AI Singularity, or Can Zipping a Zipped File Give Zero Size?

Probably everybody who ever tried to compress a file asked themselves if compressing an already compressed file might save more space. Many tried to do it and saw that after a few iterations the file cannot be reduced in size further. Information theory states that there is some limit for any loss-less compression which we cannot overcome.

Topic of AI is becoming more and more popular, people’s attention is caught by many interesting ideas stemming from it and one of the most interesting, and frightening, of the ideas is AI singularity. Experts claim that after some progress in training AI we will become able to make AI improve itself without human control. Without control because humans will become incapable to understand what and how happens.

In Aigarth, running on top of Qubic, we are working on reaching AI singularity. We created a task for AI to solve, a useless from practical point of view task. We, like illusionists, do one thing while focusing everybody attention on another: Qubic miners prepare artificial neural networks (ANNs) and we try to improve them via fine-tuning, but the real goal is to improve the fine-tuning software which is an ANN itself. What for? To run this software with the task to improve itself one day.

How much can an ANN improve itself? This is a question similar to that one about zipping files. A Turing machine cannot solve the halting problem, what if an ANN cannot improve itself much enough for us to say that the improvement is qualitative, not quantitative? This is a valid concern because the universe likes to use universal laws and file compression, the halting problem, and ANN self-improvement may be just different manifestations of the same phenomenon. Luckily for us, evolution that happened on the Earth already showed that self-improvement may lead to a qualitative leap with help of mutations, which is randomization of ANN structure in our case.

So, we are pretty optimistic that we may reach AI singularity, what we don’t know yet is if there is some limit, similar to the limit of loss-less compression, which we may hit one day.