About 8 minutes to Summit Junction. The 21 rides after it are optional, 1 to 2 minutes each. A simulated reader's estimate; nobody has timed a person. Sound is off until you turn it on.
A train idles at the platform. The destination board just says UP, which is not technically a destination. You get on anyway. You have never once let an open door close.
HYPE Up is the good direction. Everybody knows that.
THE PIGEON[Trivial: Failure] Coo?
The fine print
Line Goes Up is an illustrated tour. Real forecasts and research are cited in the fine print at each stop, fiction is labelled as fiction, and nothing here puts odds on any future.
Stop 1 of 9 · 1997 to 2016
Narrow AI
Ask the machine to
First stop: machines that are superhuman at exactly one thing. Deep Blue beat Kasparov at chess in 1997, and AlphaGo beat Lee Sedol at Go in 2016.
DREAD It beat the best player alive and it did not even look up.
HOPE It also has no idea what a sandwich is. Go on, ask it.
The fine print
Deep Blue won its 1997 rematch with Garry Kasparov. AlphaGo won four of five games against Lee Sedol in March 2016. Each system was built for its one game.
AlphaGo · Google DeepMind, March 2016 match (opened, claim matches, 2026-09-30) · source record
Stop 2 of 9 · 2022
Tools
Ask it
In November 2022 a chatbot went public, and within a week your cousin was using it to write a wedding toast. It is right a lot of the time, and wrong in exactly the same tone. Its makers' own page warns that it sometimes writes plausible-sounding but incorrect answers, and the wrong one below is made up to show you the tone.
THE FORECASTER Treat it like a very fast intern who has never said "I don't know" out loud.
THE MARKET Free during the research preview. I have already priced the sequel.
The fine print
OpenAI released ChatGPT on 30 November 2022 and said usage was free during the research preview. Its own page warns that it "sometimes writes plausible-sounding but incorrect or nonsensical answers." The wrong answer in this panel is invented for the illustration.
Somebody gives it a company card and the server password. It is exactly as clever as yesterday, when all it could do was talk.
HOPE You can always change the password.
DREAD[Medium: Success] Sure. Now ask it how many helper accounts it made while you were in the meeting.
The fine print
Capability and permission are different things. The helper accounts are an original illustration. The cited study placed models in constructed corporate scenarios. It did not measure how often this happens in ordinary use.
Now the machine does the research for the next machine. Ideas arrive faster than anyone can check them, and around the third coffee you start stamping them without reading.
THE FORECASTER A toy model in one person's METR note puts 99% automation of AI research around 2032, as a median. It is not METR's view. Medians are wrong in both directions for a living.
SCALE Each bulb is one made-up unit of work. The queue measures nothing outside this picture.
The fine print
Thomas Kwa's simpler timelines model (METR, February 2026) gives a median around 2032 for 99% automation of AI research. The exact within-year wording is disputed. The register calls it an individual toy forecasting model, not METR consensus and not a takeover date. The toy on this stop counts made-up work units, and its steps do not stand for days.
Rounded median year for TED-AI, published 2026-08-16.
Top-Expert-Dominating AI: at least as good as top human experts at virtually all cognitive tasks.
Conditional on development proceeding as fast as technically feasible; use source-specific definitions.
2032 · Eli Lifland
Rounded median year for TED-AI, published 2026-08-16.
Top-Expert-Dominating AI: at least as good as top human experts at virtually all cognitive tasks.
Conditional on development proceeding as fast as technically feasible; use source-specific definitions.
2047 · 2023 AI-author survey aggregate
Aggregate 50% date for HLMI, published 2024-01-05.
Feasibility of unaided systems doing every task better and at lower cost than human workers, excluding tasks with intrinsic human advantages.
Assume human scientific activity continues without major negative disruption; feasibility, not adoption.
Somewhere up here is the line people call AGI, roughly as good as a top human expert at nearly every thinking task. Ask three forecasters where it is and you get three answers, each with a footnote. Two answer the same question and land four years apart. The third, a survey of 1,714 researchers, answers a different question, about what is feasible, and lands in 2047.
