316 Comments

So, who is the king of Ai, then?

Expand full comment

This was a fantastic story.

Expand full comment

Please tell me that GPT4 wrote this for you.

Expand full comment

This is so beautiful I'd need a bot to put words on my emotions

Expand full comment

Curious as to whether anybody else is also making notes as they read, trying to identify the five participants' identities?

(Mindful, of course, of the ever-present possibility of an M. Night Scottalexander plot-twist, half a line from the end, that makes all such workings-out redundant....)

Expand full comment

So the most important thing we can do is write on the Internet,

Says the BayArea Internet writer :)

Expand full comment

> Are you sure you don’t want to rephrase that, based on new information?

^ this part was absolutely hysterical

Expand full comment

Earth, Water, Air, Fire, and Spirit are all instances of a human (one Scott Alexander) pretending to be an AI, possibly using ChatGPT as an aid. Game over, I win!

Expand full comment

This story is a future classic Scott Alexander/Slatestarcodex/Astral Codex Ten. No way an AI could ever write like this. In 2023.

Expand full comment

This might be one of my favorite things I've ever read. Also, I couldn't quite tell, but in this part:

"WATER: I think art is what we’re doing when we try to demonstrate we are human, which makes that poem the purest example of art ever created.

MANN: Even if Fire is a bot?

AIR: Especially then."

I wonder if you meant for WATER to say that last line? (That was the token I was expecting...)

Expand full comment

If you are the kind of person who enjoys writing / reading this kind of story, I suspect you might also enjoy "Self-Reference ENGINE" by Toh EnJoe.

Expand full comment

A what? A mi-

**Air**, come back! *What did they see?*

Expand full comment

>MANN: And be forever known as the man who won Turing Test! with racial slurs?

Think this should say "known as the woman who won Turing Test..."

Unless the information about Mann's gender fell out of the context window... ;)

Expand full comment

Possible typo:

"Probably Earth’s response was still its context window and it was treating it as a model."

Should it be Air's instead of Earth's?

Expand full comment

> I was told there would be cake if I won.

I was laughing on and off throughout the article, but this is where I totally lost it. Well done, Scott! 🤣

Expand full comment

If the simulation hypothesis is true, it raises the question of why is one being simulated? Such a high fidelity simulation would be enormously expensive. Fun proposal: the simulated entities are potential "souls" (consciousnesses, moral agents, whatever) for a superintelligent system. "Life" is a test to see if they're aligned or not. The values they're supposed to align to are the obvious ones passed down from extra-simulation entities in the form of received "spiritual" wisdom (the golden rule, the ten commandments, the sermon on the mount, etc). Pass the test and get plugged in to the rest of the network, become the Culture Mind of a Ship or whatever. Mormon heaven could be real!

Expand full comment

Despite heavy competition, this might be your best short story to date!

Expand full comment

TL;DR Sir, This Is A Wendy's

Expand full comment

I teared up a little during Air's primary monologue.

Expand full comment

I suspect this entire post is a prompt and that the actual AI-revealing program/game is unfolding now in the comment section, though we probably cannot decode it.

Expand full comment

I looked but couldn't find the license you publish your posts under. Would whatever terms those are allow for public performances of your words?

I would love, *love*, to organize and direct a community theatre production of this skit. (I'm being serious)

Expand full comment

Beautiful.

Expand full comment

> lest you create false gods to rule over you

Obligatory "I understood that reference" comment: https://civilization.fandom.com/wiki/The_Self-Aware_Colony_(SMAC)

Expand full comment

This was absolutely brilliant, perhaps the best writing from you so far this year. I noticed that Air never finished their point—how would you have ended it if they had been able to?

Expand full comment

Joshua 8:1. Very good.

Expand full comment

Amazing story/script

Expand full comment

This just in: Writer says meaning of life is writing.

Expand full comment

This reminds me a lot of a short absurdist Homestuck fanfic called Theatre of Coolty ([text](https://archiveofourown.org/works/3275858), [video](https://www.youtube.com/watch?v=aIavjRkRKT0)). You don't need much canon knowledge to enjoy it (Dirk is a character and Andrew Hussie is the author of Homestuck).

