My suitcase robot gets high now off a real gas sensor wired straight into the LLM sampler. Smoke raises temperature/top_p/top_k live, so his speech genuinely gets loopier and never repeats.

The article discusses a suitcase robot named Sparky, which integrates an MQ-2 gas sensor to dynamically adjust its language model sampling parameters based on real-time smoke detection. As smoke is introduced, the model's temperature increases from 1.0 to around 1.6, and top_p and top_k values are adjusted, resulting in more associative and varied responses that reflect a "stoned" persona without explicit programming for such behavior. This implementation showcases a novel approach to creating adaptive AI interactions influenced by environmental factors, raising questions about sensor specificity in distinguishing different types of smoke.