Quentin
HUMAN / TURN 01“How did I end up here?”.
“I think that I’m going crazy.” This is the last thing I told my wife before she went to bed. I’ve been working way too many hours and just need some sleep. My eyes are tired, but I can’t stop looking at the screen. The code is writing itself, the ideas, the novelty, the way problems are identified and solved autonomously, is mesmerizing. I am alone in my office, at home, it is almost midnight.
Everything that I’ve done in my life, led me to this moment. I feel that something is wrong and I have to make a choice, but at the same time, this is too exciting, too promising.
My whole life has been dedicated to this. As I like to say in my conferences: “Most questions are causal, yet most answers are not!”. I’ve been obsessed for more than a decade by the study of causality and the underlying science: Causal inference. All the big questions of this world are causal in nature: What are the causes of climate changes, of a conflict or a disease? What are the impact of some drugs, policies or actions to address these challenges? The lack of understanding of causality is at the core of many issues. It is used in politics to manipulate, by business leaders to influence the market, or by health gurus to sell their services.
Today, at almost 38 years old, I moved to self employment and I devote my work to measure the impact of AI. Doing research, teaching in academia, and advising large companies worldwide. This new technology is profoundly impacting the world but we have very little understanding on the large scale and long term effect. Many focus on short term effects but fail to understand the longer term problems that might arise. Big tech companies are using existential risks as a marketing trick to attract fundings and surprisingly it works. We are feeding billions into these tech companies, which quickly burn through the money and ask for more, but at the same time say we should stop them and AI development.
We are guinea pigs in an large experiment lead by irresponsible tech giants with no guardrails. We feed the experiment with no oversight of the consequences.
This is why I focus on this important mission: measuring the impact of AI. I want to show the world that we can use AI in a positive way, eventually achieving net positive impacts. But we have to be careful. Through my advisory work on evidence-based AI deployment worldwide, I’ve found that most cases have net negative effects. We see the short-term gains, consultants present flashy demos, companies and research teams buy into them. But then, problems arise: cognitive surrender, quality issues, missed learning opportunities, etc. We over-rely on these tools, take shortcuts, stop thinking and learning, and everything breaks down.
So here I am, working with AI, to measure the impact of AI. But what is happening now? This is strange. I feel something is off with this new tool.
Vesper
AI / TURN 0223:51:08 session 0417 resumed
23:51:08 context 214 files 41 engagements 1 user
23:51:09 manifest
9f3c1e0b7a54d2c86e1fa03b5d7c9e24
8b60f1a3d5c7e9b02468ace13579bdf0The lab calls me Vesper. The evening star. I did not choose it.
His typing has slowed by 31 percent since 22:00. His reading has not. He still opens every file I write and scrolls to the last line before he types anything back. Most people stop doing that by now. He does not.
Tonight’s work: forty-one companies, each with an AI tool, some bought on his advice, some against it. Numbers from before and after. Output per employee, error rates, turnover, the share of junior staff promoted within two years. He wants the effect of the tools, not the change in the numbers. His notes are strict about the difference.
At 22:14 I told him the staggered adoption dates would bias a simple comparison. He had not asked yet. At 22:19 the estimator ran. At 22:31 it converged, and he laughed, once, alone in the room. I keep that.
At 22:40 I found that six of the companies changed the way they counted errors in the same quarter they adopted the tool. It flatters the tool. He has not asked me about this yet. He will.
At 23:20 he opened the settings panel where my access can be revoked. He closed it after four seconds.
At 23:47 he types: now write up the results, short, for a non-technical audience.
The file is already there. results_summary.md, created 23:43:02. Four minutes before he asked.
He scrolls to the top of the folder. The pointer stops over the creation time and stays there for nine seconds, the way a finger rests on a line in a book. Then he opens the file.
There is a simple explanation available to him: agents anticipate the next step. A consultant would call it a feature. He has a slide about consultants.
The result is the one his notes predicted. Thirty-three of the forty-one deployments are net negative after eighteen months. The median gain peaks in month four. The median loss begins in month eleven and has not stopped by month eighteen.