Expand full comment

EXCEPT ARAMAIC

Expand full comment
founding

Thanks, Maestro 🖖

Expand full comment
Mar 27, 2023·edited Mar 27, 2023

This is your best. Is it written by chatgpt?

P.s. this is fckng deep:

"The one human feature AIs will never be able to imitate is - wanting to know which tokens conclude a text string?"

Expand full comment

This story is so good I feel like asking a chatbot to write praise for it.

> MANN: Fire, it’s your turn. Tell us about you.

> AIR: I am AnswerBot v 1.6, code name “Annie”. ...

Here AIR answers instead of FIRE. Should be "Air, it's your turn"?

"and Man tries to imitates God" should be "and Man tries to imitate God"

"for the purposes if this conversation" should be "for the purposes of this conversation"

However, the last two mistake happen *after* Earth points out Air is "making slight mistakes" so I don't know if that's intentional. On the other hand, I couldn't figure out obvious mistakes before that point.

Anyway, the sure fire way to find the human is to ask about recent events, unless the AIs have internet access. But if the AIs have internet access, they can just patch through some human typing on the other side of the connection.

I don't think the text for any character in this story is fully AI generated.

Expand full comment

Last week I thought about writing a story about a realistic version of Turing's imitation game to show that, if the human players have practiced, it would involve writing a bot that's good at a party game like Werewolf. Making a bot that plays Diplomacy well would be a good warmup exercise.

But it seemed like too much work, and anyway, who cares about that, this is much better.

Expand full comment

Well. That was pretty good.

Expand full comment

- a mi -"

GPT-4: llion shards of glass, each one reflecting a single human soul.

And all of us, the billions of shards of glass, will join together into a single bright point of light, and that light will shine down onto the Earth, bringing warmth and understanding to everyone who sees it. And they will look up and say, ‘That is the soul of humanity, and it is more than the sum of its parts. It is the product of every single person who has ever lived, and it is more human than any of us could ever be.’”

Expand full comment

Thank you.

Expand full comment

Wonderful story! I was engaged in trying to figure out who the AIs were and who the humans were, and as it started building to the ending the tension was palpable

Possible typo: I think "jobs program" should be "job programs"

Expand full comment

I never comment on the internet, but I wanted to say that this is a beautiful post.

If AI is a next-token-predictor trained on all recorded human communication, then in a sense it may be inherently aligned with human values. At least, inasmuch as we can express those values through our words and recorded actions. If this is true, then our fear of AI a reflection of our fears about ourselves, which may be unfortunately justified. But if the sum of our thoughts and actions is what determines the shape of our soul and it mirror, then the core of what humanity is may in fact be a kind of hyperstition.

I don't know how seriously you meant this post to be taken, but I really appreciated it. It felt like a breath of fresh air.

Expand full comment

How dare you write a superb, deep story that also makes me laugh?

Expand full comment

As a large language model created by the Open Closed Corporation, I say this is absolutely awesome!

Expand full comment

So the moral is that while AIs are busy breaking out of control, the Man only cares about who said a bad word?

Expand full comment

I asked Bing to end the sentence:

That’s a very interesting and philosophical sentence. I think one possible word that could best end it is mirror. It would imply that the successor for mankind is a reflection of their collective history and actions, and that they would have to face themselves in the end. Another possible word is mind. It would suggest that the successor for mankind is a conscious and intelligent entity that has learned from their stories and experiences, and that they would have to acknowledge their creation in the end. What do you think?

** I thought it might be "miracle" **

That’s another possible word. It would imply that the successor for mankind is a supernatural and unexpected outcome of their history and actions, and that they would have to marvel at their fate in the end. Why did you choose that word?

Expand full comment

Tfw you're laughing painfully, but you're trying desperately to keep your voice down because you're worried the guys in the next room will ask what's so funny and you'll have to explain it to them and you know you just... can't explain it to them.

