What if I told you that AI can solve almost any task, right now? It’s true. Here’s the algorithm. 1. Let the AI generate a completely random output. 2. Check if it’s correct. 3. If it is, declare success. If not, go back to 1. Of course, something like this either takes a very long time to get the correct solution. Or, if you only get one try to get the answer right, it’s outright unstatisfactory. It’s not that computers can’t solve tasks, it’s that we need them to not fail at tasks; we need them to be robust. We can put fully self-driving cars on the road today; just be prepared for a lot of crashes when the cars come up against novel road conditions as benign as pothole construction zones or as dangerous as children on bicycles darting across crosswalks. AI deployment is a sliding scale - as AI becomes more capable, we can deploy it in more and more risky situations. So what is a “risky” situation? Well, we’re willing to deploy AI in situations where it fails a lot if the cost of failure is low. But if the cost of failure is high, we need to have AI that fails less and less often. In other words, the reason we don’t deploy autonomous AI right now in some situations is because in those situations we really REALLY need to get it right. And that’s where AI is going to fail first.