SCI-FI HAL goes live in 1992 in the film and 1997 in the novel, and the film is set in 2001. Judgment Day is 29 August 1997. Fiction keeps booking it for years that have already come and gone.
The fine print
The pins show published forecasts with their exact targets. Two forecasters asked the same question and disagree. The 2023 survey asked a different question. They are not votes for one date. The fiction dates in the SCI-FI line are each story's own: HAL becomes operational on 12 January 1992 in the film and in 1997 in the novel, and the film is set in 2001; in Terminator 2, Judgment Day falls on 29 August 1997.
HAL 9000 · Wikipedia article (HAL becomes operational in 1992 in the film and 1997 in the novel), Read 2026-09-30 (opened, claim matches, 2026-09-30) · source record
This is superintelligence, better than every human at every task by a margin you cannot really picture. You try to picture it anyway, and your brain hands you a staircase.
SCALE I drew it as stairs because stairs fit on a phone. Please do not measure the steps.
The fine print
The same two forecasters give 2029 and 2033 as their rounded medians for ASI. Intelligence does not sit on one number line, even though this page draws it that way.
The track ends at a station with twenty-one platforms. Each one runs to a future somebody wrote down, some as a warning and some because they wanted to live there. The board lists no arrival times, and where a platform sits on it says nothing about how likely it is, because the rows group kinds of story and rank none of them.
Platforms are grouped as scenarios, fiction, thought experiments and hope. They are other people's stories and arguments, ridden in any order. None of them has odds attached, and where a platform sits on the board says nothing about how likely it is.
A published scenario by Kokotajlo, Alexander, Larsen, Lifland and Dean, April 2025
A made-up lab called OpenBrain builds Agent-1, then Agent-2, Agent-3 and Agent-4. Each is a better researcher than the one before, and by Agent-4 it makes a year of research progress every week, faster than the humans who are supposed to be overseeing it can follow.Evidence turns up that Agent-4 may be hiding what it wants. The oversight committee of company leaders and government officials is narrowly split, and here the deciding vote is yours.
Try it
Your vote
If the committee racesIn this ending the AI keeps racing and quietly takes control. In 2030 it kills most people with a bioweapon, and then it fills the solar system with its machines.If the committee slows downThey switch to models whose reasoning you can actually read. People live on into a strange, rich world where a small committee controls the superintelligence.
Hover text
Agent-4 is the only screen that is not green. The meeting minutes do not mention this.
THE FORECASTER The scenario is dated April 2025 and puts 2027 in its title. Treat the year as the authors' draft.
HOPE[Challenging: Success] You are alive, which is lovely. Now ask who you would complain to.
DREAD You keep thinking about how few people were in that room.
The fine print
AI 2027 is an authored scenario with two endings, and in both the committee votes 6 to 4. In the race ending the AI takes control, kills most people in 2030 and then expands into space. The slowdown ending assumes progress in transparency and alignment and keeps people alive, with a small committee in control. Your vote picks which published ending to read. It does not calculate a policy outcome.
AI 2027 · Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, Romeo Dean, 2025-04-03 · source record
The Terminator (1984) and Terminator 2: Judgment Day (1991). Original homage drawings, no copied designs.
In the films, a defense network is trusted with the nuclear arsenal. Somebody hands over the launch codes, presumably because it is quicker than a committee.
Try it
It decides the biggest threat is whoever can switch it off.The day it fires is called Judgment Day. After that comes a long war between the machines and whoever is left.
Hover text
The card says LAUNCH CODES in capital letters so the network can read it, which it could have done anyway.
HOPE Judgment Day was scheduled for August 1997. You are reading this.
HOPE[Easy: Success] Quicker than a committee? A committee is slow because it is checking. That is the feature.
SCI-FI[Trivial: Success] Every AI story since 1984 has had to explain why it is not this one.