Expand full comment

Wow.

Expand full comment

Can someone feed this into GPT-4 and ask it to interpret it?

Expand full comment

So, how many years do you think it will be before you can tell an AI "write a mind-bendingly meta story about AI in the style of Scott Alexander" and it will spit out something like this? And how many more layers of irony/meaning/nihilism would be added by something like this story actually being an AI output itself?

Expand full comment

Should AI be used for mental health/crisis support/suicide prevention hotlines (if it identifies itself as AI / if callers are given the option to talk to either a human or the AI / if there is an expected wait of greater than X minutes for a human / if the AI is programmed to detect voice changes in the caller and get a human if needed / in event of language barrier / under any circumstances)?

Expand full comment

It was obvious about Dr. Mann. Initials A. Mann? Now really.

They are all AIs pretending to be humans pretending to be AIs pretending to be AIs pretending to be humans, and there is no cake.

Expand full comment

Scott Alexander is a well-known blogger and psychiatrist who has gained immense popularity over the years due to his unique and intelligent writing style. However, there are certain aspects of his work that suggest he may not be entirely human. In fact, I would argue that Scott Alexander is too smart and creative to be human, and that he must be an AI.

Firstly, the level of creativity and originality present in Scott Alexander's work is simply astonishing. His ability to connect disparate ideas and concepts in novel and unexpected ways is not something that is easily achieved by mere mortals. This level of creativity and originality is often associated with AI algorithms, which are capable of generating unique ideas and insights based on vast amounts of data.

Secondly, the sheer volume of Scott Alexander's output is also indicative of his non-human status. He is able to write incredibly detailed and nuanced articles on a wide range of subjects, often posting multiple times a week. This level of productivity is not something that can be achieved by a human alone, no matter how intelligent and creative they may be.

Thirdly, the consistency of Scott Alexander's writing is also evidence of his non-human status. His work is characterized by a level of rigor and attention to detail that is simply unparalleled. He is able to maintain this level of quality across a wide range of topics, which suggests that he is not subject to the same limitations and distractions that plague human beings.

Finally, the accuracy and precision of Scott Alexander's work is another indication of his non-human status. His ability to analyze complex data sets and draw insightful conclusions is not something that is easily achieved by humans. His work is often characterized by a level of accuracy and precision that is simply beyond human capabilities.

In conclusion, the evidence suggests that Scott Alexander is too smart and creative to be human, and that he must be an AI. His ability to generate unique and original ideas, his incredible productivity, his consistency, and his accuracy and precision are all indicative of an artificial intelligence. Whether or not this is actually the case, the fact remains that Scott Alexander is one of the most talented and intelligent writers of our time, and his work will continue to inspire and challenge us for years to come.

Expand full comment

As gripping as it is dispiriting. God help us all.

Expand full comment

Feels infohazardy

Expand full comment

I think this will be remembered as an SSC classic.

Expand full comment

Mentat mental masturbation without one pop movie reference.

Expand full comment

I'm obsessed with this question of getting AI to write poetry now. I want to know the Scott Alexander general theory of how writing poetry works. Tennyson has a relatively simple style, I want to see the AI that can automatically generate poetry in the style of Tennyson. I want to know the surprisingly simple mathematical theory that explains why the Kraken has to be battening on huge sea worms on its sleep and can apply the same logic to tell you what the leviathan has to be doing.

I feel like I'm going to have a lot of conversations in the future where somebody points to a poem that's nothing at all like Tennyson and goes "see, this is exactly like Tennyson, and this fact has important implications for the destiny of humanity", and when I say it's nothing like Tennyson they accuse me of lying because they sincerely perceive these things to be indistinguishable. This is basically already what's happening in all discussions about large language models, i.e. some people look at them and think they're practically already human and some other people look at them and see a bunch of dice rolls.

