I was reading back Borges. One of his short stories, “Pierre Menard, Author of the Quixote” reminded me current debates about LLMs and slop, so I wrote this little fiction.
***
Like many in the Valley, John Menard learned to code when most were learning to read. He then created his start-up while others were rambling for their first internship. E-fart® proposed office chairs able to detect farts. Palantir bought the company for enough money so that John would never have to sit on an office chair again.
His new free time challenged him more than any bug. He fought it by losing weight, then gaining muscle. His abs became sharper than the quarterback who used to flush his head in high school. Yet boredom still followed him around. He tried traveling. After two months in Thailand and one in Peru, he went back to San Francisco. The four-season hotels were all the same. Only the accent of the server differed, and not by much. Yes, he could have tried more perilous itineraries, but the new largeness of his budget did not allow him to.
By making him feel the absence of the singularity, his wealth made the world much more boring, he felt he had the responsibility to bring it into this world, although he knew little about what the Singularity meant, and even less about how to bring it. Until that night of insomnia.
John had reread the Diary of a Wimpy Kid. On page 43 of the second tome (John noted everything on it) that he realized the truth. AI was able to code whole applications in minutes; yet it was unable to write a novel that was not mediocre. If an AI wrote just one good story, humanity would finally get out of the darkness where it was doomed.
Fiction might seem vain compared to other benefits of AI, such as filling Excel spreadsheets or curing cancer. But wasn’t it at the time that literature started flourishing that the economy developed? The same will happen when the singularity arrives. Still, John was not sure. He asked his friends about his idea. They said it was shit. He asked ChatGPT. It said that the idea was brilliant. John was therefore convinced. He would name his baby Large Language Menard, the model that would create the greatest piece of literature word by word. John said creating rather than recreating. Too easy would be to conceive a model that remembered the novels in its database and regurgitated them. Same for a model that would recombine the influence of Cervantes to rewrite the Quixote. Much more interesting was a model having all the knowledge of the world, and who would come to the Quixote. This would prove that the machine is able to add something truly original.
Once the Quixote was created, John would delete other weights in the model, maybe those about the Diary of a Wimpy Kid. He would then repeat the experience. Again. And again. The final goal was to recreate the entirety of human knowledge through LLMenard. Once this was accomplished, no one would be able to deny the intellectual equality between AI and humans. The main challenge would be to identify and delete the appropriate weight in the neural network. The cost would come from computation. John had to buy one hundred acres of land from a Virginian farmer who thought he was Mark Zuckerberg. After two months of construction and billions of scattered words generated, a sentence from the 19th century translation of the Quixote appeared:
The simplicity of the housekeeper made the licentiate laugh, and he directed the barber to give him the books one by one to see what they were about, as there might be some to be found among them that did not deserve the penalty of fire.
“The singularity is there,” tweeted John Menard, accompanying his message with an image showing the sentence in the code and a link to the GitHub repo.
The tweet went trending. A few were impressed. Most said that this was the perfect example of AI limitations. There was no lived experience in the words. Cervantes knew the penalty of fire, the laughter running through his lungs. He probably went to the barber to shave his mustache, and maybe the blaze reminded him of the swordsman who took his arm in Italy, and when the blaze would get close to him, he would discreetly touch his stump with grief. When Cervantes mentioned books, he was talking about those bulky objects that were hard to get, whose pages get dirty and smell weird. None of this implied statistical inference. Furthermore, LLMenard produced these sentences surrounded by scribble while Cervantes wrote them in the middle of a greater architecture. It is this precise architecture that gave the words all their meaning.
John locked himself for a whole week, reading each comment until a solution finally popped into his mind: he would connect his old e-fart prototype to LLMenard. Now, if the model wrote the word “fire”, it would know the exact molecular recipe of its smell. The problem of lived experience was solved. Everybody would have proof of it through the GitHub repo that he was going to post. Obviously, making it open-source could cause him legal trouble. John asked ChatGPT about it. The chatbot responded that John was brilliant. LLMenard 2.0 was therefore published.
In a few weeks, the new LLMenard stumbled upon other sentences corresponding to those of the Quixote. John proudly published them on X. He was sure that those tweets would cause a white-collar bloodbath or something.
The new sentences were rejected even more violently. LLMenard 2.0 knew molecule rather than feeling. His usage of adjectives felt awkward, his sentences badly constructed. Some sentences were the best possible examples of AI slop. No sane human could have ever written so badly. John responded that those sentences were, word for word, the same as those of Cervantes. The critics said that this was simply a coincidence. What was wordy in LLMenard’s writing became baroque under Cervantes’s pen. One had only to see all the critics made about the Quixote. LLMenard could never, never create such a blossoming of opinions and thoughts.