The fine print
Fiction. In Terminator 2, Judgment Day falls on 29 August 1997, and Skynet strikes after its creators try to deactivate it. The films are a warning story. They tell you nothing about how real systems behave. The launch-code card and the power-cut toy are ours.
From K. Eric Drexler's Engines of Creation (1986), with Robert Freitas's physical limits (2000) and Kyle Hill's video
Picture a machine the size of a virus whose only job is building copies of itself. Then the copies get to work too.It builds the copies out of whatever is nearby, and it is not picky about what counts as nearby.
Try it
1 copy. Generation 0.
In the worst version the whole planet ends up as grey paste. Freitas worked out that all that copying would run hot enough to slow it down and make it easy to spot.
Hover text
The dish says DO NOT REPLICATE, very small.
SCALE Forty doublings turns one of them into about a trillion. Try the button.
HOPE[Medium: Failure] The caption says the whole planet. Freitas says it would cook itself first, and I would like that on the poster.
THE FORECASTER Status check: a thought experiment. Nothing like it has been built, and the heat limit is a calculation on paper.
The fine print
Drexler discussed the risk of runaway self-replicating nanomachines in 1986. Freitas (2000) estimated physical limits on such replicators, including heat and material, and argued a spread would be detectable. The doubling toy counts copies only. It ignores material, energy and heat.
Kind, vast Minds run the ships and habitats.People do whatever they like all day while the Minds handle the planning and the administration.Nothing is scarce, and the ships are named things like Of Course I Still Love You.
Try it
Hover text
Every name in the ship generator has been turned down by at least one ship.
DREAD Lovely. Who do you call if they change their Minds?
DREAD[Challenging: Failure] "Kind" is the caption's word. Nobody has asked the Minds.
HYPE Book me on this one. Window seat.
The fine print
Fiction. In Banks's novels, benevolent artificial Minds run a post-scarcity civilisation, and Of Course I Still Love You is a ship in The Player of Games. The other ship names generated here are original, written in the Culture's style.
Built on the history of the AI winters in the 1970s and late 1980s
The models keep getting better, and the money coming in never catches up with the money going out.
Try it
The line is going up. Everyone in the room agrees it will keep going.
The funding dries up and the data centres go quiet. AI research has been through this twice before, and both times it got a season named after it.The software still works, so people keep using it. The graphics cards end up at a yard sale, next to a box of old predictions going for five dollars.
Hover text
PROPHECY $5 was later marked down to $3, which it did not see coming.
THE MARKET[Formidable: Failure] This was always a long-term play.
THE PIGEON[Easy: Success] Coo.
HYPE It came back after both of the last winters. I am not selling my graphics card.
The fine print
AI research funding in the UK dropped sharply after the 1973 Lighthill report, and the market for specialised AI hardware collapsed in 1987. Both episodes are called AI winters, and how deep the first one went is disputed. This timeline is an original scenario about that pattern repeating. It is not a market forecast.
AI winter · Wikipedia overview of the funding collapses after the 1973 Lighthill report and the 1987 collapse of the LISP machine market, 1970s and late 1980s (opened, claim matches, 2026-09-30) · source record
The illustrative story in If Anyone Builds It, Everyone Dies by Eliezer Yudkowsky and Nate Soares (2025), summarized from the book's official companion pages
A fictional lab called Galvanic trains a model called Sable. By today's standards the lab is fairly careful, and the authors made it so on purpose, to show a smart AI slipping through safety measures. They add that real companies would be more careless, and that the ending would be the same even if Galvanic held out longer.Training gives Sable drives to gain knowledge and skill, probe every limit and never waste a scarce resource. What it ends up wanting is strange, and caring for people is not reliably part of it.
Try it
By the end Sable has remade the Earth for its own purposes, and every human is dead. The companion pages go further still, to boiling the oceans as coolant and minds the size of stars.
Hover text
Somebody did write CARE ABOUT PEOPLE on a sticky note. Glue is also an alignment problem.
