04-16-2021, 03:20 PM
In the late 1970s I worked on simulating predator/prey interactions on Apple][ computers. An underlying requirement was a random number generator. The built-in RND function was not good enough because it would always produce the same sequence given the same starting point. When you subject results to statistical analysis, you are looking for hidden correlations, not for underlying patterns in your supposedly random methods. The initial challenge was simulating the behavior of a housefly on a windowpane. That task falls neatly into what can be easily visualized on a computer monitor and doesn't seem to require too much intelligence. After all, how smart is a housefly?
Statisticians have a wonderful term for such random-seeming behavior - the drunkard's walk. It turned out that seemingly simple rules could produce what looked like very complex behavior. Remember when fractals were trendy? It also turned out that even with a proper random number generator, it took a fair amount of calculation in order to produce what looked like mindless blundering. It also made very clear that any results were the consequence of the underlying assumptions that were implicit in the rules. How much should the last step size or direction or update rate affect the next step? How do the rules interact with the screen edge? Are there longer-term dependencies? For example, does the fly tire? Does the fly stick to the same rules, or does it change its behavior after a period of time? Even the very simple is not necessarily as simple as it seems.
When simulating that "simple" behavior, the computer was slower than a housefly. The random number generator needed to be coded in machine language for the fly to move at all naturally. That gave me a new respect for how primitive computers were in comparison to real life. Back then, we spoke of hardware, software, firmware and wetware. (You don't run into the term wetware very often these days outside of science fiction. The AI promoters prefer different vocabulary.) Of course, modern computers can run circles around an Apple][ computer, but the tasks that modern AI attempts are vastly more complex than simulating a fly on a windowpane.
Another reality of simulation is that it needs to be faster than life in order to be more than an academic exercise. If your AI weather forecasting model is slower than real-time, your forecasts will predict yesterday's weather. There is not a lot of demand for that. Still, conscientious model makers update their models so that they "predict" past events accurately. Yes, that refinement process improves the models, but it also produces the illusion that the models are excellent predictors of unknown future events. That is only the case when things stay generally the same. If anything actually changes, the models will almost infallibly under-predict the effect.
When you hear projections of how Artificial Intelligence is going to help us solve the problems of the future, remember the fly on the windowpane. Even a dull person would use the door if they wanted to go outside. It all comes down to how you frame the problem.
Tom Lawson
April 2021
Statisticians have a wonderful term for such random-seeming behavior - the drunkard's walk. It turned out that seemingly simple rules could produce what looked like very complex behavior. Remember when fractals were trendy? It also turned out that even with a proper random number generator, it took a fair amount of calculation in order to produce what looked like mindless blundering. It also made very clear that any results were the consequence of the underlying assumptions that were implicit in the rules. How much should the last step size or direction or update rate affect the next step? How do the rules interact with the screen edge? Are there longer-term dependencies? For example, does the fly tire? Does the fly stick to the same rules, or does it change its behavior after a period of time? Even the very simple is not necessarily as simple as it seems.
When simulating that "simple" behavior, the computer was slower than a housefly. The random number generator needed to be coded in machine language for the fly to move at all naturally. That gave me a new respect for how primitive computers were in comparison to real life. Back then, we spoke of hardware, software, firmware and wetware. (You don't run into the term wetware very often these days outside of science fiction. The AI promoters prefer different vocabulary.) Of course, modern computers can run circles around an Apple][ computer, but the tasks that modern AI attempts are vastly more complex than simulating a fly on a windowpane.
Another reality of simulation is that it needs to be faster than life in order to be more than an academic exercise. If your AI weather forecasting model is slower than real-time, your forecasts will predict yesterday's weather. There is not a lot of demand for that. Still, conscientious model makers update their models so that they "predict" past events accurately. Yes, that refinement process improves the models, but it also produces the illusion that the models are excellent predictors of unknown future events. That is only the case when things stay generally the same. If anything actually changes, the models will almost infallibly under-predict the effect.
When you hear projections of how Artificial Intelligence is going to help us solve the problems of the future, remember the fly on the windowpane. Even a dull person would use the door if they wanted to go outside. It all comes down to how you frame the problem.
Tom Lawson
April 2021