Habsburg AI: What Happens When Models Train on Models

The internet is running out

In 2024, the biggest AI labs in the world quietly admitted they were running out of internet. Not bandwidth — data.

Every meaningful piece of human-written text on the public web had been scraped, cleaned, deduplicated, and fed into training pipelines. Frontier models now train on tens of trillions of tokens. Llama 3.1, the last model Meta disclosed a corpus size for, used roughly fifteen trillion. The entire estimated stock of high-quality human language on the internet? Somewhere between four and seventeen trillion tokens.

Epoch AI projects high-quality text exhaustion between 2026 and 2032.

Meanwhile, global private AI investment in 2025 hit $345 billion. Half of that — $170 billion — went into generative AI alone. The thing all that money was betting on was running out of fuel.

The answer every frontier lab converged on is the same answer DeepMind worked out on a Go board in 2017. Train the model on data the model generates itself.

Self-play. Synthetic data.

It is the most consequential bet in modern AI. And every condition that made it work in 2017 is, at best, partial in 2026.

The world before AlphaGo Zero

To understand why this matters, you have to understand what changed.

Up to 2016, every serious machine learning system was tethered to human data. ImageNet — fourteen million images, hand-labeled by humans — was the foundation of computer vision. Translation models trained on parallel human translations. Even Deep Blue, the system that beat Kasparov in 1997, was built on a database of grandmaster games. Humans produce the data. Machines learn the patterns. The ceiling on the machine was the ceiling of what humans had written.

In March 2016, AlphaGo broke the contract — partially. It beat Lee Sedol four games to one. Move 37 of game two, a placement so unconventional that commentators assumed it was a mistake, announced that something had changed. But AlphaGo was still seeded with 160,000 human expert games. The human anchor was still there.

October 2017: Zero human data

DeepMind publishes a paper: “Mastering the Game of Go Without Human Knowledge.” The system is AlphaGo Zero. The training dataset is exactly one thing — the rules of Go.

No human games. No expert moves. No opening books. Random weights, rules, and self-play. Millions of games against itself. The wins reinforce the network. The losses penalize it.

• In 72 hours, it surpasses the version that beat Lee Sedol
• In 40 days, it defeats every prior version of AlphaGo — 100 games to zero

Read that again. The version trained on zero human games beats the version trained on millions. One hundred to zero.

The lineage runs forward

Two months later: AlphaZero. Same architecture, applied to chess and shogi. Within 24 hours on each game, it defeats Stockfish — the strongest chess engine humans had spent decades tuning.

From here, the lineage runs directly into the systems you use today. AlphaFold solving protein structures. MuZero learning rules it was never told. Tesla running self-driving against synthetic edge cases. OpenAI simulating millions of robot hand manipulations before touching physical hardware.

Synthetic data arrives at language

By 2024, the technique arrived at language models.

OpenAI’s o-series reasoning models — o1, o3, o4-mini — are trained via reinforcement learning on millions of reasoning traces the model itself generates. Long chains of thought, graded by automated reward models and human feedback, used as training data. By 2025, the top frontier reasoning models were scoring above 95% on the American Invitational Mathematics Examination — a test historically passed by the top 1% of competition math students.

Meta took it further. Llama 4 Behemoth — two trillion parameters, 288 billion active — was never deployed as a product. It exists solely to generate synthetic training data for smaller models. A teacher that never talks to users. The entire point of the largest model Meta has ever built is to make the other models better by producing their training sets.

That is how far the synthetic data bet has gone.

The Phi-4 proof

The most compelling evidence that synthetic data works for language is smaller and more precise.

Microsoft’s Phi-4 — fourteen billion parameters. Trained primarily on synthetic data: 400 billion tokens, 50 types of synthetic datasets covering chain-of-thought reasoning across math, science, and code. It outperforms models 10× to 50× its size on graduate-level STEM benchmarks.

To prove this isn’t memorization, Microsoft tested it on the November 2024 AMC math competition — questions released after training ended. Phi-4 beat frontier models that are orders of magnitude larger.

