from the 21st minute on:

An artificial intelligence has been developed that controls and executes all drug shipments worldwide. Because this was so critical, a special algorithm was chosen to ensure that individual population groups are not disadvantaged under guarantee, an algorithm which is always 100% politically correct.

When the artificial intelligence was activated, things went well at first, but then the number of deaths of diabetics increased in the rich countries. Insulin is lacking everywhere in the hospitals of the industrialised countries. How could this happen?

Well, the system worked exactly as it was designed. However, the artificial intelligence took into account the need for drugs worldwide. But there were not enough drugs like insulin for everyone on earth. Underserved areas, especially in Asia and Africa, received more drugs from the artificial intelligence, while the rich countries received less. Thus the shortage was distributed evenly across the globe.

It is a similar dilemma to two burning houses with people trapped inside, but you only have enough helpers to fight one fire and save the people, not both fires. What are you doing?

These are ethical questions that an artificial intelligence cannot answer automatically. So when artificial intelligence is used to sustain life, we should look very closely. And well-intentioned is certainly not always well done, as the story of the drugs shows.

ML Job Interview

There are many variants of this joke floating around It is so cool. Here is my favorite which I found on Twitter:

Interviewer: What’s your biggest strength?

Me: Expert in Machine learning

Interviewer: What’s 9 + 10?

Me: 5.

Interviewer: Nope. 19.

Me: It’s 14.

Interviewer: Wrong. 19.

Me: It’s 19.

Interviewer: What’s 2 + 2?

Me: 19.

Interviewer: You’ll overfit right in!

Like a Broken Marriage

There are cases where I think IT and business are like a broken marriage, and my work is that of a family therapist. What makes me think so?

Well, at a meeting of IT specialists I once asked which of the IT experts has the pressure to provide infrastructure faster. No one had come forward. They all said their work was okay.

A week later at the Machine Learning Conference I asked a few people where they run their applications. They said in the public cloud. When I asked if they were considering doing the same in their own data center, they gave me a big look: “I would never ask my own IT department if they were running machine learning applications. They are way too slow to deploy.”

No wonder IT staff feel no pressure, that business doesn’t even ask for faster deployment because they have given up.

It’s like in a marriage where the spouses have given up communicating. If you want to solve this, it’s hard work.

Now there are certainly IT departments that are working well with their customers in their respective business areas. But, dear IT people, are you sure that you know all the requirements of the business units? Are they still talking to you, or have they given up? It might be a good idea to validate assumptions explicitly. Maybe there is still unused potential for improvement. And dear business departments, have you asked your IT department lately if they could react faster? Perhaps you have overlooked potential in your own company?

If IT reacts too slowly to business requirements, then it has very little to do with technology, it’s all about processes and team structures. And above all it is about communication. Maybe you are getting help to get communication back on track.

Between Hype And Extraordinary Benefits

Last week I was at the ML Conference in Berlin. The conference was very instructive for me. The lectures were about the technology behind machine learning, data science, security and algorithms. But then it was also discussed how to set up an ML project, which mistakes can be made, and why container technologies like Docker or Kubernetes make ML projects much easier. Most exciting were the lectures about concrete applications of machine learning and artificial intelligence. And of course the discussions with other participants did not come too short.

My Conclusion

Most companies are currently experimenting with machine learning. There are only a few ideas on how these new technologies can bring real business benefits today. The hype in the media – both positive and negative – was far away. At the same time, everyone I spoke to agreed that in a few years, any company will have to use machine learning to stay in the market. It will be the lowest common denominator.

It will be used both in the public cloud and in the company’s own data center. The biggest problem in your own data center seems to be the slow processes. “It takes far too long for me to get a system set up. I’m much faster in the public cloud,” several people told me.

From VMware’s (my employer’s) point of view, I noticed that some participants knew us, but none associated VMware with machine learning and artificial intelligence. VMware offers interesting options when it comes to the flexibility and security of ML applications. Well, I’ll deal with that another time.

Here are two more highlights from my point of view:

How we avoid maritime accidents

The Keynote Data to the Rescue! Preventing Accidents at Sea by Dr. Yonit Hoffman of the Israeli startup Windward inspired me. Did you know that 10% of all ships have an accident once a year? With big data analysis and machine learning, Windward has developed models to predict and prevent accidents. 150 million data points on ship positions, ship types, weather information, water depths, proximity to ports, and many more have been brought together and translated into deep knowledge and understanding of what happens at sea. The lecture was really captivating.

The results have led to a very fruitful and interesting cooperation with insurance companies. Imagine which worlds meet. Maritime insurance companies have been working in a similar way for 350 years – often with paper. Now something new is emerging. And it is clearly driven by business. This is how I imagine the future of machine learning.

Rainwater drainage pipes for China

The most exciting conversation was with an employee of a company that manufactures sanitary systems for large buildings and rainwater drainage pipes. A special challenge are large flat roofs, e.g. for industrial plants or shops. The drainage pipes must have exactly the right diameter. If they are too small, too little water flows off. If they are too large, there is no suction effect. With the optimum diameter of the pipes, a suction is created in the pipe, which optimally drains off the rainwater.

I have learned that many years of professional experience are necessary to design these optimal pipes. This is a problem when you want to enter new markets. Now the company has managed to determine optimum rainwater pipe diameters much more quickly using machine learning. Now, for the first time, it is possible to enter new markets such as China without having employees there who have decades of experience.

If machine learning and artificial intelligence already deliver value in this extraordinary way today, then we will certainly see even more examples in the future. The limits are our creativity and imagination. Knowledge of technology is not enough. This is only a necessary, but not a sufficient condition for the successful use of machine learning. Identifying business ideas is the real challenge.

Good Code is like Good Literature

The basis of many innovations is software. “Software is eating the world”, said Marc Andreessen already in August 2011. Software is also the basis of all machine learning. And what does software consist of? It is code. Now, I think it’s important to understand what good code is.

A great colleague recommended the book “A Philosophy of Software Design” by John Ousterhout. The author’s name looked familiar to me. But I didn’t immediately know why. Only in a later chapter, in which John wrote about the Tcl scripting language, did it become clear to me. I had used the same author’s book “Tcl and the Tk Toolkit” a lot in the 1990s. I then developed Linux device drivers (loadable modules for the Slackware Linux 0.99 kernel) for a measurement system in my physics studies, and I used Tcl/Tk for the application interface.

“A Philosophy of Software Design” is cool. It reminded me very much of another book I had read this summer: “On Writing Well” by William Zinsser from the early 1970s. The book explains how to write texts of all kinds. Be clear and use short sentences. Write for yourself, as you would like to read it. Good code is like good literature – it’s timeless. The software design tips are like tips for writing any kind of literature – keep it simple and avoid complexity. Define good interfaces with the right level of abstraction. Think of other developers who will work with the code after you.

These rules haven’t changed for decades. And even the seemingly agile and fast innovations end up abiding by these rules because they are based on code. And ideally on good code.