AI tools in the past

 One possible approach to get an overview over a discipline is to ask which tools are important. Over the decades, the AI community has devewloped lots of tools which are promising to investigate robotics and artificial intelligence in detail. The sad news is, that most of the tools in the past were useless. There are certain hardware design available, e.g. parallel processors, there are cognitive architectures available, and some major programming languages. Also some algorithms and dedicated AI libraries were invented.

None of them was very succesful so it doesn't even make sense to name some of them. Instead of questioning the tool concept in general, a better idea would be to ask if a certain category of tools might be more useful than other. Tools in the past were mostly an answer to a problem. For example a parallel cpu should make the computer faster, while a cognitive architecture tries to emulate a human brain. The alternative concept is to use the computer to define the problems more precisely. With the advent of deep learning this problem oriented concept has become very successful.

Today, there is only a single tool available which is a dataset. A dataset encodes a problem in a problem in a table. There are datasets available for image recognition, for motion capture, for voice recognition or for game playing. What these datasets have in common is, that they never provide an answer, but they are asking questions. The concept is similar to test driven devolopment in software enginering. The idea is, to program at first a test routine. This routine is applied to computer code and can evaluate if the code is correct or not.

In the original meaning, a dataset was used to benchmark a certain neural network. For example, a neural network can recognize OCR images with an accuracy of 71.4%. The measuörement of 71.4% is the result of applying the neural network to the given dataset. Instead of asking how to program robots or intelligent computers, the new task is how to measure the software written by someone else. The idea of a dataset is, to postpone the task of creating an AI software to a later moment in time. This reduces the complexity. Creating a problem oriented dataset is much easier than programming an intelligent software which is able to pass the test. A good example is a motion capture dataset, which is a recording of motion tracker coordinates combined with the correct label like "walking, jumnping, standup".

There is a reason available why datasets have become the most dominant tool in AI research, because they are filling the gap of a missing problem definition. Before its possible to discuss advanced philosophical ideas including advanced AI algorithm, there is a need to define a concrete problem first. The problem guides the debates, because any argument or software provided in the discourse has to improve the problem solving process. This allows to decide if a certain paradigm makes sense or not.