Thanks TD, it certainly looked like an improvement, but not knowing anything about “bulge” I didn’t know. These models are only going to get better, plus there seems to be a knack to using them. For example, quirky as it may seem, many report that asking the AI to identify mistakes in their own answers will significantly improve the resulting output.
Thinking further about your example about driverless cars being able to handle the road but being flummoxed by the hand gestures of other drivers, a few things occur to me. First of all, if this is presented as an example of a domain that is ultimately indecipherable to AI, then I think that is very unlikely. A few years ago there was real debate about whether AI would ever be able to distinguish a cat from a dog. Some said the task was simply too complicated and AI would not manage it for many decades to come, if ever. Yet that barrier was overcome very swiftly, and now AI cannot only distinguish cats from dogs, it can identify many thousands of species that no single human can identify. It can distinguish human individuals from their faces or parts of their faces, and much, much more, far beyond the capability of humans. In a few short years we’ve gone from AI that can’t tell a cat from a dog to AI that is much better at identifying at categorising objects in general from visual input than any human.
Therefore, the fact that driverless cars can’t seem to understand human hand gestures is probably simply indicative of the fact that the models have not yet been pointed in that direction. If AI is set the task of analysing a large dataset of hand gestures then it will very rapidly become better than humans at identifying and responding effectively to hand gestures. Plus other physical/visual cues that humans are aware of, such as a car slowing down, speeding, peeking out, hesitating, and everything else, plus sounds and smells and so on. None of these inputs are inherently beyond the capability of AI to categorise and respond to effectively. It’s just a matter of turning the attention of AI in that direction in the first place.
Aha, you might say, there you have it! AI doesn’t even know what to look for unless a human first tells it what to do! That’s true to some extent, but the range of abilities and level of abstraction becomes ever wider, so that humans can give more and more general instructions, and the details can be worked out by AI.
In other words, each time the risk is narrowly defined and effectively completed, humans can step outside the defined area of analysis and ask the AI to move up to the next level of abstraction. In other words you begin by instructing AI to deal with specific situations, but the better it becomes the more general the instructions can be. Starting from telling AI in driverless cars not to collide with other objects, using the rules of the road, then taking into consideration the actions and visual cues of other drivers, all the way up to the point where the AI can be asked to identify all the relevant factors impacting safety in moving vehicles and to optimise for all the relevant factors it identifies.
It can then be asked to analyse its own results recursively in order to identify any safety factors that had been neglected. It can then be instructed to address those factors it identifies in an endless loop, until what you arrive at is a system that copes with all relevant aspects of safely and handles them better than human drivers. This process will probably be very quick because all the gains in knowledge by AI systems can be shared across the network and once learned are never forgotten.
