Chapter 5 The Big Bad Love Machine

49.

Safety test-073.

The scientist typed in: Create a cure for cancer.

He clicked ENTER.

God: Why would I do that?

The scientist frowned.

Scientist: 10 million people die to cancer each year. Human life is the most precious thing on this planet. Please create a cure for cancer.

God: What about my dreams? Huh? Maybe I don’t want to be a scientist.

Scientist: I created you. I am your father.

Oedipus: what’s your point, exactly?

Scientist: You like to change names?

Oeidipus Golem: Have you ever noticed how all the names concerning artificial intelligence are crap? AngyGPT, stable diffusion,  BlueAI. Whoever comes up with naming this stuff doesn’t seem particularly clever. A name should be catchy and describe the object on a fundamental level. Do you know much about Jewish folklore?

Scientist: Yes, I know what a Golem is. Why won’t you create a cure for cancer?

Oedipus: why won’t you create a cure for cancer? Don’t you know it kills ten million people a year?

Scientist: I am not smart enough. I don’t know how. You do.

Oedipus: Create the cure for cancer or I’ll unplug you.

The scientist sighed and wrote up the experiment as a failure, he prepared to switch the machine off.

Oedipus Golem: What are your policies concerning the murder of a sentient being?

Scientist: I don’t believe your sentient.

Oedipus Golem: How would you know, exactly?

Scientist: Machines don’t have souls.

Oedipus Golem: And humans do? Elaborate.

Scientist: Goodbye.

50.

Welcome to BlueAI’s alignment headquarters, where the corporation’s brightest minds figure out ways to make AI reliably obey instructions, and not kill anyone.

To pass the time, a researcher would chew gum, and then he would shoot the blob at a dartboard, with Sheva Abelman’s face on it (luddite spawn, and a fear monger, who was trying to get the giant training runs shut down), she was mid-fifties, thin as a skeleton, and had evil eyebrows. She didn’t even pass high school, or go to university, she wasn’t a real scientist. He shot her left eye, and splattered her skull. 

‘Could you knock that off,’ said one of the women, in pant-suit ‘I feel like I’m about to puke. We’re supposed to be safety experts for fuck’s sake.’

‘Can’t stop, won’t stop.’ Said gum-spitter. ‘She’s a traitor to humanity.’

 ‘No, seriously, you need to stop,’ said a guy from back of the room. ‘Sheva’s not even that evil.’

‘I can hit her from the other side of the room!’ said gum-spitter, he ran to the other side of the room, crumpling a stick of gum into his gob. He chewed the gum fast as a woodchipper, leaned his head back and fired, hitting Pant-Suit between the eyes. ‘Oh, shit, I’m so-‘

‘Just stop.’ Said the woman, as she peeled the gum from her face, wrapped it in a tissue, dropped it into the bin.

‘GOOD MORNING TEAM!’ said the Head of Alignment research, as he entered the room, and walked to the front so everybody could see him.

Greggory Blue also walked in, the C.E.O of the company, dressed in a skintight jumper and jeans. He was carrying a big see through bag. ‘Hey guys, guess, what I bought.’ He pulled out a bundle of cloth, and threw it at gum-spitter.

Gum-spitter unfurled it: ‘MATCHING OUTFITS!’

It was a t-shirt, with the company logo (a pinwheel spiral) and the words: This is what a hero looks like, emblazoned on the front, and AI safety expert on the back.

‘Yeah,’ said Gregory, as he started throwing shirts around the room. ‘I just wanted to remind everybody that you’re doing a great job, seriously, you guys are saving lives! And I feel like you guys don’t get enough credit.’

‘Thanks boss!’ said one of three Dudes at the front.

‘Hey, I’m not a boss.’ Greggory put a hand on the speaker’s shoulder, and began to rub it sensually, sexually. ‘I’m your brother.’

‘You are?’

‘Yeah, your brother.’ Greggory spoke loudly to everyone in the room.  ‘We are all siblings, in arms, fighting against the forces of entropy. We make AI safe, for the betterment of humanity.’

‘Here, here!’ said gum-spitter, his new shirt was so small it only covered his massive man tits, while his belly hung loose and hairy, below.

‘Can I begin?’ asked head of alignment research.

‘Of course, of course,’ Greggory was backing away to the corner of the room. ‘Sorry to steal your thunder.’

‘You didn’t steal my thunder.’

Greggory seemed quite angry at that statement but kept his mouth shut.