Why AlphaGo Zero worked — and why language is harder

Now the architect’s lens. AlphaGo Zero worked because Go has three properties almost nothing else has:

Self-play works because the model is grinding against a perfect oracle — the rules — that checks every move and issues an uncontested verdict after every game. Errors do not accumulate. The system cannot drift, because the oracle keeps pulling it back to reality.

Language has no such oracle.

The honest answer: synthetic data works best where a domain has its own verifier. Code that compiles. Math that proves. Games with rules. This is why Phi-4’s gains are concentrated in STEM — those domains have verifiers. It’s why the o-series excels at structured reasoning. The closer a domain gets to “open-ended judgment,” the more the AlphaGo analogy breaks down.

The shadow: Model collapse and Habsburg AI

The field has a name for what happens when you train on the output of models trained on the output of models. Model collapse. Or, in the more vivid term that has stuck in the research community — Habsburg AI.

The historical parallel

The Habsburg dynasty is the case study in what happens when a closed loop reinforces itself too long. Inbreeding over generations — traits that started as minor variations became exaggerated, fixed, and eventually harmful. The famous Habsburg jaw.

Habsburg AI is the same dynamic in distributional form: minor errors in generation get baked into the next model’s training set and amplified by its successor. The distribution progressively converges to a smooth, confident, average bell curve. Distinctive reasoning disappears. The system gets more fluent and less correct.

The foundational research

Shumailov et al., Nature 2024 — demonstrated model collapse on text and image models across multiple generations. The tails of the distribution, where rare but important information lives, get cut off first. The model appears healthy by standard loss metrics. The collapse is silent.

Strong Model Collapse, ICLR 2025 — hardened the finding. Even one synthetic token per thousand in a training set can still cause degradation. Larger training sets in this regime do not help. More data made from contaminated data is still contaminated data.

The critical nuance

The research is not uniformly bleak. A separate body of work, including NeurIPS 2024 and 2025, found that model collapse is primarily triggered by one specific practice: mass deletion of prior real data, replaced with synthetic data.

If you accumulate — if you add synthetic data alongside the real data rather than substituting it — the collapse does not materialize at the same rate. The factual anchor of the original human data stabilizes the loop.

Which means the labs that scraped the internet before 2023 — before the web filled up with AI-generated content — have a structural advantage. They own the largest pools of pristine human data.

By January 2025, more than half of all newly published online content was estimated to be AI-generated. The AI Index 2026 — Stanford’s annual stocktake of the field — puts it bluntly:

There is still no definitive evidence that synthetic data can fully offset real data depletion in pre-training.

The pristine era hasn’t just closed. It inverted.

The $345 billion open question

DeepMind solved the data problem for Go in 2017. Whether the same trick generalizes from a 19×19 board to the full surface of human language — that is the most expensive open question in computing.

Global AI compute capacity has been growing 3.3× per year since 2022. Meta built a two-trillion-parameter model that no one will ever talk to, purely to generate data for the models you will. The synthetic data bet is not a research paper. It is running, right now, at a scale measured in gigawatts.

The Architect’s Lens takeaway

Three things from a Go board to a 2026 AI strategy:

1. Synthetic data is not a free lunch — it’s a domain-conditioned tool. Where a verifier exists, it will keep paying off. Code against compilers. Math against proof checkers. Structured reasoning against known-answer benchmarks. Where the grader is another model and the domain is open-ended, you’re running an experiment, not following a recipe. The Phi-4 result is not a general license for synthetic data — it’s evidence that STEM domains have enough oracle structure to make it work.

2. Accumulate, do not replace. The research consensus on collapse is that the practice that triggers it is deleting real data. If you’re building a system that trains on synthetic output, treat the original human data as the factual anchor and never retire it. The loop needs something outside itself to pull it back to reality — the same principle as the AlphaGo oracle.

3. We are inside the experiment. Pay attention to which models stay sharp and which start sounding fluent and slightly wrong. That is the leading indicator. $345 billion a year says synthetic data works. The Habsburg jaw says: be careful.