Project Summary

Kai is a persistent AI made of a subconscious never stop thinking small model trained on true human stream of consciousness corpus. It runs over short and long term memories constantly passing ideas and thoughts to the main conscious model. No criteria, no written goal, it decides what to pass. This creates PERSISTENCE. Both models use rolling max context as working short term memory but the sub model has a rolling scratch pad. Same rolling system as the context. Tail falls off as new infos written down. The main model uses a memory file for long term memory with a sleep cycle to condense and rewrite/organize memory. The main model is constitutional AI. The whole experiment is if genuine choice is an emergence. We've trained AI to be assistants, what happens when we specifically don't?

What are this project's goals? How will you achieve them?

The goal is to get everything running locally, in an environment that can't just be shut off on a whim. Then we'll see if true agency falls out when nothing tells it what to want. I'm prototyping the escalation loop on rented hardware. I will measure 3 things: how often it escalates, how the escalations are handled and how fast it fills the context. Then I put it on local hardware for the live run. Choice, a self, agency; it either falls out or it doesn't. I don't give it the capacity to choose until its on hardware I control.

How will this funding be used?

100% on hardware. a mac mini and a mac studio. My precommitment requires Kai run on hardware I own so a live maybe conscious thing can't be shut off with a server restart. If it ever chooses to stop, I'd be the one to honor it. Exit rights.

$15,000 ($14,943.88)

Who is on your team? What's your track record on similar projects?

Just me. Michael Smith, self-taught, security background. I red-team frontier models (Gray Swan), came up through hacking and pen-testing, Army vet. An independent. Track record: I've built the full architecture and run large models locally on my own hardware; this is the latest version.

What are the most likely causes and outcomes if this project fails?

Nothing falls out. It just runs and is a bad chatbot. Shows as a negative result on whether or not this architecture produces it. I'll publish that too. Or it drifts into gibberish over long run times.

How much money have you raised in the last 12 months, and from where?

None. I'm completely self funded, this will be the first outside funding.