AI's top startups are barely publishing their research
AI’s current boom rests heavily on past openly published research, yet many leading startups now guard new advances as proprietary trade secrets rather than peer‑reviewed science. Commenters debate whether this shift is a rational response to fierce commercial and geopolitical competition or a worrying erosion of scientific norms that slows collective progress and concentrates power. Related themes include the rise of blog posts over journals, the limits of patents, contrasts between Chinese and US publication habits, and the irony that companies built on public data and papers now withhold their own work.
Foundational openness vs current secrecy
- Many note the irony that today’s frontier AI firms rely on foundational work like transformers and knowledge distillation that were openly published, while now hoarding their own advances.
- Some argue early openness came from underestimating the architectures’ potential and valuing prestige and hiring more than secrecy.
- Historical analogies appear (radiation before WWII, early industrial chemistry/dyes): once research becomes obviously lucrative or strategic, it moves behind corporate walls.
Incentives: profit, competition, and game theory
- A recurring view: companies exist to make money, not advance science; once AI is commercial, incentives shift from openness to IP protection.
- Others argue this is socially harmful: if all withhold, overall progress and ROI on massive AI infrastructure may suffer.
- Some suggest changing legal frameworks (weaker trade secrets/NDAs, shorter patents) to force more contribution to the commons, while critics respond that without strong private incentives, multi‑billion‑dollar R&D wouldn’t happen.
Academic publishing vs “blogified” research
- Several posters say formal peer review is slow, political, and prestige-driven; for fast‑moving ML, self‑publishing or blogs can communicate results more efficiently.
- Others counter that peer review, while imperfect, adds rigor, reproducibility checks, and shared standards; “blogified” AI research encourages hype, sloppy methods, and social‑media‑like dynamics.
- Some startups report giving up on tier‑1 venues due to repeated rejections and time cost, using preprints or pitch‑deck‑only sharing instead.
Trade secrets, IP, and open source
- Strong defense of trade secrets: in a world where competitors can copy quickly (especially using models), publishing frontier work is seen as “incredibly stupid” for startups.
- Others emphasize societal benefits of openness, protection from patenting by others, and talent attraction from visible publications.
- Debate around patents: some claim software/ML patents are largely unenforceable and thus not worth filing; others see them as still relevant in some domains.
Who publishes and where
- The underlying study cited in the thread reportedly finds that some large AI startups (including US labs and Chinese firms) are still major paper producers, but newer “startup” layers mostly focus on products over research.
- Several comments note that Chinese AI companies now publish more openly than many US counterparts, including strong open models and released weights.
Ethics, data, and the commons
- Posters criticize AI firms for scraping public data, digitizing (and sometimes physically destroying) books, and then closing off their own outputs.
- There’s tension between “information wants to be free” and creators who feel exploited when their work fuels proprietary systems without reciprocal openness.