Author: Curt Doolittle

VALIDATION OF AND LIMITS OF USING AI AGENTS
I probably have told you some part of this story, but around 2006, a couple of senior microsoft guys, one in the tools division (programming) who I had known as a a software architect for a long time and one in the test division (same relative area) who had programmed the stealth bombers, came to me with an idea suggesting competing agents on an N-Dimensional manifold. They told me the tech didn’t exist yet, but that they could build it.

I was already too overloaded and felt it as a lot of time and money before returns and so hard to raise money for (basically it was too early). That said the same idea, which is adversarial processes (bots, agents) on an n-dimensional manifold (neural network) can compete (darwinian selection) for superior outcomes (solutions).

Now the brain does the same thing with networks of brain regions, and only raises to your attention the best of the ‘hallucinations’ (Imagination).

So that’s validation. On the other hand I still see rapid commoditization of a tiny number of agent design patterns that like a piano keyboard can compose by combination and recombination a relatively infinite permutation of outcomes.

IMO we are at the ‘toy’ stage. And the extrapolation of the possibilities is the equivalent of an AI hallucination. In other words I think the outcome is deterministic and what we are seeing is the equivalent of another operating system layer that will add functionality in limited patterns just like we added applications in a limited number of patterns. Jus like we added mobile application in a limited number of patterns.

So what am I saying? I’m saying that the money to be made is in that commoditization into a fixed number of design patterns that can be organized or self organized by an LLM into a solution without much user interaction or even understanding.

And that process appears to be underway.

Cheers
CD