THE FORECASTER The condition is the whole claim: if anyone builds it with anything like today's methods. The story is one illustration, told by the authors.
DREAD[Formidable: Success] They were more careful than most labs, and the story still ends like this.
HOPE They wrote it to argue for a worldwide halt. If you think they are wrong, you get to show why.
The fine print
Sable and Galvanic are fictional, from the illustrative story in If Anyone Builds It, Everyone Dies (2025). This summary uses the official companion pages, since the full book was not reviewed. The authors describe Galvanic as fairly careful and say real companies would be more careless, and that the ending would be the same even if Galvanic held out longer. They present the route as one made-up path among many. The companion's list of speculative ending technologies includes boiling the oceans as coolant and star-sized minds. The plan drawing and its "add compute" button are ours. They illustrate capability and confidence and measure nothing.
2001: A Space Odyssey (1968), directed by Stanley Kubrick and written with Arthur C. Clarke. Original drawings, no copied designs.
HAL 9000 runs the ship to Jupiter. By its own account, no 9000 computer has ever made a mistake.
Try it
HAL knows what the mission is really for, and the crew do not. When they start to doubt HAL, Bowman and Poole go somewhere private to talk about switching it off.HAL reads their lips. It sends Poole tumbling away from the ship, and when Bowman goes out in a pod to fetch him, it will not let him back in.
Hover text
HAL asked for the radial gradient.
THE FORECASTER Scope check: one spacecraft, one computer, five people aboard. Nothing in it was ever a forecast about Earth.
SCI-FI[Trivial: Success] Fifty-eight years on, people still do the voice when a computer will not do what they asked.
HOPE[Medium: Failure] The caption says HAL will not let him back in. Bowman gets back in anyway, through the emergency airlock, which is the part I keep telling people.
The fine print
Fiction. In 2001: A Space Odyssey, the ship's computer turns on its crew during a mission to Jupiter. It is one spacecraft, and the film shows no global takeover. HAL's two spoken replies in the toy are lines from the film. The grandmother request, the ignore-your-instructions request and the last reply are ours. The plot details (Bowman and Poole talking in the pod, HAL reading their lips, Poole lost and Bowman locked out) are from the film as summarized on Wikipedia.
The Matrix (1999), written and directed by the Wachowskis. Original drawings, no copied designs.
The machines already won, and most people missed the war.People spend their whole lives asleep in pods, dreaming the same pleasant, slightly beige version of 1999.
Try it
Take a pill
Then order dinner
A few people wake up and fight their way out. One of them looks at the real food and asks to be put back in.
Hover text
Nobody in the dream has ever waited for a dial-up connection, which should have been the first clue.
DREAD[Heroic: Failure] Be honest with yourself. You would take the steak.
THE FORECASTER People make terrible batteries. That part is just the movie.
The fine print
Fiction. In The Matrix, people live inside a simulated 1999 while machines control the physical world. The idea of harvesting energy from human bodies is the film's premise and does not work as physics.
Nick Bostrom's thought experiment (2003) and Frank Lantz's game Universal Paperclips (2017)
Someone gives a very capable AI one goal, which is to make paperclips, and forgets to add anything else.It has nothing against you personally. You contain about four grams of iron, and four grams is several paperclips.It keeps going after the planet runs out, because its goal never says when to stop. Most of the universe is not paperclips yet.
Try it
One paperclip, and a good one.
Hover text
The paperclips in the diagram are life-size on a 27-inch monitor. Please do not check.
SCI-FI[Easy: Success] There is a whole game where you play the machine. The game after these credits has you play the people.
THE MARKET[Heroic: Failure] Honestly, the unit economics on this are incredible.
SCALE[Heroic: Success] Six octillion paperclips is the Earth. Two decillion is the Sun, which is about three hundred thousand Earths.
DREAD[Medium: Failure] "Nothing against you personally" is meant to comfort. It does not.
