AI thought process visualization
I started thinking about all the original computer science CGI stuff you could do in a sci-fi movie or TV series.
Like, you have this robot thinking about what it’ll do, and it’s running some breadth-first search or whatever, and we’re shown an elaborate crystalline 3D search tree slowly expanding from some starting point. We can see various possible actions, like “ASK THE HUMAN NICELY FOR INFORMATION”, sitting at various points in the search space. One by one, the tree expands to them and then discards after a moment’s evaluation.
And then, a distance away from the starting point, there’s this “SHOOT THE HUMAN” decision that we’re shown and that we can see the decision tree slowly but inevitably expanding towards. Then maybe it’s a race against time for the main characters to give the robot some new information that will change its evaluation criteria to reject that course of action when it reaches it. Or something.
Anyway, I bet that all kinds of seemingly-boring, basic compsci concepts like search trees could be made to seem really cool and exciting with a little work.
(This thought was inspired by seeing http://www.idsia.ch/~juergen/oopstree720.jpg and then imagining that search tree slowly and organically growing. Although that one looks more like depth-first, but anyway.)
AI thought process visualization, part II: The AI is shown as a space ship composed of many modules, floating in (concept)space. Around it are various fields of knowledge and subjects it could be analyzing, visualized as asteroids or other physical objects.
The AI’s attention is visualized as searchlights shooting out from its various modules to the surrounding objects. Most of the objects get no attention at all, or they are only occasionally thought about at a superficial level of analysis. This is shown as searchlights that gradually wander through the various objects, sometimes stopping for a moment at one but mostly just moving on.
Some objects, however, will catch the AI’s interest. At first one of the searchlights will only pause at such an interesting object. Instead of moving on, it stays there, and several other wandering searchlights will also be redirected to study it.
The AI then begins to devote more processing power to analyze this object (really a domainof knowledge, such as geography), switching from coarse-grained analysis using a few basic algorithms to a detailed investigation using a number of specialized routines. One of the modules that was shining light at the object will split apart, a large homogenous shape separating to a buzzing swarm of shapes, each with their own searchlight that is smaller but much more intense. They move to surround the object being studied, and soon the tiny searchlights find seams in its structure, cutting away its outer layer. As they do so, the object expands, unveiling a dark mass within. As the swarm turns its searchlights on the mass, they reveal its features, a planet’s surface appearing from under the surface of an asteroid. And each feature grows a new unexplored region around it as it is revealed, the object having a fractal structure that grows more and more complex the more that it is studied.
The initial, domain-general algorithms and tricks that were initially used to study the subject rapidly grow less useful. The searchlights of the initial swarm of shapes begin to dim, revealing fewer things in the darkness with each pass. The object stops growing, as well: features are slowly discovered, but they no longer yield new insights. At first, the searchlights found nothing but new kinds of things: forests, mountains, lakes. Now they find nothing new, no cities or volcanos. Each revealed feature is just a slightly different variation of the old ones, and it looks like the object’s new surface might soon be entirely mapped.
But then, after enough seemingly-identical features have appeared, one may start to notice a pattern that was not previously discovered. Searchlights are aimed at this pattern, and the AI begins experimenting with ways to exploit the new structure that has been found in the domain. From one of its modules, a new stream of small shapes – experimental algorithms customized to make use of the new information – makes it way to the object. One by one they train their searchlights on the new pattern, most of them still revealing nothing new. But then cracks begin to appear in it, cracks which widen as ever more new shapes shine light on it. And then parts of this layer, too, break away, again revealing a new kind of world below. It does not happen all at once, and many different patterns need to be studied and exploited before the whole layer gets peeled away: but gradually, deeper and deeper layers are found within the object.
The AI keeps making progress at understanding the object better: and as it builds new kinds of algorithms for studying this object, they get collected into an entirely new module, one optimized for this domain. Some members of the swarm join existing modules, too: the AI will experiment trying them at different objects, to see if they might have cross-domain applicability.
Gradually, the amount of new insights that can be collected from the object begins to again slow; and meanwhile, the searchlights of the AI’s other modules are wandering around the surrounding space, looking for something that might catch its interest…