AIArticle A Stanford-led analysis puts hard numbers on the frontier's retreat from publishing, and on who filled the gap.
Mariana Souza The vibe has been obvious for a while: frontier AI labs raise billions, ship models behind APIs, and tell you how they work in a blog post. Now there's data. A new analysis led by Stanford metascientist John Ioannidis — reported this week by Science — looked at all 317 AI unicorns that have existed since 1998 and found that more than half have never played a leading role in a single scientific paper or preprint. Collectively, companies worth over a billion dollars each accounted for roughly one in every 1,000 AI papers published in 2025.
Ioannidis knows this pattern. He was the researcher who, back in 2015, publicly flagged that Theranos had essentially no peer-reviewed footprint. His verdict on AI's version of stealth research: for a field supposedly reshaping science, having almost no scientific documentation is "a very weird paradox."
He's right that it's a paradox. But for developers, the more useful question is what replaced publication — because something did, and it's reshaping where you learn new techniques from.
Read the fine print before quoting the stat #
The study (a preprint, not yet peer-reviewed itself — savor that) counted journal articles, conference papers, and preprints where a company researcher was first or last author. That's a deliberately narrow lens, and it excludes most of what frontier labs actually put out: system cards, model cards, evaluation blog posts, open-weight releases with technical reports where authorship is a company-wide alphabetized list. By this measure, OpenAI — about 4,500 employees — has only eight researchers who've authored five or more qualifying papers.
So the "1 in 1,000" figure overstates the silence somewhat. What it doesn't overstate is the concentration. The top 5% of firms account for more than 90% of all citations in the dataset, and OpenAI alone claims nearly 40% — followed by Chinese computer-vision firm Megvii and Hugging Face. Here's the detail worth dwelling on: OpenAI's citation dominance rides almost entirely on papers from its publishing era — GPT-3, InstructGPT, CLIP, Whisper. The lab is coasting on the reputation of a practice it abandoned. The GPT-4 technical report in March 2023 marked the turn explicitly, declining to disclose architecture, model size, training compute, or dataset construction, citing competition and safety. Every major US lab followed.
Why the incentive flipped #
The comparison the study draws with pharma is the load-bearing one. A drug company can publish because a patent protects the molecule; disclosure and monopoly coexist. Model weights have no equivalent protection. Publish your data pipeline and RL recipe and a competitor replicates you in months — DeepSeek proved that replication is cheap when the method is known. The only thing publication bought US frontier labs was recruiting credibility, and once compensation packages hit seven and eight figures, they didn't need papers to hire. The talent-market pressure that forced Google to publish "Attention Is All You Need" in 2017 — the paper the entire current industry is built on — simply no longer exists at that tier.
That's the uncomfortable ledger: the transformer, RLHF, and scaling laws were all public research. The companies that benefited most from openness are the ones who closed the door behind them.
The documented frontier moved — mostly to China #
The study's geographic finding matches what any practitioner reading arXiv has noticed: Chinese firms consistently publish more than their US counterparts, and they've paired papers with open weights. DeepSeek took it furthest — its R1 paper went through three rounds of peer review and landed on the cover of Nature in September 2025, the first mainstream LLM to do so, complete with a disclosed $294,000 RL training cost. Alibaba's Qwen team ships detailed technical reports with every release. For a challenger, publication is a distribution strategy: you can't out-market OpenAI, but you can out-document them, and developer trust follows reproducibility.
The result is a two-tier knowledge economy. The capability frontier is closed and American; the documented frontier is open-weight, six-to-twelve months behind, and increasingly Chinese — with a thin fully-open academic layer (Ai2's OLMo, Hugging Face's training runs) that publishes everything including data. If you want to actually understand how a modern reasoning model is trained, your syllabus today is DeepSeek, Qwen, and Kimi tech reports, not anything out of San Francisco.
What this costs you in practice #
Three concrete losses. First, benchmark claims from closed labs are now unfalsifiable — no training-data disclosure means no contamination checks, so "state of the art on SWE-bench" is marketing until an independent eval reproduces it. Budget your model-selection decisions accordingly: weight third-party evals and your own task-specific harness over launch-post numbers. Second, technique replication lags. When a closed lab ships a capability jump, the open ecosystem reverse-engineers it from behavior, which is slower and lossier than reading a method section. Third, the learning path for engineers has changed shape: the skill of reading papers matters less than the skill of reading open-weight tech reports and, frankly, the code and configs on Hugging Face.
Is the retreat from publication a genuine structural shift? Yes — the incentive analysis says it's rational and therefore permanent for closed labs; nothing short of regulation or a talent-market reversal brings the old norms back. But the study's implicit framing, that AI research has gone dark, is wrong. It's gone asymmetric. Publication didn't die; it migrated to the players who profit from openness. The practical move for developers is to treat that migration as settled: build your understanding, your evals, and your fallback plans on the documented open-weight stack, and treat closed-lab claims as vendor claims — because that's now precisely what they are.
Sources & further reading #
[AI's top startups are barely publishing their research](https://www.science.org/content/article/ai-s-top-startups-are-barely-publishing-their-research)— science.org -
[AI's top startups are barely publishing their research (discussion)](https://news.ycombinator.com/item?id=49103285)— news.ycombinator.com -
[Secrets of DeepSeek AI model revealed in landmark paper](https://www.nature.com/articles/d41586-025-03015-6)— nature.com
[Mariana Souza](https://sourcefeed.dev/u/mariana_souza)· Senior Editor
Mariana covers the fast-moving world of machine learning and generative AI, with a particular focus on how these technologies are reshaping development workflows. When she isn't stress-testing the latest foundation models, she's usually at a local hackathon.
Discussion 0 #
No comments yet
Be the first to weigh in.