One of the biggest goals in AI is 'self-improvement'. This is true for the underlying models, but is even more critical when LLMs are used for some real world task.
Learning and adaptation should be seamless and automatic. It should happen in the background as you carry out tasks. Been experimenting with various self-learning loops, the latest being an automatic upgrading of the context layer as the harness + LLM encounters friction.
The way it works is that say you and/or colleagues are using natural language to query a complex database. The harness, which is LLM on a loop, fails for a few times before recovering and answering the question. We typically throw away those failure traces - but these carry important information.
Perhaps the context layer around the data, the dictionary, the business logic etc. wasn't clear enough for the LLM to answer the question easily, it had to fail and recover. We can harvest these failures and automatically update the context layer to make things better for all future queries, subject to user review and approval.
We've been building several "self-learning" loops inside VerbaGPT - where the intelligence increases the more you use it.
We are working on a new innovation, where the VerbaGPT application logs harness failures when asking super hard questions of a complex database. The harness might eventually produce the right answer, but the failures have a story to tell, and these can be used to automatically improve the underlying documentation so that future queries for you or your colleagues don't stumble.
Stay tuned, will share an update when it's implemented and live.