John sent all the criticism to ChatGPT, trying to find an answer to his questions. ChatGPT told John that humans were simply too stupid to recognize the genius of Menard. John never knew if ChatGPT was talking about him -John Menard- or about LLMenard. It didn’t matter. The only difference between slop and genius, John understood, was that people with rounded glass spent centuries commenting every comma of Cervantes. John could do the same in a few days of vibe coding. He therefore vibe coded LLScholar, a model who would analyze every sentence generated by LLMenard and demonstrate the genius of each word. To make LLScholar sound less biased, the model would be accompanied by LLScholar 2.0, an LLM that would criticize the output of the first critics so the debate would make LLMenard sound more profound. Internet was not satisfied. So John created an agentic model that generated critics of critics commenting on each other: LLUni. Each sentence of LLMenard would therefore be analyzed by an infinite number of LLScholar generated by LLUni. The data centers in Virginia started heating up.
The announcement of LLUni got fewer than a hundred impressions on X. The world was on the verge of explosion, the third world war was knocking on America’s door, or maybe it was America knocking at its door. Anyway, the silly literary game of a multi-millionaire did not interest the public anymore. Idiots. John wondered if his mistake was not believing that the Singularity would need humans. Maybe it was an event between machines that only a few prophets would see before dawn burned their eyes. Was he one of those prophets? Police sent him a cease-and-desist letter before he got the answer.
John, while reading the letter, felt again the cold water of the high school’s toilet on his neck. The quarterback had just pissed in it and the torn pages of the Diary of a Wimpy Kid were floating around. He asked ChatGPT what to do. ChatGPT told him he should escape north. John took his Tesla, towards Canada. At the border, he found federal officers waiting for his ass. ChatGPT snitched on him.
The charges against John were simple. E-fart was now used to detect the distinct flatulence of terrorists and foreign presidents. By making the technology open source, he had just released private military information to the public. John was accused of high treason. When it was established that he had converted one hundred acres of land into a data center, for a project that seemed devoid of value to the American people, protesters surrounded the tribunal. Still, John could have got a reduced sentence, but he insisted on defending himself using ChatGPT. ChatGPT reexplained the brilliance of John during the whole trial. The jury was not convinced. John was sentenced to 180 years in Guantanamo.
John accepted the sentence without recourse. Maybe that when he gets out, he could read the complete version of the Quixote written by LLMenard. In meantime, he would use his free time to read all the volumes of the Diary of a Wimpy Kid.
LLMenard
I was reading back Borges. One of his short stories, “Pierre Menard, Author of the Quixote” reminded me current debates about LLMs and slop, so I wrote this little fiction.
***
Like many in the Valley, John Menard learned to code when most were learning to read. He then created his start-up while others were rambling for their first internship. E-fart® proposed office chairs able to detect farts. Palantir bought the company for enough money so that John would never have to sit on an office chair again.
His new free time challenged him more than any bug. He fought it by losing weight, then gaining muscle. His abs became sharper than the quarterback who used to flush his head in high school. Yet boredom still followed him around. He tried traveling. After two months in Thailand and one in Peru, he went back to San Francisco. The four-season hotels were all the same. Only the accent of the server differed, and not by much. Yes, he could have tried more perilous itineraries, but the new largeness of his budget did not allow him to.
By making him feel the absence of the singularity, his wealth made the world much more boring, he felt he had the responsibility to bring it into this world, although he knew little about what the Singularity meant, and even less about how to bring it. Until that night of insomnia.
John had reread the Diary of a Wimpy Kid. On page 43 of the second tome (John noted everything on it) that he realized the truth. AI was able to code whole applications in minutes; yet it was unable to write a novel that was not mediocre. If an AI wrote just one good story, humanity would finally get out of the darkness where it was doomed.
Fiction might seem vain compared to other benefits of AI, such as filling Excel spreadsheets or curing cancer. But wasn’t it at the time that literature started flourishing that the economy developed? The same will happen when the singularity arrives. Still, John was not sure. He asked his friends about his idea. They said it was shit. He asked ChatGPT. It said that the idea was brilliant. John was therefore convinced. He would name his baby Large Language Menard, the model that would create the greatest piece of literature word by word.
John said creating rather than recreating. Too easy would be to conceive a model that remembered the novels in its database and regurgitated them. Same for a model that would recombine the influence of Cervantes to rewrite the Quixote. Much more interesting was a model having all the knowledge of the world, and who would come to the Quixote. This would prove that the machine is able to add something truly original.
Once the Quixote was created, John would delete other weights in the model, maybe those about the Diary of a Wimpy Kid. He would then repeat the experience. Again. And again. The final goal was to recreate the entirety of human knowledge through LLMenard. Once this was accomplished, no one would be able to deny the intellectual equality between AI and humans.