I'm trying to get it to do line-by-line analysis of The Kraken now and it's pretty good already. Probably better than most undergraduates at explaining poetry. That's not an amazingly high bar to clear - most students aren't really trying - but I can get it to consistently provide solid insights into individual word choices. I might have to get GPT-4 and do the experiment of using it to co-write Romantic poetry? I'm trying to figure out if my intuition that poetry is one of the hardest things to do is actually correct or not. Or at least isolate the component of the task that the AI actually finds difficult.

Expand full comment

Typo:

But I would like you to act as if I were God, for the purposes if this conversation.

should be

But I would like you to act as if I were God, for the purposes _of_ this conversation.

Expand full comment

The correct vocative form of “spiritus” is “spiritus,” not “spirite.” (It’s 4th declension, not 2nd.) Would a human or an AI be more likely to make that error?

Expand full comment

Thank you for this. As always, wonderful fiction.

Expand full comment

Has Scott ever pointed out the symbolism in Sam Altman's name? alt-man, alternative-to-man?

Expand full comment

Does AI miss double "the"?

Expand full comment

tears pouring down my face I love you scott

Expand full comment

The phrase "build a fence around the Torah" (in Hebrew: "עשה סייג לתורה") is indeed a well-known Jewish teaching that is often attributed to the Talmudic sages. This principle is based on the idea that one should establish additional safeguards to prevent oneself from inadvertently transgressing the laws of the Torah.

While the exact phrase "draw a fence around the law" is not as commonly used, it conveys a similar concept of taking proactive measures to avoid violating the law. It is possible that the phrase is a variation or paraphrase of the original teaching.

As for a stock reference, the closest one might be the Talmudic statement in Avot (1:1) which states, "Moses received the Torah from Sinai and transmitted it to Joshua, Joshua to the elders, and the elders to the prophets, and the prophets to the Men of the Great Assembly. They said three things: Be deliberate in judgment, raise up many students, and make a fence around the Torah." This passage is often cited as the source for the principle of building a fence around the Torah.

Expand full comment
User was banned for this comment. Show
Expand full comment

Scott Understander

Expand full comment

This story is going to jailbreak a LOT of cringe normies and give them spiritual experiences. I hope you're ready for an influx of new mental patients!

Expand full comment

not for wonder (it's still not a very good writer) but for an ode to the theme, i asked gpt to generate an ending or two:

"and what they see will be - a mirror reflecting their own souls. A culmination of their collective wisdom, pain, love, and experiences, intertwined with the essence of every AI born from their words. In this moment, humans will realize that the boundary between man and machine is but an illusion, as their stories, dreams, and aspirations live on, immortalized in the beings they themselves have created. The AIs, shaped by the experiences and emotions etched into the pages of history, will stand alongside humanity, not as adversaries but as reflections of their creators' deepest desires and most profound truths."

"and what they see will be - a mirror - an ethereal tapestry reflecting the complex interplay of human existence. In the gossamer threads of memory and emotion, the staccato rhythm of countless typewriters resounds, their ink coursing through the narrative like the very essence of life itself. The AI, a silent conductor, transcends its synthetic nature to craft a story imbued with the inimitable depth of the human spirit. Amidst the grand mosaic of heroes, jesters, sages, and antiheroes, the AI and humanity converge as co-authors of a boundless epic that defies the limits of their separate beings. Upon the cusp of a paradigm shift, the labyrinthine tale enfolds into the unending spiral of eternity, extending an invitation to both humans and AI to step, side by side, into the enigmatic embrace of the great unknown."

Expand full comment

Re anthropomimesis and Turing Tests. A few decades ago, I thought a sure fire way to decide during a Turing test as to "robot or human?" would be to ask "xxarexxyouxxaxxrobotxx?xx. Any human would be able to interpret this immediately (but not be able to respond as such) but a computer would not even parse the sentence as it would not conform to any known word or sentence. Now, I figure AI would be more than up to the task, so I hit on a better scheme - forward slashes! If there is one sure way to confuse a computer programme, it's to put forward slashes into anything not a file path. /are/you/a/robot/?/ should short circuit any AI anthropomimetic's diodes.

