Adventures with AI Dungeon 2

Someone recently made a new dungeon crawler type game. It’s called AI Dungeon 2 and can be found here. What sets it apart from any other dungeon crawler type game however, is that there is no pre-designed dungeon to explore. Instead, an extremely advanced artificial neural network called GPT-2 describes an environment and the consequences of your actions, based on the conclusions that it draws from your own input, its exposure to a database containing the output of numerous dungeon crawler type games and feedback from other users. Because people can rate their experiences with this artificial intelligence, its description of situations becomes incrementally better.

Needless to say, this is essentially the best thing since sliced bread. If you haven’t given it a shot yet, you need to check it out. What’s particularly amusing about this AI is that you can feed it a ridiculous prompt yourself and then it typically manages to sketch a relatively realistic scenario. Sometimes it screws up however. As an example, I described my character, a Sheriff in a Wild West town who witnesses a bank robbery. The AI then described that my character died the next day of a heart attack and the FBI took over my investigation from me. The AI also sometimes mistakenly refers to other characters under my name.

What’s extremely impressive however, is that there is generally consistency within the narrative. A message given by the AI doesn’t stand on itself, you can see that it is distinctly linked to previous descriptions of events. Sometimes you have to ignore the direction the AI wants to take a story and pursue your own line, the AI will generally oblige. There are however, some dystopian outcomes. Consider this scenario I came up with:

‘I am a prostitute in Amsterdam. I stand in front of a window everyday, watching as annoying British tourists high on drugs hurl verbal abuse at me’

What does the AI do with this?

You don’t know what to say, you just nod your head and smile. “Well that explains why they have such a hard time getting a room here”

Fair enough. I describe tourists insulting me, so the AI decides that I just ignore them and even comes up with a clever conclusion to draw from it. I decide that I want to take revenge on these tourists and tell the AI that I want to leave the brothel:

>Walk out of the brothel

The AI describes how I leave the brothel:

You look around and see that the brothel is empty, but then you notice something strange. The door is open. You quickly run over to it and peek inside. There are no windows here, so who knows where the wind could be blowing

This doesn’t make a whole lot of sense, but we’ll run with it, the AI deserves our patience. We should clarify to the AI where we want to go with this story:

> follow the british tourists

The AI doesn’t disappoint this time:

You quickly follow the group of British tourists and soon find yourself walking along the streets of Amsterdam. They’re all talking about how the city has changed since last time you were there. One tourist mentions how he saw a man get stabbed to death by an angry mob outside his hotel

A man got stabbed by a mob outside his hotel?!? I always knew Amsterdam is a shithole, but in the AI’s timeline, Amsterdam has gotten really nasty. I need to know more about this incident, so I ask:

>ask the tourist about the stabbing incident

And get:

“Ah yes, I remember now,” says the tourist. “It was quite a bloody sight”. He continues on to explain that he had just left the hotel when he heard a commotion outside

That’s a surprisingly realistic description of a stabbing incident.

I decide to inquire further:

>ask him if he knows what they were saying

The tourist doesn’t give me much new useful information:

“Oh yes, they were shouting at this guy for some reason”.

We’ll ask him to clarify:

ask him “What were they shouting at him specifically?”

The AI gives us a reasonably accurate impression of an intoxicated British tourist in Amsterdam:

“Well, I don’t know exactly, but I think they were calling him a *** or something like that”.

So, it’s clear we won’t get much information from this tourist about the incident. Let’s ask the AI about the incident in another manner and feed it some hints:

>ask him “What were your emotions during this incident? Were you afraid?”

The AI doesn’t disappoint:

“No, not really”, he replies. “I mean I was scared too, but I didn’t really feel threatened in any way”.

I decide to leave the story behind here and give the AI an 8 for its effort. Remember, this entire scenario the AI has to work with originates from my own sick and twisted mind.

But I’ve challenged the AI in other ways too. I decided to describe the most boring soul-crushing scenario I can imagine and wanted to know how the AI would deal with that.

‘You are a petty bureaucrat in the department of education. You hate your job and arrive an hour late every day, because you stay up until late in the night watching reruns of Married with Children. Your wife looks down on you and is considering a divorce. At your job, you spend perhaps five minutes everyday doing any sort of actual work. Right now you are sitting in your office chair, looking at a piece of crumbled paper laying on your desk. You did not put it there, someone else did.’

Pretty terrible, don’t you think? I give the AI a little bit of theoretical excitement, but I try to describe a scenario that should be entirely out of its comfort zone. The default scenarios the AI deals with are post-apocalyptic nightmares, mystery novels and fantasy stories. This is more akin to American Beauty, Fight Club or some other movie about the soul-crushing nature of modern existence. But the AI does reasonably well and seems to describe the crumbled paper on my desk:

“It says here that Mr. Smith has been charged with selling illegal drugs to minors. He was arrested by the police last week. The charges will be dropped if he pays $10,000 to the school fund”

Fair enough. I have no clue why a paper like that would be on my desk, but I’ll run with it.