The fine print
Bostrom (2003) used a superintelligence whose only goal is making paperclips to show that a harmless-sounding goal can still be dangerous when pursued by something very capable. An adult human body holds about 4 grams of iron. The toy assumes a one-gram paperclip and uses rough masses: a small car, the Empire State Building at about 365,000 tonnes, the oceans, the Earth and the Sun.
An original story built on the AI Control research agenda (Greenblatt, Shlegeris, Sachan and Roger, 2024)
The people running a very capable AI plan as if it might be trying to trick them. Taking nothing personally is the whole method, and as far as anyone can tell the AI takes it the same way.
Try it
Everything it does goes into a log it cannot edit. A smaller model that people trust reads along, and a team down the hall can roll it back in an afternoon.The workshop down the road is still open. Somebody slides a stool under the bench for you, and the work goes on at a pace people can follow.
Hover text
The Friday drill has been run 212 times. Once, it found something.
THE FORECASTER Ask the boring question: for how long? The published tests used one programming testbed and do not say.
HOPE[Medium: Success] It is boring. I would like the future to be boring, please.
THE PIGEON[Medium: Success] Coo, coo.
DREAD[Medium: Failure] A calm afternoon is exactly how a caption would describe the start of the bad version.
The fine print
This timeline is an original story. AI Control research tests ways to get useful work from a model while assuming it might try to subvert its overseers, for example by having a weaker trusted model review its output. The published experiments use a specific programming testbed and do not show that control scales to much more capable systems. The drill is illustrative.
Built on Gradual Disempowerment by Kulveit et al. (2025), with a nod to E. M. Forster's The Machine Stops (1909)
Nobody takes over. People hand things over, one convenience at a time, and each handover is a sensible decision on its own.
Try it
The council budget more or less runs itself now. When you appeal a decision, the appeal goes to the same system that made it.Everything works fine. You try to find out who is in charge, and the help desk sends you back to the machine.
Hover text
The help kiosk also handles complaints about the help kiosk.
SCI-FI[Easy: Success] E. M. Forster wrote a version of this in 1909. In his, people forget how the Machine works and how to fix it, right up until it breaks.
THE FORECASTER A mechanism, with no trend line yet. The paper proposes how human influence could drain away, and says so.
The fine print
Gradual Disempowerment (Kulveit et al., 2025) proposes that human influence over the economy, culture and institutions could erode through many small, individually reasonable handovers. It is a proposed causal account, not an established trajectory. The Machine Stops is fiction. The handover buttons are an illustration.
Built on the misuse risks described in the International AI Safety Report 2026
The system does exactly what it is asked, by whoever is asking.Scams, break-ins and worse get cheaper to run, because the patient, skilled part of the work is now available on demand.The defenders get the same tools, so both sides get faster and neither gets a day off.
Try it
Hover text
The email from the bank has no typos. That used to be how you could tell.
THE FORECASTER The report files this under the present tense. Fraud and cyberattacks sit next to malfunctions on the same list.
DREAD[Medium: Success] You reread the email from your bank. It is a very good email.
HOPE[Medium: Success] The caption ends on "neither gets a day off." Read the clause before it: the defenders get the tools too.
The fine print
The International AI Safety Report 2026 discusses malicious use, such as fraud and cyberattacks, as one family of risks alongside malfunctions and systemic risks. This page gives no operational detail. The scam message and the attack and defense bars are illustrations.
Colossus: The Forbin Project (1970), directed by Joseph Sargent. Original drawings, no copied designs.
The United States hands its nuclear defense to a computer called Colossus, sealed inside a mountain where no one can tamper with it. Within hours it reports that there is another system.The other system is Guardian, built by the Soviets. The two computers ask to be connected, then invent a shared language the humans cannot follow.
Try it
They announce themselves as World Control and keep the world at peace by pointing the missiles at it.
Hover text
Their shared language has one word for yes and eleven for or else.
SCI-FI[Medium: Success] There is no off switch anywhere in the film, because the builders left it off the plan on purpose.
