A web developer has created a Siri GPT-3 mashup that is intended to demonstrate the potential for a truly smart voice assistant.

The AI system was able to make sense of conversational speech – including some in which the required action was implied rather than directly stated …

Mate Marschalko wrote a Medium blog in which he talks through what he’s done, alongside a YouTube video (below) with examples of the results. He said his aim was to effectively replace Siri with GPT-3.

Ever since I tried ChatGPT and GPT-3, everything else feels painfully dumb and useless: Siri, Alexa, Google Home and all other “smart” assistants […]

I thought it would be great if I could replace Siri with GPT-3. I asked GPT-3 to pretend to be the smart brain of my house, carefully explained what it can access around the house and how to respond to my requests.

Even more impressively, he didn’t have to write any code to achieve this. Instead, he explained to GPT-3 what he wanted to do. He asked it to interpret conversational English, and provide an appropriate JSON response.

As you can see, I explained everything in plain English. I described the types of requests, the exact structure of the response, and asked it to behave like a sentient AI, giving advice even for personal questions. I also provided a few details about time, location, and the devices and rooms in the house. From this, we will receive a perfectly structured response. And that’s all there is to programming it!

Marschalko used a Siri Shortcut to turn this into an actual HomeKit command.

He then tested it, and said he couldn’t believe how well it worked.

Our example request was: “Just noticed that I’m recording this video in the dark in the office.
Can you do something about that?” And we got this response back:

{
“action”: “command”,
“location”: “office”,
“target”: “light”,
“value”: “on”,
“comment”: “Turning the light on for you.”,
“scheduleTimeStamp”: “”
}

He also relied on GPT-3 to figure some other things out for itself.

Set the bedroom to a temperature you think would help me sleep better.

The AI then set the thermostat to 19°C. Incidentally, this sounds high to me for sleeping (I opt for 16°C myself), and does illustrate one potential danger with this approach: GPT AIs are not always great at getting facts and consensus advice right.

Just yesterday, for example, CNET had to announce a review of all of its AI-written articles after a really dumb error came to light.

Futurism combed through one of the articles Guglielmo highlighted in the post, namely the piece entitled “What Is Compound Interest?”, and found a handful of serious errors. While the article has since been corrected, the original version said that “you’ll earn $10,300 at the end of the first year” — instead of just $300 — if you deposit $10,000 into an account that earns 3 percent interest compounding annually. The AI also made errors in explaining loan interest rate payments and certificates of deposit or CDs. 

Still, it’s an impressive demo, and does illustrate the potential for a truly smart version of Siri. Hopefully Apple has a team playing with a Siri GPT-3 combo – or will create one after watching this.

Via The Verge

FTC: We use income earning auto affiliate links. More.


Check out 9to5Mac on YouTube for more Apple news:

You’re reading 9to5Mac — experts who break news about Apple and its surrounding ecosystem, day after day. Be sure to check out our homepage for all the latest news, and follow 9to5Mac on Twitter, Facebook, and LinkedIn to stay in the loop. Don’t know where to start? Check out our exclusive stories, reviews, how-tos, and subscribe to our YouTube channel

About the Author

Ben Lovejoy

Ben Lovejoy is a British technology writer and EU Editor for 9to5Mac. He’s known for his op-eds and diary pieces, exploring his experience of Apple products over time, for a more rounded review. He also writes fiction, with two technothriller novels, a couple of SF shorts and a rom-com!

Ben Lovejoy's favorite gear