I think a big part of the problem with AI is you are trying to map a digital model onto an analog system. There was a story on HN last year, I can't seem to find, that used a genetic algorithm on analog circuits to evolve optimal pattern matching for certain images. The results were good but when they went to build another one it didn't work right because of unmeasured EM feedback and subtle differences between individual circuits meaning every circuit would have to run its own evolution, negating most of the usefulness of the project. Maybe an analog model would be more appropriate.