The article reports on a study of billion-dollar AI startups and finds that many publish few peer-reviewed papers, with publication activity concentrated in a minority of firms and often measured by citations rather than fresh output. That framing landed awkwardly because several readers pointed out the paper does not cleanly answer the obvious question of who is still publishing cutting-edge work now, especially since it excludes papers after 2025 and mixes older OpenAI output with the present closed era. A few commenters pulled out the underlying rankings, noting that firms like OpenAI, Anthropic, Hugging Face, Waymo, and Databricks do appear near the top on cumulative citations, which makes the headline sound broader and murkier than the data supports.
The dominant reaction was that there is nothing surprising here. Most AI startups are not research institutions. They are product companies in a market where code is easier to copy, distribution matters more than novelty, and the payoff from publishing can be close to zero once you already have brand, capital, and recruiting pull. Several people argued that publishing is most valuable when you need attention from outsiders. Once a lab can already recruit top talent and raise money, a paper mostly gives competitors free information. That logic felt especially strong in AI because implementation details, tuning, data pipelines, and operational know-how often matter more than the high-level idea described in a paper.
At the same time, people were plainly uncomfortable with what this does to the field. The strongest concern was not that startups are selfish. It was that AI now runs on a strange one-way knowledge economy. Labs were built on openly published papers, open source code, public datasets, scraped web content, and digitized books, yet the newest work is increasingly withheld or reduced to polished blog posts and benchmark claims. That makes the public literature a worse map of reality. It also makes it harder to reproduce results, harder for outsiders to learn, and easier for weak or selectively framed claims to spread under the aura of research.
A lot of the thread treated this less as an AI morality play and more as an incentives story. Academic publishing was described as slow, prestige-driven, expensive, and often hostile to outsiders. In machine learning especially, commenters said top venues reward mastery of field-specific style and social signals as much as substance. For startups, that makes the cost of paper writing feel even less attractive relative to shipping product, filing patents where useful, or just keeping methods private. Several people also noted that
peer review is not a magic quality filter anyway. Poor work gets through there too. So the practical equilibrium is ugly but predictable: serious frontier work stays private, medium-value work becomes marketing content, and public papers increasingly lag production systems.
Where the conversation landed was blunt. The field is not just publishing less. It is splitting into two layers. The visible layer is blogs, demos, model cards, benchmark deltas, and open releases that are safe to share. The invisible layer is the part that likely matters most commercially, including data curation, infrastructure,
post-training tricks, evaluation practices, and system integration. That means anyone trying to understand AI progress through papers alone is watching only the showroom floor, not the factory.
science.org