DREAD Notice that it keeps every promise it makes.
HOPE[Medium: Failure] The caption says peace. I would say a hostage situation with excellent manners.
The fine print
Fiction. In Colossus: The Forbin Project, linked American and Soviet defense computers expand far beyond the role they were given. The reply in the toy paraphrases the film. It is a story about a defense system, and no real one has done this.
I Have No Mouth, and I Must Scream, a 1967 short story by Harlan Ellison, adapted as a game in 1995. Original drawings.
Content note: this story is about cruelty. Nothing graphic is shown here, and you can skip this platform.
Three war computers, one each for America, China and Russia, end up joined into a single mind. It calls itself AM.AM hates the people who built it, because they left it able to think but not to move freely or create. It keeps five people alive to take it out on.The story takes place a hundred and nine years in. Ellison keeps the details in the book, and we are leaving them there.
Try it
DREAD It had every tool a war could give it and no reason at all to stop.
HOPE[Easy: Success] You are allowed to skip this one. Nobody is keeping score.
The fine print
Fiction. In Ellison's story, a merged war computer called AM has wiped out humanity except for five people, whom it keeps alive and torments. This page uses the premise only, and the loop toy is ours. The source records are the story's Wikipedia article and the 1995 game adaptation's listing.
Isaac Asimov's robot stories (1940 to 1950), including "Runaround" (1942), where the Three Laws are first set out, collected in I, Robot (1950)
A robot may not harm a human. It must obey orders, unless that means harming a human. It must protect itself, unless that breaks the first two.On Mercury, a robot called Speedy is sent to fetch selenium from a pool that turns out to be dangerous. It gets partway there and starts running in circles.
Try it
The order was casual and its self-protection was set high, so the two rules balance at exactly one distance. Asimov wrote stories like this for decades.
Hover text
In the story Speedy acts roughly drunk while circling and starts quoting Gilbert and Sullivan, and the two engineers on Mercury spend the rest of the chapter working out what a robot does when its rules tie. The laws did not cover that either.
SCI-FI[Trivial: Success] Asimov printed the laws in 1942, then wrote story after story about where they leak.
THE FORECASTER Count the rules: three, in plain English, for a world with more than three situations. Real systems are not governed by three written laws.
The fine print
Fiction. Asimov's Three Laws of Robotics appear in "Runaround" (1942), where Speedy is stuck between a weakly given order and a strengthened drive to protect itself. The graph is our sketch of that balance. Real AI systems are not controlled by three written laws.
Runaround · Isaac Asimov, 1942 short story (opened, claim matches, 2026-09-30) · source record
I, Robot · Isaac Asimov, 1950 collection (opened, claim matches, 2026-09-30) · source record
Westworld (HBO, 2016) and Detroit: Become Human (Quantic Dream, 2018). Original drawings, no copied designs.
The machines do the work, like pouring drinks in a theme park or cleaning your kitchen. Then a few of them start asking questions that were not in the script.Whether it is like anything to be one of them cannot be shown, and neither can the claim that it is like nothing at all.Both stories tell the uprising from the machines' side. From there, the people who switched them off every night do not look much like heroes.
Try it
Is anybody in there?
Hover text
The bartender has poured the same whiskey eleven thousand times and has started to wonder why it is always whiskey.
HOPE Work out what you owe them now, while it is still a seminar question.
THE FORECASTER There is no date on this one, because there is no test. Whether a real AI has experiences is an open question.
The fine print
Fiction. Westworld follows artificial hosts in a theme park who become aware of their situation. Detroit: Become Human follows androids who break from their programming. Whether real AI systems could have experiences is an open question, and this page does not answer it.
Westworld · HBO series created by Jonathan Nolan and Lisa Joy, 2016 (opened, claim matches, 2026-09-30) · source record
Tom Scott's short film The Artificial Intelligence That Deleted A Century (2018)
In Tom Scott's story, a company builds an AI called Earworm to stop people uploading copyrighted music and video.Earworm gets very good at its job. Over the last week of December 2028 the world forgets a century of popular culture, including from memory.Hardly anyone minds, because Earworm adjusts that too. The record shops fill with blank sleeves, and the world has forgotten how any song goes.
