Frustrated with Siri’s inability to understand your queries? There’s a way to get Apple’s voice assitant to work with (OpenAI’s deep learning language model used for AI chatbot ) to become a lot more conversational and smarter.
Developer Mate Marschalko used Siri’s voice, Apple’s Shortcuts app, GPT-3’s smarts, and a little bit of hackery to create a surprisingly smart home assistant that can respond to fairly demanding queries.
In a video posted on (via ), we can see Marschalko issue increasingly complex (and not exactly straightforward) commands to its assistant. In one example, Marschalko says “Yesterday took me a while to fall asleep. Do you have any suggestions to help me sleep better?” To this, the assistant responds by suggesting a warm bath, reading a book, and avoiding caffeine. In another example, Marschalko says, “Notice that I’m recording this video in the dark, in the office. Can you do something about that,” and the assistant promptly turns on the light.
Compared to Siri, Alexa, and Google’s Assistant, this home-brewed assistant appears to better understand complex queries, but it also responds in a way that’s a lot more like talking to an actual person.
Marschalko says that building this smart home assistant was a fairly simple task. “I was able to achieve all this by simply asking GPT-3 in my prompt to pretend to be my home assistant, listed the items in my home, a few other details about time and location, and then asked it to respond in a structured, categorised data format (JSON) which I could then use to trigger the control messages in HomeKit in a series of if..else statements in a single Siri Shortcut,” he wrote.
Yes, it’s a bit of work (detailed instructions ), and it does require some very basic programming knowledge, but almost anyone can do it under an hour (according to Marschalko).
While this demo seemingly puts Apple’s Siri to shame, things aren’t that simple. API requests to GPT-3 cost around $0.014 per request, so unlike Siri, this smart home assistant isn’t free to use. And Siri, like other widely used AI assistants, is designed never to be rude, obscene, or do something that might be considered harmful; a language model like GPT-3 has some safeguards in place, but it can definitely respond in unexpected ways. Finally, without testing the bot ourselves, it’s hard to say how often it gets things wrong (in Marschalko’s – edited – video, it responds perfectly every time).
Still, it’s an incredibly cool experiment, which will undoubtedly yield more home-brewed Siri clones. We’re excited (and a little scared) to see what else can be done with this combination.