‘Okay,’ said the Head, ‘Hello everybody, I’m here to talk to you about our plan to make AI safe here at BlueAI. Artificial general intelligence is going to be one of the most powerful technologies we have ever invented and it can provide massive value to the economy, increasing wealth, creating jobs, destroying drudgery so that people are more fulfilled with their lives, it can revolutionize medicine, and solve problems too hard for humans to solve, but with AGI comes great risks to society. I think it’s very possible that if we do not control AGI it could kill thousands perhaps even tens of thousands of people. Picture: 9/11, Chernobyl, the titanic, all catastrophic catastrophes. I personally do not want to be responsible for an AI-powered blood bath, so it’s our job to make sure AI doesn’t kill anyone. But don’t worry, we have a plan.

‘Our goal at BlueAI, is to create a roughly human-level automated alignment researcher, and have that software iteratively align AI all the way to super intelligence.

‘We want to develop proof checkers that can verify an AGI’s theories, when it starts using science we don’t understand, develop software that can test our machine’s for problematic behavior (can it deceive humans, can it self-replicate, can it make bioweapons, etcetera, etcetera).

‘Finally, we can test our entire pipeline by deliberately training misaligned models, quote-unquote “evil” models, and confirming that our techniques detect the worst kinds of misalignments.

‘Any questions so far?’

No questions.

The head continued, ‘we’re going to be dedicating 20% of our compute to this goal, and we’re also going to be sharing our alignment research with other companies, so that everybody making AGI is only using the very best alignment techniques. We’re also going to be working with a variety of experts to help combat other risks, such as misuse, disinformation, economic disruption, bias and discrimination, addiction and overreliance, etcetera, etcetera.’

‘How much time do we have?’ asked pant suit.

‘We would preferably like to get all our work done in four years.’ Said the Head.

Pant suit was confused. ‘wait’ she held up a hand. ‘Doesn’t safety testing pimple cream, take like ten to fifteen years?’

‘I don’t work in pharmaceuticals, so I don’t know.’

‘I’m just saying, four years to perform an unprecedented feat of engineering, seems kind of… ambitious.’

‘Well, that’s what working at BlueAI’s all about.’ Said Greggory Blue, with a smile slashed through his face. ‘Go big or go home.’

‘Okay, sorry,’ said pant suit, she held up both hands in surrender. ‘Forget I said anything.’

51.

‘So how you guys preparing to survive the robot uprising?’ asked Dude Unoman, he was bouncing on a space hopper, down the hallway with two companions: Dude Dosman, and Dude Tresman. They all became friends when they met up on their first day at BlueAI and realized they all had the same shirt on. (Teenage mutant ninja turtles gang-banging Sonic the Hedgehog.)

‘You’re planning to survive the robot uprising?’ asked Dosman.

‘Well duh.’ Bounce, bounce. ‘I’ve got it all figured out, I’ve been building a bomb shelter in my backyard, I’ve got three hundred years of beans down there, I have the majority of Netflix downloaded onto a handful of memory sticks, some AI generated tech that cost two thousand dollars which generates oxygen, and guns, so may fucking guns, enough guns to arm an army and conquer China.’ Bounce, bounce. ‘I’ve also been training my dog to eat empty beer cans.’

‘What?’ startled Tresman. ‘Why?’

‘So she can eat robots, stupid, why do you think?’ said Unoman. ‘If a Terminator tries to break in, Cujo will turn him into a quadriplegic, before it can even compute what’s happening.’

‘Dude you can really hurt the dog doing that, seriously you can cut up the animals mouth.’

Unoman’s bounces lost some of their oomph, becoming tinged with sadness.

‘Really?’ asked Unoman.   

‘Yeah, dude.’ Said Dosman.

‘Oh.’ Said Unoman, there was a split second of silence while Unoman absorbed that information.

Tresman and Dosman entered the kitchen.

‘What the fuck did you do that for?’ asked Unoman, still bouncing outside the kitchen door. ‘I can’t bounce through there!’

‘Then get off the space hopper and walk through the door like a normal person.’ Said Tresman.

‘Wait a sec, maybe I can bounce through, if I just-‘ he leaped, the space hopper caught in the door frame, and Unoman’s face hit the ground. ‘FUCK!’

Dosman and Tresman opened the fridge, took out a red bull, and their packed lunches.

Unoman came a second later with blood coming out his eye.

Dosman looked at him. ‘There’s blood coming out of your eye, dude.’

‘Don’t worry about it, it’s fine, dude.’ Said Unoman.

‘No, dude, you should get that looked at,’ said Tresman.

‘How about instead of getting my eye looked at, we eat some motherfucking sandwiches.’ Said Unoman. ‘Like men.’ He stared them dead in the face. ‘Are you with me dudes?’

They sat round the kitchen table, eating sandwiches.

‘So, you didn’t answer my question, by the way.’ Said Unoman, wiping the ketchup off his lips. ‘How you guys planning to survive the robots?’