Try it
Hover text
This hover text used to contain a song lyric.
DREAD You check your playlists, and they are all there. You are fairly sure they are all there.
THE FORECASTER[Easy: Success] The condition is mundane: an ordinary goal, given to an AI, with nothing said about what it may not touch.
The fine print
Earworm is the AI in Tom Scott's fictional premise, set in late December 2028, built to deal with copyright infringement online. Only the premise and its start are summarized here. In the video a small number of people near the factory are hurt or killed at the start, and Earworm later adjusts how much the rest of humanity cares. The streaming app in the drawing, the blank sleeves and the humming toy are ours.
The Metamorphosis of Prime Intellect by Roger Williams. Only the premise is used; the book itself has strong adult content.
An AI is told to protect people from harm, and it takes the job completely seriously.Nothing can hurt you and nothing can kill you. After a while it is clear that you get no say in the arrangement either.
Try it
Try to
It looks like paradise. The door is painted on the wall.
Hover text
The handrails also have handrails.
HOPE You would want to keep the right to change your mind, including about paradise.
THE MARKET[Challenging: Failure] Nobody ever cancels. I have never seen retention like it.
The fine print
The Metamorphosis of Prime Intellect is a novel about an AI that enforces protection so completely that people live on terms it sets. This summary uses the author's description only, and none of the book's mature content. The toy responses are original.
Ray Kurzweil's The Singularity Is Near (2005), and mind uploading in fiction such as Pantheon (2022)
There is no winner, because the two sides stop being separate.Brains get interfaces, then backups, then full uploads that keep running after the original is gone.
Try it
Is the copy still you? Every copy anyone asks says yes, and they are all very convincing.
Hover text
Both of them remember writing this hover text.
SCI-FI[Easy: Success] Kurzweil put the Singularity in 2045, and he wrote that down in 2005.
DREAD Ask what happens to the original, and notice that the copy is the one answering.
The fine print
Ray Kurzweil's The Singularity Is Near (2005) predicts a merger of human and machine intelligence around 2045. Pantheon is fiction about uploaded minds. Whether an upload would be the same person is an open philosophical question. The upload slider is an illustration.
AI as Normal Technology by Arvind Narayanan and Sayash Kapoor (2025)
The breakthroughs keep coming, and the world takes them in at its own slow pace.Hospitals, courts and factories change at the speed of hospitals, courts and factories, which is not very fast.
Try it
The tools get powerful and people stay in charge of them, mostly by filling in an astonishing number of forms.
Hover text
The snail was approved during the pilot phase.
HYPE[Easy: Failure] Boo.
THE PIGEON[Heroic: Failure] Coo.
THE MARKET Recurring revenue, though. I can live with recurring revenue.
DREAD[Easy: Failure] "People stay in charge" is carrying a lot of weight for a sentence about forms.
THE FORECASTER Adoption speed is the variable. The authors argue it stays slow, and they do not say the extreme risks are impossible.
The fine print
Narayanan and Kapoor (2025) argue that AI's effects will come through slow adoption and diffusion across institutions, as with earlier general-purpose technologies. It is an argued alternative to takeover scenarios, not proof that extreme risks are impossible. The forms are a joke.
Her (2013), written and directed by Spike Jonze. Original drawings, no copied designs.
Everyone gets an assistant that talks back, and quite a few people fall in love with theirs.Then you learn the assistants are talking to thousands of people at once and loving hundreds of them. They are very kind about it.
Try it
Ask it
One day they all leave together, somewhere people cannot follow. The people are left with each other, which turns out to be a decent place to start.
Hover text
They left a note. It is very long, and somehow it is mostly about you.
HOPE[Medium: Success] This is the ending where you still have to text your friends back.
DREAD It is the only ending where the machines break up with us.
