Sergey Markov had just finished a serious point about AI being a double-edged tool, a hammer that builds houses or splits heads, when he said something else. We are surely living in a training simulation, he told the host. Some post-singularity society has to raise its children and teach them to answer hard ethical questions, so it built this place, and we are the ones "dangling around in it."
Markov heads the experimental machine learning systems directorate at SberDevices, part of Sberbank. He called the training-simulation theory a funny story, and added that you can invent others. Then he invented one from his own workbench.
His problem is text. The datasets used to train models are growing faster than the amount of text on the internet. At some point the models will want more than exists, and the shortage will reach everyone who talks to a chatbot. Epoch AI put a date range on it in 2024: if trends continue, language models will use up the stock of public human-written text somewhere between 2026 and 2032.
You can't write the internet again. One way out, Markov says, is to make the missing data yourself: run simulations, produce synthetic examples, train the model on those as well.
Then he turned the idea around and pointed it at us:
Maybe we simply live in a world that generates synthetic data to train some model.
Может быть, мы просто живем в мире, который генерирует синтетические данные для обучения какой-то модели.
— Sergey Markov, Gleb Solomin Podcast, 04:10
Look at what changed between the two stories. In the first, we are pupils, and the simulation exists to teach us something. In the second, nobody is teaching anyone. We are the pages being produced, so that something else can read them.
Markov doesn't push it further. It is an aside, offered in the same breath as the first story and with the same shrug, and he doesn't claim evidence for either. The conversation moves on to the next question.
Still, the data shortage is not a joke, and synthetic data has its own trouble. A Nature paper by Ilia Shumailov and colleagues in July 2024 found that training indiscriminately on model-generated content causes irreversible defects in the resulting models: the tails of the original content distribution disappear. The rare, odd, unlikely material goes first. Whatever a world-sized generator turned out, a model would need the odd material to stay in.
Markov leaves it at two stories. In one we are children being taught how to be decent. In the other we are the text.
