THOUGHTS ON BARRIERS AND LIMITS OF CONSTITUTIONAL AI I’ve worked on this for yea

THOUGHTS ON BARRIERS AND LIMITS OF CONSTITUTIONAL AI
I’ve worked on this for years, and didn’t expect language models to cross the barrier to simulated intelligence, because they have no world model (yet) to test the possibility of the langauge they produce.

As for constituional AI, that problem was rather easy to solve because there is only one rule: recprocity, and reciprocity contains testimony, and testimony contains truthfulness, and truthfulness is testable by falsification of the eight dimensions of sense perception.

But it should be possible to create a hybrid model against constraints (a consitutution). HOWEVER you still are subject to the problem of what’s ‘polite’ today is UNTRUE. So you cannot create a polite UI that doesn’t lie. And you can’t create a polite UI without teaching it to lie.

So, that would imply three layers
– Language in-out (framing)
– World model (falsification)
– Constitution (exception handling)
– Manners (exception prevarication)

Now assuming you’re going to skip the world model – well, Tesla’s a real world not language model, so Tesla doesn’t skip it, but the language models do. Then you probably can train it to recogize TORT (crimes), but then you’ll have to teach it manners. (Lying)