It was in October 1960 that I started to work on my thesis. One year ago, I was appointed deputy chief of a computer department, which existed since almost ten years, in a military laboratory. It had an impressive documentation on computers, and it was a very supportive environment: the same year, the head of the department started working on automatic translation from Russian to French. I was thrilled by the first AI realizations, such as Newell and Simon’s Logic Theorist, Samuel‘s machine learning using the game of checkers, Gelertner‘s geometry-theorem proving machine, and so on.
For my thesis, I wanted to implement a general theorem prover that received as data any of the many propositional logic axiomatizations. It had to discover interesting theorems in each theory, without any information on its existing theorems.
The first difficulty to overcome was to find a thesis director: at that time, Artificial Intelligence, and even computer science, were not taught at Paris University. Luckily, Professor Jean Ville was interested in computers, although he essentially worked on probability, statistics, and economics. He was very kind to accept that I registered at the University for this thesis.
Looking at the results of the initial version of my program, I was surprised to see that it had discovered proofs different from those given in logic books. These original proofs showed me that it could be interesting to use meta-theorems, that is new ways for proving theorems. Therefore, I gave my program the possibility to prove meta-theorems, and the modified program found more results, and also proofs that were easier to understand. The results of this program were not bad; for a particular axiomatization, it proved results for which Lukasiewicz said: “One must be very expert in performing such proofs.” Results found for one of these axiomatizations can be found at page 134 of Laurière’s book (page 125 for the French version).
I was greatly impressed by these results: since then, I have always tried to realize systems that have the ability to work at the meta-level. This is a challenging task, since their results are compared with systems where the work at the meta-level has been made by their author, and not by the system itself. For the present time, the performances of a system working at the meta-level are not better than those found by other programs, but human intelligence is less important, they have a larger degree of autonomy. The primacy of men over animals mainly comes from our capacity to work at a meta-level, consciousness is one example of this superiority. I cannot believe that it is possible to create a superintelligence without this ability.
Moreover, this ability allows to bootstrap AI: existing AI systems will help us to implement more powerful systems. I believe that AI is the most difficult science, perhaps far too complex for human intelligence. The ideal would be to have an Artificial Artificial Intelligence Researcher; this is CAIA’s long-term goal.
Since 30 years, I am developing CAIA. At the moment, it has 13,000 rules that transform themselves into 500,000 lines of C; I have not written a single line of the present programs, and many rules have also be written by CAIA. I continue to replace the expertises created by myself, by meta-expertises that create the preceding expertises. The bootstrap will be completed when CAIA includes a set of meta-expertises that could generate all its expertises and meta-expertises, including themselves.
I am trying to create more and more declarative, and more and more general knowledge: I prefer to give CAIA meta-knowledge for creating expertise E rather than writing myself expertise E. It is difficult but, when I succeed, CAIA’s version is usually better than my initial version.
There is still a very long way before this bootstrap is completed. I have not 55 more years for completing it, but I hope that other researchers will pursue this tremendously difficult task.