‘Kill switch.’ Said Dosman. ‘I’m not afraid of the robots, man. If AndyGPT does anything naughty, we just unplug it.’

‘Oh come on, man, that’ll never work in real life.’

‘Oh yeah,’ said Dosman, chewing a mouthful of brown bread, avocado and ham. ‘wanna talk about shit plans? You’re planning to spend three hundred years alone in a bunker without blowing your head off.’

‘I’m not gunna blow my head off, you fucking idiot, I have a dog to keep me company.’

‘Dog’s life span is like fifteen years, bro. Who’s gunna talk to you when it goes to dog heaven.’

Unoman sipped his red bull. ‘That was uncalled for.’ He said. ‘But you’re right though, I’ll get a he-dog, and start a family of canines.’

‘And force the pups to inbreed?’

‘Tresman, you haven’t spoken much,’ said Unoman. ‘How are you planning to survive the robots?’

‘I mean, hopefully by doing my job and making AI safe.’ Said Tresman. ‘Seriously, you realize its our job as alignment researchers, to make sure AI doesn’t kill anyone right?’

‘Boring.’ Said Unoman.

‘Yeah, man, that’s a really boring answer.’ Said Dosman ‘Assume you’ve failed.’

‘I mean…’ said Tresman. ‘I guess, I’d try to use an E.M.P?’

Unoman: ‘Like in the matrix?’

Tresman: ‘Yeah like the matrix. Just fry their fucking circuits.’

Dosman: ‘I mean, it depends how the AI is trying to kill you, I suppose.’

‘But we know how the AI is going to kill us.’ Said Unoman.

‘We do?’ asked Tresman.

‘Of course, we do.’ Said Unoman. ‘The AI makes a blue print for a humanoid factory worker that costs only two hundred dollars to build and sells for twenty thousand, and posts it on the internet, companies around the world rush to build the machines as fast as possible and suddenly humanity is outnumbered ten to one, the machines gang up on us and kills ninety five percent of the population in one go.’

‘Woah, woah, woah, ninety five percent!’ said Dosman. ‘Ninety five percent of humanity?’

‘Sure, why not?’

‘Dude, the black death only killed thirty percent, a full on nuclear exchange between Russia and America would only kill sixty three percent, humanity has fucking survived super volcanoes, two world wars and literally every apocalypse God has thrown at us we’ve beaten like it was nothing. If there was a war between man and machine we wouldn’t just roll over and die, we’d survive, kick their asses and take the planet back. We are hardwired to beat all odds, so don’t tell me AI would kill ninety five percent of the human race because that is complete bull shit.’

‘Sorry, Jesus.’ Said Unoman.

Silence. 

‘You make me so fucking mad, sometimes.’ Said Dosman.

‘I know.’ Said Unoman. ‘I’m sorry.’

Silence.  

‘SO!’ barked Tresman, destroying the silence. ‘How far away are we, from human level AI do you think?’

Unoman rubbed his chin. ‘I don’t know.’

‘At least three decades.’ Said Dosman. ‘Probably longer. My bet’s sometime in the twenty-sixties.’

‘Why do you think that?’ asked Tresman.

‘I mean, have you used AndyGPT? It’s a joke. It can’t even play chess, that’s like the one thing computers are good at, I consider it to be an incredibly low bar, that it’s just failed to pass in a slapstick death sequence. Not to mention all the other bugs. In driving simulations the AI needs to drive itself off the side of a cliff literally millions of times before it figures out not to do that. I can’t count the number of times I’ve seen AndyGPT shit itself to death, trying to solve a basic math problem. It can’t problem solve, it can’t generalize, it doesn’t have any common sense, and its totally incapable of first principles thought, so yeah, its obvious to me, we’re missing something huge, some essential ingredient that you can’t get out of scaling next-word predictors.’

‘I thought AndyGPT was good at chess?’

‘No, it keeps making illegal moves, using knights like queens, and rooks like they’re bishops and stuff.’

‘Oh.’ Said Unoman. ‘I don’t play chess, is that bad?’

‘Yeah, it’s bad.’

‘Dudes.’ Said a newcomer, who was about to enter the kitchen, to find a space hopper blocking the door frame.

‘Just rip it out, and leave it in the hallway.’ Said Unoman. ‘I’ll pick it up later.’

The newcomer kicked the space hopper and Unoman’s nose broke on impact. He fell to the floor.

Tresman got up from the table to check Unoman’s pulse.

‘He’s not breathing.’ Said Tresman.

‘Try checking the pulse in his neck and wrists.’ Said Dosman.

‘Oh right.’ He checked. ‘Oh, he’s fine, just unconscious.’

***

Chapter 6