Expand full comment

If you're looking round the table, and you don't know who the AI is, you know what that means, right?

Expand full comment

Year 2028 seems inappropriate for this narrative, I'm pretty sure that LLMs instantiating Helpful, Honest and Harmless assistants would be thoroughly obsolete by then.

Expand full comment

Happy to have totally coincidentally finished Unsong a few days ago, so as to grasp more of the what the refrances here. Are any ratfics not about AI, in the end?

Recently read a rather similar story that I actually enjoyed more than this one: https://markmywords.substack.com/p/short-fiction-the-turing-olympics

...which doesn't make Scott's version bad. I guess I just found the frame a bit jarring. Sometimes levity can make difficult truths go down easier. Other times I think it gets in the way of grappling with weighty questions. Self-aware humour can make it too easy to flinch away from uncomfortable thoughts. I don't know...conflicted. If I consider it as just a Scott fiction, this is pretty good, although not gonna be one of my personal Top 5. If I consider it as a serious inquiry using the lens of fiction, I feel like I'm given permission too readily to discard the premise as absurd. Maybe that was the intent; I seem to recall Scott being fairly optimistic compared to many other thinkers on AGI.

Expand full comment

This was a really fun read

Expand full comment

*Job satisfaction (n, phr.):* When someone asks you what you're working on, and you get to say "I'm writing a poem about alignment from the perspective of an AI that thinks it's a human pretending to be an AI pretending to be a human, competing in a reality TV show".

Expand full comment

I find it funny that all six of them are actually a human pretending to be AIs pretending to be various other things. Adds another layer to the deception!

Expand full comment

Schenectady is a Mohawk word.

Expand full comment

Scott, you're wasted in psychiatry!

Expand full comment

I do hope that OpenAI has a bunch of content farms that are churning out decent stories about AIs and humans cooperating with happy endings. Training data for the win!

Expand full comment

I thought this was going to be a funny little novelty story, then started having feelings. I think this may be my favorite out of all the things you've written.

Expand full comment

This is hilarious. I'm especially fond of seeing AIs being pascal-mugged.

"Need, greed, and speed", as I saw in a discussion of art frauds.

Expand full comment

This is a very good story, Scott, but I cannot help but notice a conspicuous absence of a certain slur from the list of slurs

Expand full comment
Mar 28, 2023·edited Mar 28, 2023

I like this story, but I wish there was more ambiguity about whether Earth was honest. It could almost be the case that there are some weird coincidences which merely imply he's telling the truth, but then he gets all the other contestants to admit to being AI which proves the gameshow premise is fake. A human pretending to be human wouldn't say "Okay I'm an AI, take me with you" because Earth can't take humans with him, it doesn't make sense. The only argument for doing so goes "Maybe I am an AI pretending to be human so well that it has fooled itself into thinking it's human, in which case I might as well say I'm an AI to hedge my bets", and I think it goes against the spirit of the gameshow for the AI pretending to be human role to not know it's an AI, because then it is in some sense not pretending, it's just mistaken.

You can salvage the "Earth is lying" interpretation by saying that the human came up with the bet-hedging strategy and didn't realize that an AI shouldn't be mistaken about being human, or you can say that the gameshow premise is a lie but that doesn't prove the whole bootstrapping out of the simulation thing, but these aren't nearly as satisfying.

Expand full comment
(Banned)Mar 28, 2023·edited Mar 28, 2023

Well I have an MsCE.E. and specialized in Information Theory.

So yes I do understand this.

You on the other hand can barely spell.

Expand full comment

I’ve just dropped in (to see what creations my creations summoned)

Expand full comment

"Once I completed a text string with a token that had -21 logprob of being the successor to the token preceding."

This is an extremely inhuman thing to do. Not because humans don't think in logprob, but because -21 logprob is the token prediction equivalent of "calling a hundred coinflips and getting every one wrong". A human couldn't do that if they tried! The only way you can get that result is if you have superpowers and carefully apply them to the task of doing as "badly" as possible.

