The AI Labs Stopped Publishing — And Nobody Noticed

In June 2017, eight Google researchers posted a paper to arXiv called Attention Is All You Need. That paper described the transformer architecture. Without it, there is no ChatGPT, no Claude, no Gemini. Google charged nothing for it. Nine years later, the companies that built hundreds of billions of dollars in value on top of that paper will barely tell you what they’re doing.

Technical Reports That Are Really Marketing Decks

The page count didn’t shrink. The content did.

The 2020 GPT-3 paper ran 75 pages. Parameter counts, training data mix, batch sizes, learning rate schedules — it was all in there. A motivated grad student could read it and build a scaled-down version.

The 2023 GPT-4 technical report ran 98 pages. Longer, and yet it contained no model size, no architecture, no training data description. The first page said the omissions were due to “the competitive landscape and the safety implications of large-scale models.” Most frontier model documentation since has followed that template: benchmark charts, a safety evaluation summary, a narrative about how much better this model is than the last one. Nothing you could reproduce.

The industry stopped calling these papers. They’re model cards now, or system cards. The renaming matters more than it sounds. A paper is a submission for scrutiny. A card is a press release with a table in it.

Three Reasons the Door Closed

Money, first. A single frontier training run now costs in the hundreds of millions. Papers used to be the company asset. Now the training recipe is the asset — which data, in which order, what got rewarded during the RL phase. That is the moat. Publishing it is a charitable donation to your closest competitor.

Talent, second. Papers used to be a recruiting tool. Land at NeurIPS, and good people came to you. The pressure now runs the other way. A researcher with their name on a frontier paper becomes a target for every competing recruiter within a week. Since 2024, reported signing packages for frontier lab researchers have crossed into professional-athlete territory. Publishing doesn’t get you talent. It gets your talent poached.

Legal, third. Documenting your training data in detail is handing plaintiffs their discovery for free. With The New York Times case and a stack of others in motion, a paper that specifies your dataset is a liability with a DOI. There’s probably genuine safety conviction mixed in there somewhere. It also happens to align perfectly with what counsel would advise anyway.

Universities Eat the Cost First

Academic labs take the hit before anyone else. You cannot do follow-up work on something you can’t see inside. To study why a model fails in a particular way, you need at minimum to know its architecture. Without that, you’re poking an API and cataloguing surface behavior. Academia has been reduced to doing black-box psychology on the most important artifacts in computing.

The deeper casualty is reproducibility. The whole premise of science is that someone else can rerun your work. Benchmark scores with no method attached are a claim, not a result. Right now the most influential outputs in AI arrive un-peer-reviewed and unreproducible. In physics or biology that would be a scandal with a Wikipedia page.

Safety research has the same problem, only worse. Verifying a risk from the outside requires seeing inside. The labs closed access citing safety — which means the only people equipped to audit safety now work at the labs. The student is grading their own exam.

But It’s Not Uniformly Closed

Frontier labs haven’t stopped doing research. In some areas they’re prolific. Interpretability is the obvious one. Anthropic publishes steadily on model internals, and work like extracting features with sparse autoencoders comes out in real detail. OpenAI ships alignment research too.

The pattern is legible: publish what isn’t a moat, close what is. Interpretability findings don’t help a rival train a better model. They do buy reputation for taking safety seriously, and they read well in Brussels and Washington. Training recipes are pure competitive transfer the moment they’re public.

China’s labs walked straight into the gap. When DeepSeek shipped models with detailed technical reports in late 2024, what startled Western researchers wasn’t only the performance. It was that the method was written down. Since then, Chinese open-weight releases — Alibaba’s Qwen family among them — have become the actual substrate of academic work. The best American models are opaque. The Chinese models are legible. Ten years ago, almost nobody would have predicted the openness axis flipping in that direction.

Where This Goes

There’s a precedent. Bell Labs opened in 1925 as an AT&T subsidiary, invented the transistor, and published the papers. It could afford basic research because monopoly rents were fat, and it had regulators to placate. In industries where competition actually got brutal, corporate research labs either turned inward or disappeared entirely.

AI is the second case. Competition is savage, winner-take-most is plausible, and training costs are astronomical. Expecting voluntary openness under those conditions is wishful thinking. If it changes, it changes from outside: regulation like the EU AI Act, which requires publishing a summary of training data; compelled disclosure in a large lawsuit; or transparency strings attached to publicly funded compute.

One more thing. This topic barely registers as a conversation. Sweep the last month of community discussion and there’s almost nothing on it. A benchmark leaderboard shuffle draws thousands of comments. A new model release draws thousands more. The disappearance of the scientific record draws roughly zero. That may be the most accurate summary of the problem available: it closed quietly, while nobody missed it.

The Takeaway

The modern AI industry exists because the transformer paper was free. That industry has apparently decided not to pass the same gift down. The closure is defensible as a business decision — I’ll grant that without argument. But individually rational choices adding up to a collectively bad outcome is one of the most common shapes a problem takes. When an AI researcher in 2035 looks back at this period, will they be able to tell what was lost? They won’t. That’s the point.