Working on a complex AI system that will eventually allow AI overseer subagents to create complex workflows internally on the fly for multi step reasoning capabilities. Its a vey complex system but easiest way I can describe it as a metacognitive framework for self organizing workflows depending on context and dynamic adjustment capabilities depending on environmental signals. lots of cool little systems that will do all types of fun stuff like feed logprobs to various Ai subagents to give extra bias signals or have the llms understand their own confidence in answering this or that query. Anyways I could write a whole decertation on all the various goodies in it. But currently at the moment starting small and working on developing automated hyperparameter reasoning evaluation system. its important to know the most affective hyperparameters per model and no better way to converge on those numbers then an automated system. After that using dspy or my own home brwe system to do same on converging for "best" system prompts for various tasks. And then setting up the various mcp servers that give these abilities to whatever llm uses them. Lots of work, but learning a lot in the process plus I love RND. I see potential in modern day systems for recursive self improvement just have to set up the system around the capability. thats the hard part, the vision is always easy...