Retired Ai Scientist - Tumblr Posts

Long before the recent AI explosion, engineers were not only developing the technologies, but also actively looking for the weaknesses. To be clear, the visual AI work we've been witnessing recently is generative - more or less convincing/compelling images. But the AI that drives your car or diagnoses your illness, these are decision makers. It's a different application of the same class of technologies - modern artificial neural nets. And this is what I want to highlight here. The above demonstration shows how easy it is to confuse these types of system about what they are looking at. The point is, some object we leave by the road could have a very unpredictable effect on the self-driving car. It wouldn't even occur to us that that would be a problem. I got the picture from this article, where the author goes into more detail.

I remember one talk I was in where the speaker had rigged their AI to highlight the parts of the image that it based its decision on in identifying fish. Turns out, the AI had just learned, if there's seaweed in the background, it's a fish. You see the problem.

I do actually believe in AI. But what we have now is somewhat fragile, and has no concept of our values, and we should bear this is mind in deciding how we use it.