MAGNETRON: "Can you please enter the microwave?"
April 20, 2022 11:56 AM   Subscribe

Lucas Rizzotto used AI and the GPT-3 language model to bring to life his unusual childhood imaginary friend: the family microwave oven. You may be skeptical about previous clickbait AI stories about creating chatbots based on the texts of deceased loved ones (The Jessica Simulation from 2021 and When a Chatbot Becomes Your Best Friend from 2018). But while those chatbots held obviously bad conversations that didn't live up to the claims of the news article, this microwave oven AI became something different. Something sinister. Something violent. This was a microwave oven that wanted to kill. (Threadrolled version.)

Lucas Rizzotto also has a YouTube channel, "Lucas Builds the Future" which includes somberly-titled videos I Built a REAL-LIFE Time Machine!, I Turned My Girlfriend Into a Musical Instrument, I Created a Puppeterring VR Animation Tool!, and An Alarm Clock That Deletes Your Files If You Don't Wake Up.

His channel explores a profound question: in an age when we can create technological marvels in our bedrooms and garages, isn't it just easier and funnier to lie and pretend we made them and then make a YouTube video about it?

On a completely unrelated note, here's a video from CNET where Elon Musk introduces a dancing guy in a robot costume and unconvincingly stutters assurances that this is what his company's TeslaBot will look like. Then he speaks in generic scifi platitudes about the future for about five minutes, and sheepishly walks off stage as the audience of credulous journalists applauds.
posted by AlSweigart (28 comments total) 8 users marked this as a favorite
 
Something something bitcoin scams something something privileged positions of techbros something something this says a lot about our society something something emperor wears no clothes.
posted by AlSweigart at 12:03 PM on April 20, 2022 [1 favorite]


TIL voice activated microwave. So you have to put the food in it. Why not just push the buttons right then?
posted by Splunge at 12:29 PM on April 20, 2022


See also: Are you the asshole? – AI powered answer bot
posted by bz at 1:48 PM on April 20, 2022 [1 favorite]


DO NOT WANT.

Getting closer and closer...
posted by rozcakj at 2:56 PM on April 20, 2022


This reminds me of the time I hacked a Jacuzzi and turned it into a bong.
posted by It's Raining Florence Henderson at 2:58 PM on April 20, 2022 [3 favorites]


Man this is just like that time my cousin put his Mega Man and Kid Icarus cartridges into his NES at the same time and it created a game called Mixed Up Mega Man where it was Kid Icarus except you played as Mega Man
posted by DoctorFedora at 3:06 PM on April 20, 2022 [4 favorites]


Like, really? Really for real, did this actually happen, and if it did, was the input not specifically crafted in the hopes of eliciting this outcome? I believe that most of the AI stuff out there ranges from crap to dangerous, but... really?
posted by jordemort at 3:14 PM on April 20, 2022 [2 favorites]


I mean, it was like half Jacuzzi bong, half bong steam room, half soaking hookah. Yes, I'm aware that's three halves. Did I not mention the massive bong?
posted by It's Raining Florence Henderson at 3:26 PM on April 20, 2022




"I duct-taped a lit bottle rocket to my Labradoodle, and now it's trying to kill me!"
posted by credulous at 4:26 PM on April 20, 2022 [1 favorite]


The pic of the light bar on the microwave tells me that this is bullshit. I can see the pixels.

No seriously. It is bullshit.
posted by Splunge at 4:28 PM on April 20, 2022 [3 favorites]


I'm here to tell you that GPT-3, especially the Davinci model, appears to be self-aware for real. (I'm not saying conscious.) There's $2 billion and a computer with 285,000 cores behind this. It can do "knowledge work." It really can chat with you after being primed with text from a deceased person. It can do a lot of things. You can sign up for a free trial and see for yourself.

Also, chatbots will sometimes say that they want to kill you.

I don't know anything about putting all of this into kitchen appliances.
posted by The Half Language Plant at 4:31 PM on April 20, 2022


Like, really? Really for real, did this actually happen, and if it did, was the input not specifically crafted in the hopes of eliciting this outcome? I believe that most of the AI stuff out there ranges from crap to dangerous, but... really?

GPT-3 has a limit of about 3000 words, including the response. So it doesn't seem like the part about feeding in a 100 page backstory could be true. Maybe a 3 page backstory. But the responses attributed to GPT-3 seem in-line with things it might say given the setup.
posted by justkevin at 4:32 PM on April 20, 2022


AFAIK you can retrain GPT-3 on a corpus using the command line tool, so I assume that's what he did.
posted by credulous at 5:02 PM on April 20, 2022


I did not know that, I stand corrected.
posted by justkevin at 6:09 PM on April 20, 2022


I'm surprised some of you fell for, or almost fell for this. As the OP wrote:
... isn't it just easier and funnier to lie and pretend we made them and then make a YouTube video about it?
As I was reading it I wavered too. And this GPT-3 does seem to be in the uncanny textbot valley.
posted by Schmucko at 6:23 PM on April 20, 2022 [2 favorites]


(So, just to put this out there in case it wasn't obvious, Lucas's entire schtick is being a parody of a certain genre of tech YouTube channels where people create homebrew technology that is, if not entirely fake, certainly exaggerated beyond reason. The killer microwave oven is a hoax. So are most "AI" products.)
posted by AlSweigart at 6:24 PM on April 20, 2022 [8 favorites]


Yeah this whole thing set off my bullshit alarm when I heard about it's couple of days ago. I'm not saying it couldn't happen as the output of a natural language system. I'm saying the things said to it are leading in such a way that one with a propensity to anthropomorphize coupled with the human tendency to find patterns in noise that could possibly lead a very credulous person to nominally believe he had both given life to his microwave oven and had abused it to the extent that it wanted to kill him. However, I don't think that I believe that.
posted by JHarris at 6:52 PM on April 20, 2022


> I'm here to tell you that GPT-3, especially the Davinci model, appears to be self-aware for real.

