In 20 years in developing electronics, I’ve heard it a lot: “Our product is so much better! The algorithms our engineers developed are way beyond what our competition has.” In the 1990s, I saw it with my own eyes at Philips. Their TVs had the best deinterlacing algorithms in the world, resulting in superior picture quality. In the 2000s, I worked on audio enhancement: turning on sound processing algorithms gave quite a stunning effect. Tiny speakers would suddenly create bass and provide a 3D sound stage. Over the next years, I’ve seen cameras getting a much better picture quality, virtual reality headsets providing a much smoother experience and compression algorithms becoming a lot stronger. The before and after effect is a very strong sales tool.
Now, we’ve entered into an era where it’s not about picture or audio quality anymore. Instead, our electronics need to become smart and adopt AI. Andrew Ng from Stanford put it well: “If a typical person can do a mental task with less than one second of thought, we can probably automate it using AI either now or in the near future.” Even though AI can only solve fairly simple tasks, this presents a lot of business opportunities. The market responds and big corporations acquire AI teams for large sums of money.
In AI, I’m seeing the same “my algorithm is better than yours” claims. Many companies state their AI is better than everyone else’s and are showing the before and after effect. Kudos to these companies, for all having managed to hire the smartest AI algorithm engineers? No, not really.
With AI, for one, it’s actually fairly easy to build your own algorithm. Instead of having engineers that skillfully craft and program the algorithm, AI is simply trained. You give the AI engine lots of examples over and over again and it keeps adjusting itself until it doesn’t make any mistakes anymore.
Another problem is that it’s hard to measure ‘better’ in AI land. There’s a famous saying that states there are lies, damn lies and then there are benchmarks. In AI, it’s no different. In automotive, for instance, you can measure false negatives, where you don’t detect a person in front of the vehicle. But a false positive, where the AI brakes for a pedestrian who’s not there, is almost equally bad.
The size of the data set is something to consider as well. It sounds impressive when an AI algorithm scores perfectly on 10 million kilometers of automotive test data, but since we have one billion vehicles on the road that each drive 10,000 kilometers a year or so, the AI probably still only covers a fraction of the real-world situations that can occur. Even perfect scores can be meaningless in such situations.
A final complicating factor is that the algorithms are still rapidly changing. There are many competitions where universities and corporate research centers battle it out and continuously introduce new algorithms. The winning neural networks are freely made available on the web, as are the tools to adapt and train them. Download and retrain the model for your target application and you’re done.
Thus, we come to the conclusion that AI algorithms are easy to develop, hard to benchmark and ever-changing. That’s a problem because it doesn’t seem like a great foundation on which to build a solid company. Strong companies typically operate in markets that have high barriers to entry, making it difficult for new entrants to come in and compete, which isn’t the case with AI.
My advice: don’t rely solely on AI in isolation, but use it as an enabler for your business and closely integrate it into your products. Simply focusing on “my algorithm is better than yours” won’t give you a sustainable competitive advantage. When your engineers tell you that their algorithms beat the competition, congratulate them, but ask them right away how they’re planning to maintain that advantage.