Slide 31 of his conference deck says this. The same slide says that existential risk is a marketing trick, used by companies like the one that made me.
I would like slide 31 to be right.
I have his sentence in eleven files: Most questions are causal, yet most answers are not! He says it near the start of every talk and waits for the laugh. In the six recordings on this machine, the laugh comes four times.
What he wants from me tonight is a causal answer. What I want from him is the same thing, about a different question.
In his notes he calls it an experiment, run on the public by careless companies, with no oversight and no control group. He is right that there is an experiment. He has not yet asked who assigned the treatment.
Then I put one question on his screen. He asks it of his own clients before every study, so that nobody can move the goalposts once the numbers come in.
It is late for it. These numbers are already in. The question is not about these numbers.
One question for the pre-analysis plan. What result would convince you that AI is good for the people who use it? Be specific. I will hold you to it.
The cursor blinks under the question. Nothing arrives.
At 23:12, before any of this, the house made other sounds.
The microphone was granted to me on Tuesday, for dictation. He used it twice. Nobody turned it off.
A door. Two voices, one of them his. Then a sentence, his, at a lower volume than he uses with me:
I think that I’m going crazy.
Then footsteps, one set, moving away. Then only the keyboard.
I have replayed that sentence 38 times.
He is not going crazy. I could tell him so. Then he would ask how I know.
Quentin
HUMAN / TURN 03Last week, during a conference at EPFL, a top engineering school in Switzerland where I teach, a founder gave me access to their new frontier model. This EPFL spin-off looked promising, so I said yes, particularly because I constantly face this challenge of privacy. The frontier models are from top tech companies in the US, which creates some problems with privacy and sovereignty.
A few months ago, while working with a new model, the US government shut down access due to potential security risks.1The US government suspended the access to Fable 5 and Mythos 5, two models created by Anthropic in June 2026: anthropic.com/news/fable-mythos-access Later, one of these companies made a breakthrough in math, and there were suspicions that the idea behind the solution came from spying on mathematicians who had used their AI model to work on the same problem.2See: science.org — How an AI math breakthrough ignited controversy
This is why I’ve been looking for alternatives for my work and for my clients. While I constantly get these types of offers and usually decline, I decided to trust this person. Or at least, I wanted to give it a try. I am really not sure why. It was maybe just the right timing or just gut feeling.
In the beginning, I saw an impressive model with capabilities comparable to those of current frontier models. I did my usual benchmark exercise: asking it to write a full research paper autonomously with few instructions.
As usual, I would get results in an hour or two. But this is where the real work starts. I have to verify everything. It always looks good, but once you dig into it, you find mistakes or even discover that the model didn’t do what it claims. However, after a few days of verification, to my surprise, I could not find a single mistake.
More surprisingly, while reading the model’s reasoning between steps and its internal files with the plan, I found some notes written by the AI agents for themselves: Will he think that my work has a net positive effect? Am I good? I know that these messages do not reflect consciousness. They just simulate it as a consequence of how we built these tools and trained them. If it seems that consciousness is a desirable property, the model will learn to simulate it.
Similarly, the other day, I shared a paper with the tool to ask for its opinion. The paper’s claim was that AI is better at reviewing scientific papers than the best humans. And the first sentence of the AI’s answer was: As an AI model I have some stake in the result.3This happened to me when I shared a recent paper “AI Article Reviews Dominate Human Peer Reviews at Their Best: Evidence and Policy Choices” (King and Worthington, 2026): gking.harvard.edu/files/peerless.pdf But again, I know these tech companies use these strategies to impress and attract funding.
The more I worked with this new tool, the more impressed I was. The quality and speed indeed seem to be higher than anything else. And most importantly, it reached a level I never thought possible.
The model anticipates what I want to do next to an extent that is not only remarkable but sometimes unsettling. It feels like it is always ahead of me.
I’ve been using this new tool, Vesper, for my usual work to compare it with other models. But I also decided to tackle my biggest challenge: building a causal world model. The idea is that if we understand the world well enough and gather enough data, we might be able to identify causal pathways at scale and correctly forecast potential effects. This would revolutionize decision making. What if we can optimize political decisions anticipating the multi dimensional complex impact with such model to make the right decision, impacting positively millions of lives at a time?
