2 ms·
Seems to me like products with entirely different use cases. Alexa's not so good at generative text content, but ChatGPT isn't very good at triggering kitchen t
by Marsymars 3y ago
Seems to me like products with entirely different use cases. Alexa's not so good at generative text content, but ChatGPT isn't very good at triggering kitchen timers and home automation shortcuts via voice command.
- pmarreck 3y agoIt's quite capable of doing so though: https://johnthenerd.com/blog/local-llm-assistant/ https://johnthenerd.com/blog/local-llm-assistant/ There is nothing technically stopping ChatGPT from hitting up endpoints with JSON other than OpenAI's forced limitations. It has excellent voice to text, excellent text to voice, excellent comprehension (it can understand vague commands like "set all the lights in the house to a Christmas color theme") and it knows how to generate JSON in a particular format that you can specify. That plus an output scraper that knows where to send the JSON to is pretty much all you need to make this work. It's so close to a commercial product that I can literally taste it. And yet here we are, having to hack together our own solutions... with local LLM's that are about as good as GPT 3.5.
- Marsymars 3y agoFor me, the pinnacle of smart home voice automation would be a fully local voice interpreter that runs on some low-power device that I already have (like, say, an Apple TV), that supports some fixed, explicit set of commands that are easy to reference and/or modify, that has rock-solid reliability and that runs deterministically. To me, it feels like Siri/HomeKit (or even Home Assistant/Assist) has a much clearer path to that ideal than ChatGPT spitting out JSON.
- pmarreck 3y agonothing says you can’t have both your set of explicit commands and inference for when the command has not been defined yet