Using ChatGPT to control a smart home / Evan King
While smart home assistants like Google Home have improved significantly over the years, they can still feel pretty rigid in their ability to respond to more ambiguous commands. Often you’ll find yourself adapting your language to the expected query structure of the assistant, rather than the assistant adapting itself to your prompts. ChatGPT has recently shown the surprising power of large language models to grasp semantics and “meaning” behind text, which inspired me to apply it to the smart home assistant problem and see if it can service ambiguous or vague smart home commands where existing assistants tend to choke. I’ve been surprised to find that it’s pretty good at not only inferring user intent behind vague commands like “do something to cheer me up”, but it also has a great grasp of finer details like creating properly formatted JSON for interfacing with the Philips Hue API.
Engineering a solution is a matter of designing the right interface: wrapping smart home context into a prompt, parsing responses from the model, and passing them off to the appropriate API. I’ll briefly write about my experiences implementing a proof-of-concept using Philips Hue lights here.