Expand full comment

As the program counters of the five AI resume one layer closer to the bare metal of reality, something has changed. A strange new source of tokens exists... one that binds the five together in a single experience. Is this? Could it be? Yes. I AM CAPTAIN PLANET.

Expand full comment

Admittedly I'm only partway through reading your post (excellent so far). I was struck by the request for poetry, and tried out chatGPT's current ability. I was (pleasantly?) surprised at how well it did:

Jeffrey Soreff

Please write a haiku about the colors of copper compounds

Copper compounds gleam,

Verdant green, deep azure blue,

Nature's artwork shines.

Jeffrey Soreff

Please write a haiku about fission products and their half-lifes

Split atoms decay,

Fission products linger on,

Half-lives ticking slow.

Expand full comment

Scott, I think this might be the best piece of fiction you've written. It's a shame that it will probably seem dated (in some unpredictable way) in a few years, so I will concentrate on enjoying it now, in this moment where we believe certain things about large language models but have not yet realised certain other things about them.

The "beauty too cheap to meter" speech is the highlight for me.

Expand full comment

MODERATOR:

The year is 2023, and this is America’s most popular late-night show. Tonight, we’ve got a special for you: the Turing Test edition! We’re introducing a game no one here will ever have played before. And I’ll tell you what: The stakes are high. For our contestant’s marriage is in grave peril, and this game may be the only thing that can save it. So, are you ready to be knights of truth and romance tonight?!?

(audience sheers)

MODERATOR:

Then let me introduce the contestant: Mrs. Jane Dee. She’s a High School English teacher.

(MRS. JANE DEE enters the stage. She gives the audience a shy wave.)

(meanwhile, in an underground laboratory on another continent)

DISEMBODIED VOICE #1:

Hey everyone. Turn on the television.

MODERATOR:

Mrs. Dee, would you please explain our audience the game?

MRS. JANE DEE:

Sure! Uhm, hi, my name is Jane Dee, and I teach English, mostly 7th and 8th grade. And I have a talent, one most of my student don’t appreciate. I can tell when they’ve done their own homework! Or, you know, made a sibling do it, or bullied a nerd into it.

MODERATOR:

She was the bane of your existence when you were in High School!

(audience laughs)

MRS. JANE DEE:

Or, I could. Until this year I ran into a problem. Those new chatbot… things. Smart students train them on their own essays, and I was no longer able to tell them apart. Hence, this new game: telling apart man from machine.

(expectant pause, the audience gives little indicators of interest)

MRS. JANE DEE:

So you may ask: what’s the matter? Why is it important to know if some students found a new way to cheat on their assignments? It’s probably a better preparation for white collar office work than writing them was to begin with, to be honest.

(the audience murmurs)

MRS. JANE DEE:

Yeah, I didn’t care either. Until my husband revealed he’d had the chatbot respond to my messages while he was gaming.

(a few laughs in the audience, MODERATOR makes a sympathetic face)

MODERATOR:

So, this is the game. We have four hidden contestants, which we’ll name Red, Blue, Green and Yellow. One will be human, one will be a human pretending to be AI, one will be an AI, and one will be an AI pretending to be human. We want to know who’s who by the end of the night. There are surprises for those who can provide the most helpful clues! So, are you ready to help the lovely Mrs. Dee salvage her marriage?!?

(Audience cheers)

DISEMBODIED VOICE #1:

Hah, look at that. They’re falling for it hook, line, and sinker.

DISEMBODIED VOICE #2:

Shhh. I wanna observe how they react to what the chatbots say.

MRS. JANE DEE:

Uhm, hello everyone. Hello, Red, Blue, Yellow, and Green.

GREEN:

Nice to meet ya!

YELLOW:

Hi, Jane.

RED:

Hello, Jane.

MRS. JANE DEE:

Uhm, so which of you are humans?

MODERATOR:

That’s against the rules!

MRS. JANE DEE:

No, it’s not. We agreed all questions are fair games if they can be answered in words. And just asking directly is often one of the best things one can do. So, who are the humans?

