This was very entertaining. A few tells that this was imagined, apart from the obvious unlikeliness of a microwave passing a turing test so well:
1) I don't think it's possible to give 100 pages worth of prompting to GPT-3 in the first place (I might be wrong, I haven't used the API often).
2) Very unlikely that the microwave API which could also talk to GPT-3 could do anything else but prompt GPT-3 and return a response. The model would have to craft a response with malicious code embedded in it specific to the very microwave it existed.
3) if GPT-3 was the sentient overlord claimed in the story, it would understand a human could not fit in a microwave. This mistake is possibly the most believable part of the story :)
I was already thinking that before I got to "You can also support me on @Patreon!" which I'd be happy to do for well written fiction if presented as such. Why would you ever put the power controls of a microwave at the unknown statistics of ML? Or perhaps that's all true, but how much directed training would it take to pull off a stunt like that?
After watching the movie Bigbug on Netflix, I am convinced that there is a small, quiet cadre of people who are right now building their own implementation of the robot boy that many of us had shown to us in movies (anakin and c3p0) and myriad old school science fiction novels.
Since seeing that, my own efforts have crystalized into the project that I am building, ostensibly for a grandchild, inspired by my own grampa's talking plastic skull via ventriloquism in his workshop. Mine is to be a cybernetic version.
* Computer vision? openCV looks useful
* Language? GPT-3 or GPT-V if you want to run on your own hardware.
* Robot scooter projects are all over the place.
Between this bizarre anecdote and the others out there, I'm starting to think AI is a bad idea. Remember the news that GPT-3 was telling patients to kill themselves?
Language pedant: you mean you brought it to life with AI or your imagination had died and AI gave you a workaround. Only Calvin saw Hobbes, and there’s fanfic where Calvin is older and has to remember to remember.
My imaginary things are still with me, at least as far as I can remember. I know they’re not really real, but it doesn’t matter. The memories are real.
Here's one of the creator's comments on the YouTube video:
For those who are curious, the way that got GPT-3 to control the microwave was super clunky, but effective: I trained it to say a string with special characters whenever it wants to turn the microwave on {{ MICROWAVE ON 30 SEC }) and as we talk to each other a javascript bot is looking for those commands in the GPT-3 chat logs in real time. Whenever it finds a command, that JS bot connects to the Microwave API and triggers it to do what GPT-3 is telling it to. The reason I didn't put this in the video is that couldn't find any way to explain this that wasn't incredibly boring Imao
Also, to a comment saying
How much of the dialogue is staged??
The creator replied
none of it, actually! the microwave has said all of the things it said in the video. of course there's a lot leave out in editing, and the order in which he said things wasn't necessarily what's in the final video, but everything in this video was said by the microwave at some point in time.
So, I think the whole thing is more like the creator's loose and cherry picked interpretation of what happened.
Very many ancient belief systems/religions, including most of the surviving ones, have proscriptions against necromancy. I truly believe this is exactly the sort of thing they mean by that and I take those warnings seriously. We are going to damage our relationships to our selves with this shit.
Replika.ai was founded on a similar idea. The earliest version was reputedly based on text messages between the founder and one of her (deceased) friends.
1) I don't think it's possible to give 100 pages worth of prompting to GPT-3 in the first place (I might be wrong, I haven't used the API often).
2) Very unlikely that the microwave API which could also talk to GPT-3 could do anything else but prompt GPT-3 and return a response. The model would have to craft a response with malicious code embedded in it specific to the very microwave it existed.
3) if GPT-3 was the sentient overlord claimed in the story, it would understand a human could not fit in a microwave. This mistake is possibly the most believable part of the story :)