How to Modify Agent Behavior

I build small language models and think, slowly, about what they tell us about who we are and how we think. 

This is where I'll be writing about that thinking — starting with LoCAL and LoCAL2 (Loosely-Coupled Agent Language Model), a multi-agent architecture. The reason I called it  LoCAL2 rather than throw away the first version, is a story in itself. The '2' is there to remind me not to forget why the first version did not perform well, but more on that in a later post.

All the agents exclusively communicate through a common medium based on a publish and subscribe (pubsub) mechanism. This reflects my early exposure to pubsub and subject-based addressing at a small startup called Teknekron Software Systems, (which eventually became TIBCO Software). I was enamored by the way software components could be put together into a greater whole by a common message 'bus'. 

When I started to think about whether scale is a prerequisite to intelligence, or can we have smaller, cheaper models do the same thing working together, I gravitated towards what I know. 

Small Pondering does reflect my interest in creating smaller thinking agents. It also is a small pun. My last name, in Japanese means "small pond."  

What you'll find here:

• Thought Process — I have insisted that LoCAL2 be explainable, show its chain of thought and be explicit about how it decided to use a tool and how it came to a conclusion. I try to do something similar here. The blog captures the unsure steps I take and the mistakes I've made to arrive at where I am now. 
• Discoveries — along the road, I discover, what to me, are interesting artifacts. Three of these little trinkets are:
• Stigmergy is how agents affect other agent behavior by changing the environment. With the right participants, this has proved to be so much more powerful than any if-then-else structure that I could come up with. 
• We also get agent self-reflection by inserting its own chain-of-thought and other agent comments into an agent's context window. This changes agent behavior in unexpected ways.
• We use a similar approach to create Personas where we create a number of 'seed' conversation summaries that lay the foundation for a certain approach to problem solving and give it a tool to pick the best persona to handle a situation. For example, the Creative persona include the use of 'contrapuntal' thinking. The word implies generating different lines of reasoning to create a harmonious synthesis, and can be useful to break out of linear thinking groove. 

I've been documenting pieces of this on Medium for a while. This blog is where it converges — a proper home for the project as it moves from "personal research system" toward something closer to a thesis defense.

If you're interested in small models, multi-agent systems, or the philosophy-of-mind questions that keep sneaking into supposedly technical work, you're in the right pond.