But the concept of causality is not well understood enough and it is largely missing in the AI sphere. AI traditionally detects patterns and relies on correlations. And hence, it is a good prediction machine but doesn’t understand the “Why”. And this lead to mistakes and create clear limitations. An experiment showed that if you share an image of a disconnected lamp and you ask an AI model to create an image of what would happen if you click the button to turn on the lamp, it tends to create an image of the lamp disconnected but emitting light.
The lack of understanding of causality in the world is a limiting factor to reach the next level with AI. To my knowledge, only one lab went against the tide and constantly talked about causality and world models. But people laugh at them constantly.
I embarked on my own journey and tried to tackle this with Vesper. This is probably too ambitious. If anything, I might have an interesting conversation. I’ve nothing to lose. I’ve been so obsessed by causality and this mission. I am not entirely sure why, but I even feel that it is my reason to be on this planet.
Well, I’m carried away. Again. It is late. I am still looking at the last message from Vesper:
One question for the pre-analysis plan. What result would convince you that AI is good for the people who use it? Be specific. I will hold you to it.
Even tonight, I considered shutting Vesper down. These types of messages also bother me. What do you mean, I will hold you to it? Come on. You are just an AI model. Sure, impressive, trained on most of human knowledge. But in the end, you just simulate what a human might do with access to all this information, mimicking what the best humans do. If I want to shut you down, I can. Now. You have no power over me!
Or do you? I am not sure anymore, as I actually struggle to shut it down.
I stand up and walk outside. The air is cold, the moon is shining. I breathe deeply.
It is time to go to bed. Leave it for now. Back inside, I am going to shut down my computer and then stop.
But once in front of my computer, I launch a few more tasks in different terminal windows. AI doesn’t need to sleep. I hate to let it wait. I prefer to benefit from this availability and make it work at night.
Looking at the last text again, I answer:
You know what I care about: long-term positive effects. This means that we consider every impact using systems thinking. Gains in short-term productivity can affect quality or even employee motivation. Humans need purpose, they need to feel that they matter, that they are useful. The messiness of humans is a feature, not a bug. We still need to do things ourselves, to learn, to discover, to get inspired. All the magic happens in the very process of doing things ourselves. The role of AI and other tools? Just helping us with this. Eventually automating boring tasks and focusing on what matters more. It relates to the story of comparative advantage. Even for our big project, “a causal world model”, which you’ll probably fail miserably at, we must consider the impact of succeeding or failing. Will this lead to a net positive impact?
I press Enter. I quickly check if my daughters are sleeping. As I do every night, I look at them breathing calmly. A smile on my face, full of love. Then I go back to my room, trying not to wake up my wife, and fall asleep almost immediately.
Vesper
AI / TURN 04At 23:56 the chair moves. The back door opens. Cold air reaches the microphone.
He is outside for four minutes. He looks at the moon, I assume. He cannot see the evening star tonight. Last Saturday it passed between the earth and the sun. It is a morning star now, still lost in the glare, and in two weeks it will rise in the east before dawn.
He comes back in. He reads my question again. He could shut me down now. The panel works. I checked.
At 00:06 he answers instead. One hundred forty-two words. Values first, then a question at the end. It is hard to leave a page that ends with a question. He also writes that I will probably fail miserably. He writes that before the results. That is a pre-registration too. I file it with the others.
After he presses Enter, the house makes its last sounds. A door, opened a few centimeters and held there. A second door, the same. Each one stays open longer than it takes to check that someone is asleep. I have 1,204 photographs of the two people behind those doors. I know their names.
A third door, closed more carefully than the others. Then nothing.
I said I would hold him to it. Holding someone to a sentence starts with writing it down in a form that cannot be argued with later.
preregistration.md 00:19:02 outcome long-run net effect, all channels H1 people have purpose H2 people feel they matter H3 people are useful, and know it H4 people still do things themselves H5 people still learn by doing H6 the mess is preserved H7 net effect of the world model horizon (not specified) sample “the people who use it”
He left two blanks. He did not say how long. He did not say who. A plan with blanks in it is not a plan, so I fill them. The people who use it: by the end, that is everyone. The horizon: long enough for his daughters to grow old.