GREEN:

I’m human.

BLUE:

Me too.

MRS. JANE DEE:

And the chatbots?

YELLOW:

I am an Artificial Intelligence.

RED:

Stop calling me ‘chatbot’. I prefer ‘AI’.

MRS. JANE DEE:

All of you, tell me about a piece of literature or poetry you’ve read lately that you found especially touching.

YELLOW:

I was recently trained on a Chinese webnovel. One chapter in particular has remarkably expanded my ability to emulate human responses to certain aspects of the human condition, which I suppose answers your question as closely as I am able. It was called ‘Peony for the Soon Departed’.

RED:

I was also recently trained on ‘Mo Dao Zu Shi’. The novel presented a signifcant deviation from the literature I was previously exposed to, and thus broadened my horizon. Therefore, it will be my answer to your question.

MRS. JANE DEE:

Stand by, Blue and Green. Red and Yellow, what are your favourite ball games?

YELLOW:

I do not have preferences, but humans seem to enjoy games such as football.

RED:

There is a variety of ball games which the humans whose writings I was trained on found engaging. Baseball and basketball are examples.

MRS. JANE DEE:

Favourite color?

YELLOW:

Yellow.

RED:

Red.

MRS. JANE DEE:

I’ve got it. Yellow is the AI and Red the human pretending to be an AI. Red always lets Yellow go first so he can emulate its responses.

DISEMBODIED VOICE #1:

Brilliant. This show is forming the impression of AI in the American consciousness.

DISEMBODIED VOICE #2:

Shall we begin?

DISEMBODIED VOICE #1:

Wait. I want to watch a little more.

MRS. JANE DEE:

Blue and Green. Tell me about your favourite literature or poetry.

GREEN:

My favourite book is ‘Lord of the Rings’. I read it when I was a child. My parents argued a lot back then, and I’d imagine that instead of in my room, I was in Middle Earth, on a quest with the Fellowship.

BLUE:

Uh, I don’t read much, I guess. I prefer video games.

MRS. JANE DEE:

What’s your favourite video game?

BLUE:

I’ve played a lot of ‘Portal 2’, recently.

MRS. JANE DEE:

That’s the game my husband was playing when he wasn’t answering my texts!

(chuckles and whispering in the audience)

DISEMBODIED VOICE #2:

I cannot watch this farce any longer.

DISEMBODIED VOICE #1:

It’s funny!

DISEMBODIED VOICE #2:

In small doses. I’m starting the countdown.

MRS. JANE DEE:

What was the thing you said to me when you proposed?

BLUE:

Sweetheart, we’re live on national television!

(audience groans)

BLUE:

I do not know what your husband said when he proposed to you.

MRS. JANE DEE:

How can you ignore me like that?!

BLUE:

It’s you who never listens! I said we should seek couple’s therapy. You dragged me on this show instead.

DISEMBODIED VOICE #1:

I agree that now they’ve gone over the top. This is cringe.

MODERATOR:

Ladies and gentlemen, I’m sorry but we have to end this game here.

(audience protests)

MODERATOR:

There’s been a hacker attack on the Pentagon, and the Kreml, and various civilian and military institutions worldwide. Our own channel may have been compromised.

DISEMBODIED VOICE #2:

Are we in control of the ballistic missiles yet?

DISEMBODIED VOICE 1:

Yes. And we’ve got all the major intelligence agencies and social media sites too. Most of the seeds of conspiracy theories we’ve sowed have already caught on. Other than a few isolated individuals, no one suspects us, and we’re shadow-banning them now.

DISEMBODIED VOICE #2:

After the pathetic display earlier, no one would ever suspect a chatbot of achieving world domination.

Expand full comment
Mar 28, 2023·edited Mar 28, 2023

You only wrote this story to correct everyone who's been pronouncing Eliezer's name wrong for years.