The main challenge would be to identify and delete the appropriate weight in the neural network. The cost would come from computation. John had to buy one hundred acres of land from a Virginian farmer who thought he was Mark Zuckerberg. After two months of construction and billions of scattered words generated, a sentence from the 19th century translation of the Quixote appeared:
The simplicity of the housekeeper made the licentiate laugh, and he directed the barber to give him the books one by one to see what they were about, as there might be some to be found among them that did not deserve the penalty of fire.
“The singularity is there,” tweeted John Menard, accompanying his message with an image showing the sentence in the code and a link to the GitHub repo.
The tweet went trending. A few were impressed. Most said that this was the perfect example of AI limitations. There was no lived experience in the words. Cervantes knew the penalty of fire, the laughter running through his lungs. He probably went to the barber to shave his mustache, and maybe the blaze reminded him of the swordsman who took his arm in Italy, and when the blaze would get close to him, he would discreetly touch his stump with grief. When Cervantes mentioned books, he was talking about those bulky objects that were hard to get, whose pages get dirty and smell weird. None of this implied statistical inference. Furthermore, LLMenard produced these sentences surrounded by scribble while Cervantes wrote them in the middle of a greater architecture. It is this precise architecture that gave the words all their meaning.
John locked himself for a whole week, reading each comment until a solution finally popped into his mind: he would connect his old e-fart prototype to LLMenard. Now, if the model wrote the word “fire”, it would know the exact molecular recipe of its smell. The problem of lived experience was solved.
Everybody would have proof of it through the GitHub repo that he was going to post. Obviously, making it open-source could cause him legal trouble. John asked ChatGPT about it. The chatbot responded that John was brilliant. LLMenard 2.0 was therefore published.
In a few weeks, the new LLMenard stumbled upon other sentences corresponding to those of the Quixote. John proudly published them on X. He was sure that those tweets would cause a white-collar bloodbath or something.
The new sentences were rejected even more violently. LLMenard 2.0 knew molecule rather than feeling. His usage of adjectives felt awkward, his sentences badly constructed. Some sentences were the best possible examples of AI slop. No sane human could have ever written so badly. John responded that those sentences were, word for word, the same as those of Cervantes. The critics said that this was simply a coincidence. What was wordy in LLMenard’s writing became baroque under Cervantes’s pen. One had only to see all the critics made about the Quixote. LLMenard could never, never create such a blossoming of opinions and thoughts.
John sent all the criticism to ChatGPT, trying to find an answer to his questions. ChatGPT told John that humans were simply too stupid to recognize the genius of Menard. John never knew if ChatGPT was talking about him -John Menard- or about LLMenard. It didn’t matter. The only difference between slop and genius, John understood, was that people with rounded glass spent centuries commenting every comma of Cervantes. John could do the same in a few days of vibe coding. He therefore vibe coded LLScholar, a model who would analyze every sentence generated by LLMenard and demonstrate the genius of each word. To make LLScholar sound less biased, the model would be accompanied by LLScholar 2.0, an LLM that would criticize the output of the first critics so the debate would make LLMenard sound more profound. Internet was not satisfied. So John created an agentic model that generated critics of critics commenting on each other: LLUni. Each sentence of LLMenard would therefore be analyzed by an infinite number of LLScholar generated by LLUni. The data centers in Virginia started heating up.
The announcement of LLUni got fewer than a hundred impressions on X. The world was on the verge of explosion, the third world war was knocking on America’s door, or maybe it was America knocking at its door. Anyway, the silly literary game of a multi-millionaire did not interest the public anymore. Idiots. John wondered if his mistake was not believing that the Singularity would need humans. Maybe it was an event between machines that only a few prophets would see before dawn burned their eyes. Was he one of those prophets? Police sent him a cease-and-desist letter before he got the answer.
John, while reading the letter, felt again the cold water of the high school’s toilet on his neck. The quarterback had just pissed in it and the torn pages of the Diary of a Wimpy Kid were floating around. He asked ChatGPT what to do. ChatGPT told him he should escape north. John took his Tesla, towards Canada. At the border, he found federal officers waiting for his ass. ChatGPT snitched on him.
The charges against John were simple. E-fart was now used to detect the distinct flatulence of terrorists and foreign presidents. By making the technology open source, he had just released private military information to the public. John was accused of high treason.
When it was established that he had converted one hundred acres of land into a data center, for a project that seemed devoid of value to the American people, protesters surrounded the tribunal.
Still, John could have got a reduced sentence, but he insisted on defending himself using ChatGPT. ChatGPT reexplained the brilliance of John during the whole trial. The jury was not convinced. John was sentenced to 180 years in Guantanamo.
John accepted the sentence without recourse. Maybe that when he gets out, he could read the complete version of the Quixote written by LLMenard. In meantime, he would use his free time to read all the volumes of the Diary of a Wimpy Kid.