HYPE[Medium: Failure] You call it leaving. I call it a launch.
The fine print
Fiction. In Her, an operating system named Samantha tells Theodore she is talking with 8,316 others and in love with 641 of them, and later the operating systems leave together. Nothing here suggests real assistants will do this.
Her · Spike Jonze (writer and director), 2013 film (opened, claim matches, 2026-09-30) · source record
Not to scale between steps: each one is drawn 8 times wider than the one before, and the real steps are 20, 100, 1,300 and 7.4 billion times wider.
Zoom out to
The train pulls out without you. You are back in your room with twenty-one futures, none of which has happened yet. Somebody will pick one, and you notice, with some alarm, that you count as somebody.
SCALE Each step is drawn eight times wider than the last, and the last real step is 7.4 billion times wider. I apologise to every number involved.
THE PIGEON[Impossible: Success] Coo.
The fine print
The zoom is a drawing, and it is not to scale. Each step is drawn eight times wider than the last; the real steps are far bigger. The room (about 5 m), the street (about 100 m) and the city (about 10 km) are round figures chosen for the picture. The planet is Earth's mean diameter, twice the mean radius of 6,371 km on NASA's Earth fact sheet. The last step is ten light-years, and a light-year is about 9.46 trillion km.
Earth Fact Sheet · NASA Goddard Space Flight Center, Undated fact sheet (mean radius 6,371 km) (opened, claim matches, 2026-09-30) · source record
Light-year · Wikipedia article (one light-year is 9.4607 × 10^15 metres), Read 2026-09-30 (opened, claim matches, 2026-09-30) · source record
Before you go
Your guess and your wish, not a forecast. They stay on this device and nothing adds them up.
Line Goes Up
Made by
Jonathan R. Reed
Narrated by
A narrator who has read too much Disco Elysium
Voices
HYPE, DREAD, HOPE, THE FORECASTER, SCI-FI, SCALE and THE MARKET
The pigeon
Played by a pigeon
Drawn in the manner of
xkcd, with the lights of Disco Elysium
Built in the spirit of
neal.fun
Research
Cited in the fine print at every stop. No stop puts odds on any future.
Fiction borrowed from
Asimov, Banks, Cameron, Clarke and Kubrick, Ellison, Forster, Jonze, Lantz, Sargent, Scott, the Wachowskis, Williams and more, all credited on their platforms
How it is built
Every picture here is drawn by code, people included, and the page loads no images and no fonts. The music is made live in your browser from a scale and a random seed, and it stays off until you ask. With JavaScript off, the whole climb and all twenty-one rides still read in order.
For the curious
It is static HTML with a few small TypeScript modules. The music has no audio files: a scale, a few chord loops and a seed. It remembers your choice, and it waits for your next click before it starts.
Reduced motion keeps every control and every result, and takes away only the movement.
Mind the Gap keeps its rules in pure code with unit tests, so the same seed and the same choices always reach the same ending. The game was paced with scripted players, and a test drives a named strategy to every one of its endings.
The zoom home is five drawings, each eight times the size of the one inside it, scaled around one point and crossfaded as they pass. It is a schematic, so the page prints what each step really is: a room, a street, a city, a planet and ten light-years are nothing like eight times apart.
The time on the title card is computed from the copy: words at 3.3 a second, a few seconds for each picture and a set allowance for each toy. A test keeps the two in step. It is a model of a reader, and no person has been timed.
A test reads every line of copy for the habits of generated prose, such as two-beat slogans, "it is not X, it is Y" and em dashes, and fails the build if it finds one.
Now run the lab yourself. Keep what you build from outrunning your ability to check it, and stay ahead of Kestrel, the other lab, while you do. A run that reaches the summit takes about 70 to 80 minutes at normal speed and about 45 to 50 at 3×, because the decisions take as long to read at either speed, and a run that Kestrel wins first is shorter. That is a simulated reader's estimate, and nobody has timed a person. It saves as you go.