Skip to main content

Artificial Intelligence comes with some caveats for investors

What will it be able to do? How intelligent will it be? Actually, for now, not much more intelligent than this chicken 👆(video courtesy University of Melbourne, animal behaviour group)

And thus, maybe, you’d want to dig a little deeper before making investment decisions into a hot new AI startup, or buy shares of a corporation with new “AI enabled” product innovation.

One of the new types of “emerging AI” is “Reinforcement Learning” (RL). You reinforce good behaviour by giving “it” (in this case the AI but it can also be a chicken) a reward. This approach is based on Psychologist B.F Skinner’s Behaviorism developed in the 1930s (!!). Skinner in turn was influenced by Pavlov’s experiments. That chicken is a great example of how Behaviorism works.

Reinforcement learning is how we currently train our cutting edge robots. And applications in manufacturing. Also in finance, inventory management, video games, image recognition, self driving cars, medical applications and even marketing tech.

Like the example with the chicken, the tasks the Artificial Intelligence (AI) can learn are narrowly defined. Change the colour of the disk and the chicken will have to learn again. Change anything in the scenario the AI trained for, and you will have to retrain the algorithm. Unlike a human an AI can’t make associations and draw parallels from other things it learned in the past. It can't understand a new situation. Or judge the context it operates in.

The next time a hot new startup talks about their AI it’s OK to have some reservations before you invest. Or when a corporation shows off their new autonomous robot in a demo. Because it might be that their AI is trained specifically to do this particular task really, really well so it can win over first customers and investors.

But how much training resources went into this before it learned this task? How much support was needed to do the demo? Is that scalable to similar tasks for new clients? Can they generalize their approach to different tasks?

After all, a cool demo is something totally different than a new revenue model you can invest in.