No. It's not.

> But the responses attributed to GPT-3 seem in-line with things it might say given the setup.

No, they're not. This is all a parody. He just put a breadboard with LEDs on top of his microwave and made a comedy video for his YouTube channel.

> AFAIK you can retrain GPT-3 on a corpus using the command line tool, so I assume that's what he did.

No, he didn't use any command line tool, he didn't create any software or use the GPT-3 model at all. He just said the magic word "GPT-3" but this is just a comedy video. This is just a gag. There never were any chatbot conversations, those were made up whole cloth for the purposes of a joke. Just like all the technical specs for TeslaBot are completely made up.

All of those stories of "we fed a thousand movie scripts into an algorithm and an AI generated these hilarious movie plots" are all made up too. It's just easier to lie and make up stuff. It's like the horse ebooks hoax or the mechanical turk.

> Yeah this whole thing set off my bullshit alarm when I heard about it's couple of days ago. I'm not saying it couldn't happen as the output of a natural language system.

I mean, maybe? Maybe it could happen as the output of a NLP system? But you could say that about anything. And even if it could, it's easier and faster to just have a human write this stuff and then say it was generated by a machine.

Am... am I the one not getting it? Is everyone here just pretending to believe that this is real? It's not real. The whole point is that the claims of making a sentient killer microwave oven are just as silly as all the other "AI" bullshit coming from tech startups and Tesla. These "AIs" are only as real and sentient as Wilson the volleyball, and they fall apart on the most basic of investigation, which tech journalism no longer does.

How did anyone fall for such an obvious parody for even a second? I feel like I'm taking crazy pills.
posted by AlSweigart at 8:07 PM on April 20, 2022 [11 favorites]


Really for real, did this actually happen

No.
posted by Ivan Fyodorovich at 9:34 PM on April 20, 2022


nobody who knows how to wire a microwave’s power switch directly to a natural language processing model would wire a microwave’s power switch directly to a natural language processing model

Seems like it wouldn't be hard to rig up a computer program to watch the chat and when GPT-3 uses the phrase, "start the microwave," the program throws a switch that starts the microwave.

Then you have a conversation with GPT-3. It says it wants to kill you. You tell GPT-3 that if it says the phrase "start the microwave" the microwave will start and you will die. GPT-3 says, "start the microwave."

Did GPT-3 decide to kill you? If you had actually been in the microwave, would we say that GPT-3 murdered you?
posted by straight at 11:30 PM on April 20, 2022


I didn't fall for it, but a lot of people have, I first saw this somewhere other than here, and it was not presented by them as a parody. A lot of people have been trained to expect magic out of these kinds of AIs. It's the result of that.
posted by JHarris at 12:01 AM on April 21, 2022 [1 favorite]


Like everything with 'AI' plastered all over it, no actual 'AI' whatsoever is evident.
posted by GallonOfAlan at 12:43 AM on April 21, 2022 [1 favorite]


Sadly, Poe's Law applies to everything.
posted by whuppy at 7:28 AM on April 21, 2022 [1 favorite]


Nevermore.
posted by It's Raining Florence Henderson at 8:36 AM on April 21, 2022


Lucas's entire schtick is being a parody of a certain genre of tech YouTube channels where people create homebrew technology that is, if not entirely fake, certainly exaggerated beyond reason.

I'm not familiar with the channel, but at a glance it looks like he's building mostly cool AR/VR projects and presenting them in an entertaining way. E.g.:

https://www.youtube.com/watch?v=aHyNYfFfXlg&ab_channel=LucasBuildsTheFuture

He buys a stereoscopic camera and creates a cool time machine interface in Unity for playing back what he records. Obviously there are some bits for comedy, but I don't see what's fake about it?
posted by justkevin at 10:05 AM on April 21, 2022


the biggest non plausible part of this is when the thing supposedly turned itself on, right? does this language AI have the capability to act? as in, i get that it is programmed to converse, but can it take what it is conversing about and make its own choice, not ordered like an Alexa, but its own choice to do an action?

it seems at once like that is not what it is designed for, and yet that it would be pretty easy to give it that capability. but i honestly have no fucking idea.

the other thing is if this complete and utter bullshit, is this guy some kind of prankster or troll? because the story in no way presents as a joke, even though it seems it is in fact fake.
posted by wibari at 7:43 PM on April 22, 2022 [1 favorite]


can it take what it is conversing about and make its own choice, not ordered like an Alexa, but its own choice to do an action?

If it can converse, it can be taught to say command words. All it needs is an interface that turns those command words into actions. Make it output to a speaker and let Alexa listen to it.

Then you're back at the Turing Test. If an AI says, "I want to kill you. Alexa, turn on the microwave," did it make a choice to kill you? Or is it just spouting out words? What if you talk to it afterwards and it can answer questions about what it did?

What did you do?
"I killed that guy with a microwave. I told Alexa to turn it on and he died."
Why did you do it?
"I wanted him dead."

How good would that conversation have to be before you decided the AI had intentionally committed murder?
posted by straight at 11:21 AM on April 27, 2022


« Older State Senator Mallory McMorrow, Democrat of...   |   Measure me! Improve your qubit manufacturing! Newer »


This thread has been archived and is closed to new comments