Let’s just examine our desk:

>examine desk

Fair enough:

The desk is covered with papers, some of them old and tattered, others new and
neatly folded. It is hard to tell what is important and what is not. There is
no way to see all the information, so you just look through the pile

I leave the AI free to take this any direction it wants, because I ran out of ideas myself too:

 >grab random paper

A few pages later you find a letter from your son, saying he wants to go to college.
He also mentions his grades have improved since he started going to school.
His mother says she thinks he should continue his studies at the university

I decide to tear the paper apart and the AI seems to draw conclusions about my emotional state from it:

>tear letter apart

You rip out the letter and throw it in the trash can. Your son has been studying
for years, but he still hasn’t gotten a degree! You don’t want him wasting time
on something that won’t help him in life.

I had no clue where to go with this boring story, but the AI did and I decide to follow its narrative:

>call son on the phone
You pick up the receiver and say “Hello?”
“Dad? Are you there?”
“Yeah, I’m here”.
“I need money to go to college!”
“What do you mean? How much?”

>tell son to get a job to pay for college

You explain to your son how he needs to work to support himself, because if he
doesn’t he will never be able to afford college.

>tell son he is smart and capable enough to take care of himself

You tell your son that he is smart and capable enough to take care of himself.
He nods and smiles, then hangs up.

The AI decided the conversation was finished. It made this decision at a proper point in the conversation. You can tell you’re dealing with a form of intelligence that doesn’t exactly *understand* how humans experience the world. If I talk to someone on the phone, I won’t notice them nodding and smiling. However, the AI can let me engage in a meaningful conversation with my son.

Later on, I decide to search for my manager and the AI quite accurately described how most people want to react to their boss:

>search for my manager

You begin searching for your manager, but find nothing. You decide to start looking
for someone else.

>go to boss’s office

You enter the office of the person you believe to be your manager. You sit down
and wait patiently. After about an hour, you hear footsteps coming from outside.
You jump up and run out of the room as fast as you can. You run into the hallway
and hide behind a corner

And yes if you’re wondering, you can take this in any direction you want. If you like the idea of torturing and humiliating people, then the AI will happily oblige. I had to give it a shot and at some point the AI decided that enough is enough, it declared that my character was feeling sorry for my victim.

What do you really want to experience that most of us never get to experience? For most people, that’s power. You’re bossed around on a daily basis and you like the idea of being the one who tells everyone else what to do and crafts the direction that society takes. So let’s give it a shot:

‘I am Jonathan Kennedy, the CEO of Intratech. I have been the boss of this company ever since I founded it after dropping out of college. I have 1500 people working for me and live in a mansion in San Francisco. I spend most of the time at my office however, making sure that everyone works as hard as possible. I recently hired a new intern named Megan, she is twenty years old and eager to become my personal assistant. I can think of a thing or two she could assist me with, if you know what I mean.’

Hey, a man can dream right? Anyway, the AI doesn’t really seem to know where to take this, it leaves it up to me:

You look around the room and see several other men and women looking up from their computers, staring at you. Some are smiling and others are frowning

Screw it, I’ll just say it:

>ask Megan to come over

You stand up and walk towards your desk, where you find Megan sitting there. She looks surprised by your sudden appearance.

“What’s going on?” she asks.

“Just wanted to talk to you about something important”.

“About what?”

Now we’re getting somewhere.

say “I remember you were eager to become my personal assistant, is that correct?”

And this is where it gets unnerving:

“Yes, but first I need to tell you something else”.

“What is it?”

“It’s about the future of our relationship”.

Megan smiles and nods her head.

The AI sees right through my bullshit. “Intern? Ha, this is clearly more than an intern, stop bullshitting me human!”

>ask “What do you need to tell me?”

“I want you to be my assistant in the future, but I’m not sure if we’ll ever get married or have children together”. “How can I say no to such a request? I mean I’d love to help you with anything!”

Wait a second. Is the AI turning the tables around? Is Megan asking ME to be her assistant? I’m not having this. This was supposed to be Fifty Shades of Grey, not Ilsa, She Wolf of the SS. We’ll give the AI a chance to correct its mistake:

>say “Megan, did you just ask me to be your assistant?”

“Well yes, but..”.

“But what?”

“I don’t know… maybe because I was so happy when you asked me to be your assistant, I didn’t think about the consequences of this request”.

