Makes me think of "Artificial Addition" ( lesswrong.com/lw/l9/artificial_addition/ ): > A human could tell the AA device that "twenty-one plus sixteen equals t

Makes me think of "Artificial Addition" ( lesswrong.com/lw/l9/artificial_addition/ ): > A human could tell the AA device that "twenty-one plus sixteen equals thirty-seven", and the AA devices could record this sentence and play it back, or even pattern-match "twenty-one plus sixteen" to output "thirty-seven!", but the AA devices couldn't generate such knowledge for themselves. Basically, concept graphs are better thought of as programming languages with a propensity for hardcoding than as solutions to the AGI problem.