AI as the art of finding problems

 In classical computer engineering there are many obvious problems available for example how to compress a videogame into 64kb of RAM, how to design a CPU which is 50% faster, how to render a 3d animation on the screen. These problems are creating a creativity space in which possible answers are discussed, revised or rejected. A typical programmer never interprets a challenge as a dead end but as an opportunity to find an answer.

For Artificial Intelligence there is no such a problem available. This makes it hard to discuss what AI is about. The philosophical problem space of "AI is to build an intelligent machine" isn't a problem space in a technical space because no existing tools or algorithm can be applied to this challenge.

If classical computer engineering is about problem solving, AI is maybe the art of searching for problems. A good AI problem can be solved and is related to thinking machinery. Typical examples for these problems are computer chess, micromouse challenge and a chatbot competition in which the answers of a chatbot are scored. These problems can be solved with existing tools like hardware, software, libraries and AI algorithms.

A new and very powerful problem space is a dataset which is discussed in the literature frequently after the year 2000. A dataset is a universal problem because there are datasets for image recognition, trajectory planning, OCR, question answering, motion capture and so on. MOst real world problems can be presented as a dataset. The dataset is not the answer to a problem and it doesn't controls a robot, but a dataste formulates an AI problem in a machine readable format. A single dataset can be solved with different tools e.g. a certain neural network, a handcoded software, or with a pattern matching algorithm.

What can be observed in the literature is, that the amount of datasets has exploded since the year 2000 and the difficulty has increased. From today's perspective a dataset is a universal problem generator which is used to measure every new generation of large language models.

Let me give an example how a dataset works. Suppose there is a table with 10 motion capture poses including the textual annotation. The task for a computer is replicate the shown picture to text pairs. That means, the computer sees a picture and has to print the correct annotation on the screen. In this specific case, Artificial Intelligence is defined as the ability to find matching pairs in the dataset. In other words, the definition of AI can't be discovered in nature but it has to be constructed similar to a painting. The art of dataset creation means basically to discover new sort of problems not available before.