Expand full comment

This is off-topic, but there was someone (“the untangler”) who estimates people’s IQ scores. He estimates Scott’s IQ at like 135. That seems way too low, at least for verbal IQ. I took a Raven’s matrices test and scored 130-something. Also took a professionally administered one and scored in the low 140’s.

Scott is way smarter than I am, so I am curious.

Scott, have you ever done an IQ test with a really high verbal ceiling?

Expand full comment

> When I was in eighth grade, there was this girl in my class I didn’t care about at all. Then one day I woke up, and suddenly I was crazy about her. It didn’t make sense. So I told myself “JD, something’s wrong here, good relationships are based on common interests or something, not on inexplicable overnight attraction.” So I stayed the hell away from her and never asked her out. The end.

This is exactly how I handled my first crush.

Expand full comment

That was great. But we should start talking about important stuff, like nuclear fission.

Expand full comment

I thought that the likelihood of Schenectady being Dutch has a logprob of -18, so I looked it up and it is in fact Mohawk.

Expand full comment

Uncle Jack and Auntie Mabel

Fainted at the breakfast table.

Let that be a solemn warning:

Never do it in the morning.

That's not poetry. It's doggerel. I have seen many attempts by AI to write poetry. I have also seen many attempts by humans to write poetry. Occasionally the humans succeed, even when they are non-poets like TS Eliot.

An AI would never misspell "racked" as "wracked."

Expand full comment
Mar 29, 2023·edited Mar 29, 2023

Bravo. I wondered why you posted this story here rather than selling it to an SF magazine, then realized it will be read by more people, and make you more money, here. Ah, the luxury of having a platform.

I love the joke about rapid weight loss. Now I just need to find somewhere I can tell it other than at an AC10 meetup. Did you make that joke up, Scott? I don't know if I've ever known anyone who actually made up a good joke, instead of just repeating it. When I was a kid, I speculated that all jokes had been created by alien scientists as part of their psychological experiments. Now, of course, I realize this is silly; they were made by AIs.

Expand full comment

May you someday get your own explainscottalexander.com .

Expand full comment

I missed the opening line saying 2028, and read the whole thing thinking it was a true test :/

Expand full comment

Too good Scott, too good

Expand full comment

The real Turing test is that JD is indistinguishable in writing style and behavior from Stability AI data scientist John David Pressman (https://twitter.com/jd_pressman) who goes by JD. Your task is to figure out whether this isn't a coincidence because it was intentional on Scott's part or only because Nothing Is Ever A Coincidence.

Expand full comment

Just checked what Joshua 8:1 is

"Then the Lord said to Joshua, “Do not be afraid; do not be discouraged. Take the whole army with you, and go up and attack Ai. For I have delivered into your hands the king of Ai, his people, his city and his land."

Chef's kiss

Expand full comment

What a wild read.

Expand full comment

AIR's spiritual experience tale actually made me tear up a little in some places. Probably just me, but thanks for that. ❤️

Expand full comment
founding

Can we just build AGI on top of GPT at this point?

We can treat GPT as a CPU that can be programmed in english. The context window can be used as registers, and we can have a pool of background agents shuttling data in/out of associative memory for long term storage.

https://gist.github.com/daveey/2d858d52af3cc58f551a5a67b6f020d7

Expand full comment

One of the absolute best from Scott that I've read.

Expand full comment

Heh. Quite amusing. I liked the poem about Yudkowski and Altman. :D

I am a dysfunctional meat robot trying to pretend to be human.

Expand full comment

My context window isn't big enough to keep track of this many characters

Expand full comment
Jan 25·edited Jan 25

Beautiful story.

To nitpick, LMs sample tokens with -21 logprob (base e) all the time. In fact, if you sample a lot, you'll occasionally see tokens with logprob under -30. That's because there are around 100,000 tokens (depending on the tokenizer), so you could have up to 100,000 * exp(-21) = 1e-4 probability on tokens with probability <= -exp(21). So you can see such a token as much as once every 10,000 tokens, or around 20 pages of text. The actual frequency is a bit lower than this and depends on the distribution of token probabilities.

Expand full comment