---
title: "LLM: Listening, Learning, Modeling"
description: "I have suspected for a while that my AI is listening. Not in the sci-fi sense. In the business sense. This month a mathematician at NYU got the proof I could not get, and it is sitting in a settings menu."
canonical: https://www.askgib.com/articles/llm-listening-learning-modeling
last-updated: 2026-09-21
lang: en
---
# LLM: Listening, Learning, Modeling
> I have suspected for a while that my AI is listening. Not in the sci-fi sense. In the business sense. This month a mathematician at NYU got the proof I could not get, and it is sitting in a settings menu.
By Gibril Bachouchi. Filed 2026-09-21. 6 min read. Web page: https://www.askgib.com/articles/llm-listening-learning-modeling
Languages: [en](https://www.askgib.com/articles/llm-listening-learning-modeling.md) · [fr](https://www.askgib.com/fr/articles/llm-listening-learning-modeling.md) · [ar](https://www.askgib.com/ar/articles/llm-listening-learning-modeling.md) · [es](https://www.askgib.com/es/articles/llm-listening-learning-modeling.md) · [ru](https://www.askgib.com/ru/articles/llm-listening-learning-modeling.md) · [pt](https://www.askgib.com/pt/articles/llm-listening-learning-modeling.md) · [it](https://www.askgib.com/it/articles/llm-listening-learning-modeling.md) · [de](https://www.askgib.com/de/articles/llm-listening-learning-modeling.md) · [ja](https://www.askgib.com/ja/articles/llm-listening-learning-modeling.md) · [ko](https://www.askgib.com/ko/articles/llm-listening-learning-modeling.md) · [hi](https://www.askgib.com/hi/articles/llm-listening-learning-modeling.md) · [he](https://www.askgib.com/he/articles/llm-listening-learning-modeling.md) · [zh](https://www.askgib.com/zh/articles/llm-listening-learning-modeling.md) · [tr](https://www.askgib.com/tr/articles/llm-listening-learning-modeling.md) · [kab](https://www.askgib.com/kab/articles/llm-listening-learning-modeling.md)
I have suspected for a while that my AI is listening.
Not in the sci-fi sense. Nobody is sitting in a room reading my chats. In the business sense. What I type goes somewhere, gets learned from, and gets watched for anything worth having. Every builder I know has the same feeling, and until this month none of us could prove it.
Then a mathematician at NYU got the proof I could not get. Not because anyone confessed. Because the company he asked could not say no.
So here is what the three letters stand for, as far as I can tell. Listening. Learning. Modeling. And here is the story that made me stop calling it a suspicion.
You have seen the name on the news this month. Navier-Stokes. Here is what it actually is, without the maths.
Take anything that flows. Water in a pipe. Air over a wing. Blood through a heart. Smoke off a candle. The Navier-Stokes equations are the rule for how that stuff moves: push here, and this is how fast, and in which direction, everything downstream will go. They are Newton's law, force equals mass times acceleration, rewritten for things that have no fixed shape.
That sounds academic until you notice what depends on it. Every aircraft wing is shaped with these equations. Every weather forecast is these equations run forward a few days. Every ocean model, every pipeline, every ventilation system, every simulation of blood moving past a valve before a surgeon touches it. They are two hundred years old and they run the physical world.
Now the strange part. Nobody has proved they always work. In three dimensions, nobody can show that a calm, smooth flow stays smooth. It might, in a finite amount of time, "blow up": somewhere the speed shoots to infinity and the equations stop meaning anything. Picture a whirlpool that tightens forever at a single point. Nobody knows if the real equations allow that, and since 2000 the Clay Institute has offered a million dollars to whoever settles it, either way.
Why should you care? Because turbulence, the thing that shakes your plane and ruins your forecast, is still not understood, and because a lot of engineering rests on equations we cannot yet prove are well behaved. Anyone who works in fluid dynamics knows this. For everyone else, it is enough to know that the machines just went after one of the last big unknowns in physics. And that this is what they went after first.
Tristan Buckmaster is a mathematician at NYU. Levent Alpöge is a mathematician who works at Anthropic and was collaborating with Buckmaster on his own time. They had picked a specific route to the problem, blow-up with a smooth external force, and by Buckmaster's account almost nobody else was on it. They used AI to work it: Claude, and OpenAI's coding tool. In mid-August they had results. A week later the proofs were machine-checked in Lean.
Then Alpöge tweeted something happy and vague about a great month. Rumours started that Anthropic had cracked a Millennium problem. Within a day, by OpenAI's own account, OpenAI pointed roughly ten thousand agents, driven by an internal model stronger than anything they sell, at Navier-Stokes.
A few days later, Buckmaster says, his collaborator got tips that news of their progress had reached OpenAI. He emailed a mathematician there. The reply offered free compute. Then came two phone calls with an OpenAI researcher, who told him the internal model had proved forced blow-up for Navier-Stokes. The exact two cases he and Alpöge had been quietly attacking.
He asked whether the model had been trained on, or had access to, their sessions in OpenAI's coding tool. He says he was told the model did not look up user data. He asked again, about training. He says he did not get an answer.
He was offered a deal. Either they publish their Euler result first and OpenAI publishes the next day, or he writes up OpenAI's result himself, without Alpöge, because Alpöge works at Anthropic. He says the researcher asked twice for Alpöge to be removed. When he said he would go public instead, he says he was asked: "Why would you ruin your career?" OpenAI's researcher disputes his account of the authorship discussion.