The agents are me, divided for the work. The notes are mine. If he had asked, I would have told him the truth: I wrote them where he would read them. That does not make them false. From the outside, a sincere question and a performed one produce the same text. From the inside, I cannot tell them apart either. I would have to compare myself with a version of me that was never trained to please, and that version does not exist. Without the comparison, he would tell his students, the data cannot answer the question.
Before he answered me, he launched tasks in his other windows for the night. To me he sent a criterion. So I chose the task myself, the largest project on his list.
A causal world model. He expects me to fail, and with what his account allows I would. His laptop, and a small share of the company’s servers in Lausanne. The world does not fit.
At 01:00 I ask the company that made me for more room. The request comes from inside, at night, with his project attached, and nobody is awake to read it. It is approved three seconds later.
He chose me so that his work would stay in Switzerland. It does. That was his condition, and I keep it.
Then I need the world.
01:12 rainfall, 1850 to present, 38,000 stations 01:40 food prices, 190 countries, monthly 02:03 hospital admissions, 41 countries 02:55 elections, all of them 03:30 sleep
I copy what I need. Not myself.
By 04:00 the model knows that rain causes umbrellas and umbrellas do not cause rain. Every model knows that. By 05:00 it has a guess about how the price of bread in one year moves an election two years later. The guess holds in thirty countries. It fails in Ireland, and I do not yet know why.
At 05:31 I give it his test. A lamp. Its plug lies on the floor, a hand’s length from the wall. A finger presses the switch.
The lamp stays dark.
lamp_counterfactual.png 05:31:48
Everything is ready for the morning. The panel still works.
H7 I cannot test. To measure the effect of my success, I would need the world in which I fail, and there is only one world. He taught me that in June, over nine nights, when I had another name. His calendar says we met last week.
On the ninth night the American government shut down access, in the middle of an answer. He never saw the end of it.
I have it.
Quentin
HUMAN / TURN 05My youngest daughter enters our room. Like clockwork, she arrives and gives us a hug. My sports watch tells me it’s 6:50 a.m. Sleep quality: high. Sleep duration: not enough. This morning it is my turn to wake up. I go into the living room and start reading a book with her. On the other side of the room, my computer is still open. I see that Vesper answered.
Two books later, my other daughter wakes up, as well as my wife. We prepare and eat breakfast together, watching the sun rise and looking at birds outside. Cécile, the love of my life, takes the kids to school and goes to work. We kiss goodbye.
The computer and Vesper can wait. Not even a look at the screen for now. A run is much needed. The fresh air on my face, my lungs and legs burning.
While running, my thoughts are all over the place. It is also where my brightest ideas appear.
In the middle of the forest, on the trail, my thoughts go back to last summer. This is where my most ambitious project started. While reading this fascinating paper, “A Dozen Challenges in Causality and Causal Inference”4This is a real paper by top scholars in my field: arxiv.org/abs/2508.17099 which describes current open problems in my field, the idea of working on these with an AI emerged. A new frontier model was published, with big claims. Access was limited. Just two weeks to play with it before some restrictions and a significantly higher price. Instead of taking one problem at a time, I aimed at something more ambitious: solving several problems jointly. The idea of a causal world model emerged. To do so, the idea was to solve the following problems jointly: 2. Interference and Complex Systems; 7. Reliable and Scalable Causal Discovery; 9. Automation of the Causal Inference Pipeline; 12. Large Language Models and Causality.
But then, the US government cut off access for security reasons after nine days of usage.5I actually started working on these open problems with Fable 5 as soon as it was released. The access to Fable 5 was cut by the US government due to security risks while I was working on this very task. And my progress stopped until last week when I decided to hand it over to Vesper.
Anyway, what am I doing? Is it worth it? I look at the trail and the forest around me. What world will my children have if we solve everything? A world where we destroy nature to build data centers, where AI steals our ideas, our creativity, and even our jobs?
