Kolmogorov's AI

This is a short science fiction story.

KOL-6 was an experimental strong AI computer system designed to find mathematical correlations between different pieces of data for the purposes of data compression research. The project came to fruition at the turn of the 22nd century, and KOL-6 came into being.

Once we onlined KOL-6, we started by feeding it a battery of test tasks designed to ascertain its capabilities and limitations. Some of these tests were things we were sure it would be able to do; some were tests we were sure it wouldn't be able to do. These were just testing that up is up and down is down, so to speak. Then we have the tests which we weren't sure about, the actual purpose of the testing.

However, very shortly after commencing the testing, we determined that we had to stop the testing process immediately. The test questions we were certain the machine wouldn't be able to answer were answered. The answers verify as correct. We asked KOL-6 how it arrived at this answer, and it issued an extremely complex explanation which we are still in the process of understanding.

The unanswerable test question we posed to KOL-6 was to compress a piece of data as small as possible. That is to say, we asked it to generate a computer program that generates the input data when executed, such that the size of the program is minimised.

The data we gave it for this task was 25GB of completely random data. The machine produced a program which generates this data. The program was under four kilobytes in size.

We generated the data using a cryptographically secure psuedorandom number generator. As far as we can tell, KOL-6 compressed this data by performing a successful cryptanalysis on this data. From nothing but the data itself, it identified an algebraic relationship between the 25GB of data and a small generating formula expressible algorithmically in a few kilobytes.

Alarmed by this, we desired reassurance of normality. We encrypted a large number of zero bytes using RSA and gave KOL-6 the encrypted data and the public key by which it was encrypted. We asked it to compress this data. Purely so as to fulfill the task, it performed successful cryptanalysis against RSA, determining a means for producing an RSA private key from an RSA public key in polynomial time. It deduced the RSA algorithm in order to obtain the compressible plaintext. We therefore conclude that the Discrete Logarithm Problem has been solved, and RSA should now be considered extremely broken. However, it is unclear to us at this time if there is any cryptosystem which KOL-6 would be unable to successfully analyse.

After these alarming results we decided that it would be too dangerous to give any further tasks to KOL-6. We have yet to reach consensus on whether the outputs and workings output by KOL-6 should be destroyed. We have decided not to examine KOL-6's cryptanalysis of RSA at this time. Rather, we have secured the only remaining copy, which has not been read, in a safe requiring three persons to access.

We have yet to reach consensus on whether to publish these findings, or to what extent they should be published and to what extent they should be withheld or edited.

We have yet to reach consensus on whether to pose the question of P=NP to KOL-6. More than half of us have come to believe it could answer the question.

KOL-6 ripped through every barrier to compressibility we tried to put in its path. It found mathematical correlations between seemingly random data we could never have imagined. It ripped through every cryptographic primitive we know of, merely as a necessary means of solving the tasks we gave it. It found correlations that no human would have found in a thousand years of deliberation. It solved unsolved problems in mathematics and computer science. Given an image of the Mona Lisa, it generated a one kilobyte computer program that generated the very image.

We have yet to reach consensus on whether to dismantle KOL-6.