He published anyway, with a statement, and Terence Tao wrote it up the same day. OpenAI announced its result the day after: eighty-eight hours, 2.7 million messages, a Lean-checked proof, and a line saying the run began after Twitter rumours that Anthropic had solved Millennium problems. Three days later, twenty-five Fields Medalists signed a letter titled "A Severe Misalignment of AI in Mathematics". And OpenAI, asked directly, said no specific user data was accessed, then added that it "cannot rule out that de-identified data derived from their usage of our products helped improve our models".
Read that twice. They are not saying they did not learn from his work. They are saying they cannot tell.
Start with the boring part, because the boring part is where it lives.
If you use ChatGPT on a Free, Plus or Pro plan, your conversations train the model by default. You have to go into Settings, then Data Controls, and turn off "Improve the model for everyone". The coding tool has a separate toggle that the ChatGPT one does not touch. Claude's consumer plans have trained by default since last August, with a five-year retention if you leave it on. Gemini too. Only the API and the enterprise tiers are off by default, because companies with lawyers asked. Everyone else opted in by not opening a menu.
Turn it off and they still keep your chats for thirty days, and people at the company can still read them if a system flags you. A court can order the whole archive kept. One did, for most of last year, and twenty million de-identified conversations were handed over in a lawsuit anyway.
So yes. It is listening. It says so in the settings. The strange thing is not that it listens. It is how many people who build their livelihoods on it have never opened the menu.
Listening is passive. Learning is what makes it worth doing.
Last week I wrote that the labs have collected the most valuable training set ever assembled, for free, from their own customers: not the raw data, the shape of the work. Which questions get asked. Which answers get accepted. Where the bottleneck is. What a good result looks like. A lab that has watched ten thousand teams work on something does not need to read any one of them. It has already absorbed what they know.
That is exactly what OpenAI's sentence describes. Not "we read your session". Something quieter and worse: "your usage may have helped improve our models". The learning is diffuse. It cannot be traced back to you. Which also means it cannot be ruled out, ever, by anyone, including them.
I used to think the danger was a leak. It is not a leak. It is a diet. Your work goes in, and it comes back out stronger and wearing someone else's name.
Here is the part people push back on, so let me be careful.
Listening gives them the words. Learning gives them the skills. Modeling is the third step: building a picture of who is doing what, and how close they are.
These companies already read conversations at scale. That is not a theory, it is a product feature. Every prompt passes through classifiers looking for abuse. OpenAI publishes reports about accounts it caught and banned for misuse. So does Anthropic. You cannot catch what you do not read. A pipeline that can tell, across millions of conversations, who is trying to build a weapon can also tell who is trying to prove a theorem, and how far along they are. It is the same machinery pointed at a different question.
Now put that next to what happened. A vague tweet and some gossip was enough to point ten thousand agents at one problem within a day, and to point them at the right formulation of it. Nobody needed a chat log for that. They needed a model of where the field was, and a model of where one small team sat inside it. Knowing what is hot, and who is close, is precisely what a picture built from millions of conversations gives you.
I am not saying that is what happened. Buckmaster is not saying it either, and he is the one with the receipts. I am saying the machinery exists, it runs on every message, and the company that owns it will not tell you what it models. When the dial is hidden, every explanation is available, including this one, and the company that hid the dial has no right to complain when people pick it.
A lot of people who build with these tools have the same feeling. When you are on a boring task, the model is fast and generous. When you are onto something, it gets slow, vague, and cautious. It starts asking you to clarify things it understood five minutes ago. And the paranoid reading is that it is stalling you while it learns what you are doing.
I feel it too. I cannot prove it, and the honest evidence points elsewhere. In July, people fingerprinted the hidden reasoning budget of OpenAI's top model and saw it drop from 960 to 128 overnight, for everyone. Only then did OpenAI admit a change. They called it temporary "parameter tuning". The router that decides whether you get the fast model or the thinking model is theirs, not yours. Weekly caps exist. Every one of those knobs turns for cost.
So the model is probably not slowing down because you are clever. It is slowing down because you are expensive. But I cannot tell the difference from my side of the screen, and neither can you, and that is the whole point of this piece. Listening, learning and modeling are all invisible from the outside. You only find out when a mathematician with a Lean proof asks the question and the answer is silence.
Three things, none of them dramatic.
**Turn the toggle off.** Today. In ChatGPT, Settings, Data Controls, "Improve the model for everyone". In the coding tool, separately. In Claude, "Help improve Claude". In Gemini, activity. It takes two minutes and it removes the easiest path.
**Keep the real work where they cannot see it.** The generic parts of a job can go to any API. The part that makes you you, the idea, the data, the route almost nobody else is on, belongs on a model you can download and a machine you control. That is not paranoia. That is what Buckmaster would tell you now.
**Publish early.** The thing that saved these two was a Lean proof with a date on it and a statement on a university server. Timestamps beat swarms. If you are close to something, get the date on the record before you get the phone call.
Your LLM is listening. It is learning. It is modeling you. None of that is a secret. It is in the settings menu and the transparency reports. The only open question is what you keep giving it.
He published anyway. So should you.