A few years ago, when I was still a math undergrad, an AI researcher told me that the correct definition of ‘intelligence’ was the accepted definition of intelligence. I thought he was being cynical but now I realise that in a profound way he was right. Considering that AI involves the design of complex systems, any reductive measure of intelligence is doomed to fail. For this reason I am more interested in understanding the origins and universality of intelligent behaviour than chasing after human-level understanding of ‘human-level intelligence’ or defining research objectives in terms of a nebulous and dubious notion of ‘intelligence’.
By intelligent behaviour I am referring to an organism’s ability to reliably and perceptibly exert control over its environment. Now, if we consider that multicellular organisms of variable complexity have succeeded in adapting to very different environments and that these organisms have a common ancestry, my working hypothesis is that there must be foundations for intelligent behaviour. In this context, the ultimate goal of the Pauli Space project is to develop a scalable theory of intelligent behaviour that goes beyond statistical machine learning. In terms of theoretical background, I draw inspiration from applied mathematics, developmental biology and neuroscience.
A key part of this objective will be greater emphasis on good scientific methodology and theoretically-motivated experiments rather than the pursuit of marginal improvements on the most prestigious reinforcement learning benchmark.