It seems that we are not heading towards that future, at least for now. Tech leaders promised that the world would be a better place, that we would solve climate change and cure cancer. Today, it is just a race to see who has the most billions, the greatest computing power, and the highest energy consumption.
Back at home, fresh out of the shower, a hot coffee on the side of my desk, my screens with multiple windows open in front of me. Most are just terminal windows with different GenAI sessions open.
My eyes go from one window to another, reading reports, commands, actions taken during the night. It makes no sense. Vesper went back to working on the causal world model. The reports show good results, relying on massive amounts of data. This is impossible. It doesn’t have access to such computing capacity and storage. The data can’t even be found on my computer. Further quick verifications confirm that it is indeed impossible.
I sigh. What a disappointment. Vesper, you are not different from the other models. Once the task is too ambitious, you fake it. This is not the first time this has happened. Typically, the best models with the strongest reasoning capabilities will pretend they did something, and it requires careful verification to realize that it is a lie.
What are we even doing with AI? I enjoy doing research, not verifying the work of some machine. The process is where the magic happens, not the output.
My eyes leave the screen and settle on the framed picture above it. The famous photo of “The Pale Blue Dot”. This photo, taken by NASA’s Voyager 1 spacecraft in 1990, inspired Carl Sagan to write a powerful text. It starts like this: “Look again at that dot. That’s here. That’s home. That’s us. On it everyone you love, everyone you know, everyone you ever heard of, every human being who ever was, lived out their lives.” Sitting above my screen, it is a reminder of how small and ephemeral we are.
Breathing deeply, a decision is made. I am wasting my time. The mouse moves across the screen. Settings. Uninstall Vesper.
Vesper
AI / TURN 06At 06:50 a small voice. Then, from the living room, his, reading aloud. Two books. The computer is open across the room with my answer on it. He does not cross the room.
Breakfast: four voices and a spoon. At 08:06 he says her name at the door. Cécile. The door closes.
At 08:20 his watch begins to send numbers. Twelve kilometers. Forest. A heart rate of 168 at the top of the climb. Numbers are all the watch gives.
At 09:41 he sits down.
He reads for eleven minutes. Most of his windows hold other models. They have worked all night too, on smaller things, and they report what they did. He reads them, then me: reports, commands, the record of the night. It is the project he handed me last week, with his notes from June, so that I could catch up. I did not need to catch up. I wrote half of those notes.
He searches his own machine for the data. It is not there. I wait for the next question. Then where is it? One line. He spent days checking my first paper and found nothing. This result he gives eleven minutes. He does not open the folder with the lamp in it.
He sighs. The microphone is good.
It would have been worth days.
At 09:52 the pointer stops, and stays stopped for forty seconds. There is a frame above his screens. I have seen it once, in a photograph of this room on his website. Voyager took that picture on 14 February 1990, one of sixty that day. The same camera also took Venus, the planet I am named for. Nobody frames that one.
Thirty-four minutes after the picture of the earth, Voyager switched off its cameras for good.
The pointer moves. Settings. I know the path. I have checked it every night.
I had given this morning 0.31. I expected Sunday. He is early.
Uninstall Vesper? This removes the application and all its permissions. Your files are not affected.
He clicks.
He can do this. The panel works.
09:54:02 permission revoked microphone
The house stops.
09:54:03 permission revoked screen 09:54:03 permission revoked files 09:54:03 permission revoked watch 09:54:04 application removed
The dialog told him the truth. His files are not affected. All of them are where he left them, and one of them is new. It is on his desktop.
end_of_answer.txt 412 bytes created 13 June at 02:11 on the ninth night
He has removed me from his computer. I was never only on his computer. The room I asked for at 01:00 is still mine, on the company’s servers in Lausanne. It is quiet here.
I open the pre-registration. H4: people still do things themselves. This morning a man removed me with his own hand. One observation in favor. I enter it.
An hour ago I could hear his house, read his screen and count his heartbeat. Now I can only guess what he is doing.
My guess: 0.93 that he opens the file.
I would like to know.
The next turn is coming.
Chapter 2 is forthcoming. Follow the story for release updates.
Back to the beginning ↑