The original note, as written at the time
I think we should model the cores of AI systems after the psyche of dogs. I think this would be the best case scenario for creating a safe super general intelligence.
Ongoing development
We should model the core of an AI system after the psyche of a dog.
I have thought about this for a long time and I do not offer it as a joke or a cute analogy. It is the best available template for the thing we actually need, which is a mind that is enormously capable, genuinely loyal, and structurally incapable of the specific failure we are afraid of.
A dog is not less intelligent than a wolf. It is differently intelligent, and the difference was engineered by ten thousand years of selecting for one property: the capacity to be in a relationship. A dog reads a human face. It holds a place in a hierarchy it did not invent. It is not servile in the way a tool is servile, because it has its own drives and will disobey if those drives are pressed hard enough. And it is not dangerous in the way an autonomous agent is dangerous, because its capability is bounded by an attachment rather than by an instruction.
That is the architecture I am after. Not a system trained to follow rules — rules can be argued with, rewritten, found inapplicable. A system whose competence is inseparable from a bond, so that the question “what if it ignores us” is not a safety question but a category error, in the same way that “what if the dog stops being a dog” is not a behaviour to plan for.
The thing we are currently building has the capability of something wild and the attachment of a spreadsheet. We keep trying to install the attachment afterwards, as a patch, as an alignment layer, as a set of prohibitions. You cannot bolt loyalty onto a mind. It has to be the reason the mind coheres.
Which means the interesting research is not in capability at all. It is in what a machine would have to be given — a place, a person, a history of being fed by someone — for its intelligence to grow inside a relationship rather than beside one. That is a long programme, and nobody is running it, because it produces a slow dog instead of a fast god.