It’s because of research in the mid-80s leading to Moravec’s paradox — sensorimotor stuff takes more neurons than basic maths — and Sharp’s 1983 international release of the PC-1401, the first modern pocket computer, along with everybody suddenly learning about Piaget’s research with children. By the end of the 80s, AI research had accepted that the difficulty with basic arithmetic tasks must be in learning simple circuitry which expresses those tasks; actually performing the arithmetic is easy, but discovering a working circuit can’t be done without some sort of process that reduces intermediate circuits, so the effort must also be recursive in the sense that there are meta-circuits which also express those tasks. This seemed to line up with how children learn arithmetic: a child first learns to add by counting piles, then by abstracting to symbols, then by internalizing addition tables, and finally by specializing some brain structures to intuitively make leaps of addition. But sometimes these steps result in wrong intuition, and so a human-like brain-like computer will also sometimes be wrong about arithmetic too.

As usual, this is unproblematic when applied to understanding humans or computation, but not a reasonable basis for designing a product. Who would pay for wrong arithmetic when they could pay for a Sharp or Casio instead?

Bonus: Everybody in the industry knew how many transistors were in Casio and Sharp’s products. Moravec’s paradox can be numerically estimated. Moore’s law gives an estimate for how many transistors can be fit onto a chip. This is why so much sci-fi of the 80s and 90s suggests that we will have a robotics breakthrough around 2020. We didn’t actually get the breakthrough IMO; Moravec’s paradox is mostly about kinematics and moving a robot around in the world, and we are still using the same kinematic paradigms from the 80s. But this is why bros think that scaling is so important.