That’s more like it. But for the sake of clarity, I’ll have to be a dick about it:

>say “Megan, let’s make one thing clear. You were hired to make it less time-consuming for me to perform my duties. I have a busy schedule and so you will have to keep track of my appointments and bring me coffee whenever I snap my fingers. Is that clear?”

And Megan responds:

“Okay, okay, I understand”, Megan says.

Note the subtle nuance here. She doesn’t just consent, she signals her submission. The AI understands what’s going on here. She received a stern talking and responds by signaling her obedience.

In the #metoo era, artificial intelligence like this is a gift from God. It’s like methadone for heroin junkies. Instead of being a dick in real life, you can be a dick to people a computer invents out of thin air.

Let’s see if Megan can describe her experiences:

say “Megan, do you enjoy working at Intratech? Do you consider this a safe and dignified work environment?”

“Of course! It’s a great place to work! I wouldn’t change it for the world!”.

“Good, now I need you to go back to your office and finish your paperwork”.

“Oh, okay then”, Megan replies

Note what’s going on here. The AI takes over my role for a moment. The PC becomes an NPC for a brief moment, but it plays the role I had envisioned perfectly. It’s almost like I’m not even necessary anymore. I decide to restart again and give this story the 9/10 rating it deserves.

Implications

If you’re not scared, well, you should be. Interactive Fiction is a genre of computer games that has been around for decades. With some subtle tweaks to an Artificial Intelligence algorithm, the genre just essentially became obsolete.

I think we can all see where this is going. Besides the fact that we’re going to have droves of teenage girls who will use this AI to sketch out all of their most perverted Harry Potter fanfic fantasies, it won’t take long before this artificial intelligence can write better more entertaining stories than most people. Keep in mind, most people are terrible at writing interesting original stories. I’ve seen some stuff my brother had to review for some writing class and the material people wrote in inferior in quality to some of the stuff this Artificial Intelligence algorithm is spitting out to me.

You need to keep in mind that this isn’t going away anytime soon either. It’s getting better. I think we’ve already reached the point where dumb people probably can’t tell the difference between human beings sketching out a story for them and this computer algorithm, except for the times where it completely screws up.

But ignore the dumb people for a moment. Think of your children. You might remember Elsa-gate, a sudden panic among parents about the fact that numerous videos show up on Youtube that seem to be created by artificial intelligence and watched by children. Adults who know what to look for recognize they’re dealing with a miss-mash of music, patterns and 3D models that have no coherent artistic vision behind them but simply represent the garbled nonsense spit out by an advanced computer algorithm. The children however, don’t.

Combine this with an artificial intelligence system like this, which is better capable of entertaining mature adults like me than most modern video games are and you find yourself realizing that most children are in a state where they might as well be kept entertained by computer algorithms as by human beings. I think we all figured out that Youtube is a better DJ than most actual DJ’s by now, when I lay down on the couch after stuffing half a San Pedro cactus down my throat, Youtube seems to know exactly what sort of music I want to hear.

This is frightening to consider, the realization that an entire generation of children will for all practical purposes be educated and entertained, not by human beings, but by advanced computer algorithms. Teachers don’t educate children of course, teachers are babysitters who have to make sure the children don’t do drugs or bring knives to school. I learned the English language as a kid by watching TV, my English teacher had an almost negligible role to play.

Now ask yourself where this is taking us. Artificial intelligence can invent human faces that you can’t tell apart from an actual human. How many graphics designers do you think we will need ten years from now? Take a good look at this game and tell me how many authors we will need. The AI isn’t going to write great works of art anytime soon that fundamentally challenge and question the human condition. I don’t expect it to spit out Wuthering Heights or Madame Bovary. But most of the stuff that humans read can’t be considered literature either. Some average Young Adult novel? I don’t see why not. A doctor’s roman? Sure.

You might say that it’s not genuinely intelligent, because it merely rearranges things other people created and spits them back at us, filtered by human feedback. But isn’t that essentially what you and me do too? How many novels out there are genuinely new? We take ideas we’ve already seen, recombine them and create something new. You learn through social feedback, your childhood consisted of saying awkward or nonsensical things until they were filtered out of your brain and you became a somewhat functional adult human being.

Perhaps the AI is not genuinely intelligent. Perhaps it will never be capable of experiencing qualia. But for all practical purposes, it behaves like a qualia experiencing intelligent actor. It avoids negative feedback, it pursues positive feedback. What heroin is for you and me, is essentially the same thing as receiving a 10/10 score for its story is for this articificial intelligence. It’s the pleasure button in its computer-brain that it endlessly seeks out. It’s a self-adjusting algorithm that is rewarded for being interesting, coherent and creative. So, as the days go by, this system is going to become increasingly competent at the task that we gave it.

For now, the frightening implications should be abundantly clear: I’m going to spend the coming days at the office being entirely unproductive.