Hutter Prize: A Quest for AI Understanding | Microsoft ai Course | What is Generative ai Select the Best Option | Google Generative ai Course | Turtles AI

In the field of artificial intelligence, the quest for human-level understanding has been ongoing for decades. And for the past 17 years, AI researcher Marcus Hutter has offered a unique take on this quest through the Prize for Compressing Human Knowledge, also known as the Hutter Prize. The Hutter Prize challenges AI researchers and engineers to losslessly compress a one-gigabyte snapshot of Wikipedia to a smaller size than the previous winner. The goal of this competition is to demonstrate how text compression can lead to a better understanding of the text itself. This concept will be explained in a second. The current record holder has managed to compress the one-gigabyte file to 115 megabytes. The relationship between compression and understanding can be explained through a practical example. Imagine a text file containing millions of pairs of numbers, each representing the values of x and y in a given equation. While any compression algorithm could shrink the size of this file, the most efficient way to attain the highest compression ratio would be to comprehend the principles behind the connection of x and y, i.e.: the analytic function and its formula. By doing so, a simple computer program could flawlessly recreate not only the pairs contained in the original file, but any other set of x and y values responding to that analytic rule. A similar concept can be applied to the objective of the Hutter Prize, which is to compress the contents of Wikipedia. For instance, if the compression program has knowledge about the interconnection between velocity, time, and distance, it can remove redundant information when compressing pages related to classical mechanics. The same principle holds true for the pages about economics. The more the compression program comprehends the principles of economics, such as, for instance, the relationship between inflation and monetary policy, the more it can effectively eliminate unneeded words when compressing pages about economics. Large language models, like ChatGPT, are trained to identify statistical patterns in text. They analyze extensive amounts of text from the web and learn to associate phrases with each other, such as "unemployment rate rises" frequently occurring near "economic growth slows down". If a chatbot is trained on this association, it may respond to questions regarding the relationship between unemployment and economic growth by indicating that an increase in unemployment often leads to a slowdown in economic growth. However, these models are not eligible for the Hutter Prize due to their lossy compression, meaning that they don't precisely reconstruct the original text. Despite this, their ability to make connections between concepts and respond to questions in a plausible way raises the question of whether their compression indicates a real understanding of the sort sought-after by AI researchers. We can hence assert that Hutter Prize serves as a unique and ongoing challenge in the field of AI, pushing researchers to explore the relationship between compression and understanding. As AI continues to advance, it will be interesting to see how this relationship evolves and how it can be used to achieve human-level understanding in AI. Only when a large language model will be eligible to win the Hutter Prize, it will be error-free (at least compared to the original source of knowledge).