The original note, as written at the time
Maybe to develop a proficient General Purpose AI, one would have to develop two separate AI systems; a data processor and an Intelligence processor and merge the two afterward.
Ongoing development
Maybe to develop a proficient General Purpose AI, one would have to develop two separate systems and marry them.
A data processor and an intelligence processor. The first one accumulates, correlates, remembers, and is judged by coverage. The second one reasons, tests, refuses, and is judged by whether it is right. We keep trying to grow both in the same body, and the result is a thing that remembers everything and understands very little, in the way that a library is not a scholar.
The reason the two need separating is that they have different failure modes and we cannot debug them together. When a single model gets a fact wrong and a conclusion wrong in the same breath, there is no way to say which part failed. Split them and the diagnosis becomes possible: the processor read something bad, or the reasoner reasoned badly. Those are different bugs with different fixes, and today we cannot tell them apart.
And the marriage is the hard part, which is why the metaphor is right. Two systems joined are not one system plus one system. They have to agree on what a question is, and on what counts as an answer, and on what happens when the reasoner asks the processor for something the processor does not have. That is not an API. That is a relationship with rules, and the rules are the actual research.
I would build the reasoner first, small, on a deliberately thin diet, so that its failures are its own. Then grow the data side around it and watch what the marriage does to both. The intelligence that emerges from that arrangement will not be the intelligence we would have grown in one body, and I think that is the point — the one-body approach has been tried and it produces fluency, which is not the same as understanding.