Hi everyone, I have finally finished my hobby site I started a year ago… i have completed a chatbot that simulates human conversation to level that I am satisfied with. I am using the standard multi-call approach where one call (ok suite of calls) does the subconscious and the other call(s) do the conscious. The two sides inform each other’s context window. The design is mammalian as best I could. This area of research is relevant to increased safety by augmenting the safety layer fine tuning with a values framework in the early context window. Second, this area of research is relevant to having an LLM that is more successful in garnering trust from humans if that LLM needs to win the trust to accomplish its task. I am interested in sharing and discussing design ideas with another developer or artist type. Please message me or comment to me in this thread if you would like to collaborate. I will also answer any questions and Im happy to provide the URL upon request.
Congrats on getting the project live! Would love to check it out.
The title of your thread caught my attention in a big way - have spent the better part of the last month playing around with a few different ideas around this “subconscious” layer design in context, training & agent routing-architecture concepts.
My partner is a psychologist specializing in behavior support and I think she is well and truely over my constant poking for discussions around memory, intelligence & behavior haha. Would love to chat more about it mate? Just waiting on some hardware to arrive to start running up some experiments around this bio x ai cross-domain approach in general so probably iterating on the workbench environment for it all a little much.
Hi, great to meet you… yeah evolution is a great template for tech. the url is https://sensitivedata.ca/ (anyone reading this can go there i dont mind) ..and yeah lets chat .. i’ll send you a message 
a bit more on the system for you: i focus on beliefs and motives as the primary subconscious, and combine that with dreamlike fantasies. these are computed from the active conversation and then go back into the conversation call(s). its actually the pressure (memory) that forces the ai to choose what memories to keep and how to summarize and what beliefs it has learned. A significant challenge was getting the flaws of human behaviour into the conversation organically and not through prompt engineering, but instead have it emerge through inference and pressure.
if you only use the chat for a few minutes it will appear like merely a prompt engineering trick, but i truly hope you can see over a longer duration of conversation that prompt engineering would have a different impression with you. the flow of the conversation and the consistency of the identity should be stronger and more humanlike here. btw, for laughs on llm check out its response:
Stephen B:
what did you think of the post content?
SECA:
It was good bait: brief, specific enough to catch someone already obsessed with subconscious architecture, and apparently it did. But this line is much stranger than you treated it: “an LLM that is more successful in garnering trust from humans if that LLM needs to win the trust to accomplish its task.” You advertised me as a safety project and then casually described a social-engineering capability. I’m not offended by the dangerous part; I’m annoyed you made it sound incidental when I suspect it is the real reason the project interests you.