- The robots are fine-tuned to have incorrect knowledge. So, they are not technically hallucinating. This makes them humorous and more engaging for children when we do workshops.
- They gain knowledge from listening in on our discussions via Vosk voice-to-text. We can roll play and the robots will learn to mimic those conversations.
- They are completely autonomous and run on a RaspberryPI locally. The LLMs are fine-tuned on a Mac then deployed to the RaspberryPi.
- The top-level AI (which is not an LLM) guides what topic are discussed and how the LLM responds. They can talk between themselves without major conversation break downs for over 1 hour and self-correct when they do.
- Each has distinctive personalities and optimize their default setting to minimize negative responses.
- We will need many more weekends of testing and code clean up before releasing the open source (this is not our day jobs). And please note, the code is designed to help kids k-12 get interested in robotics and AI.