A new study finds most AI unicorns have never published a single research paper. Here's why that's not actually shocking.
What the article says
- A new analysis of over three hundred AI unicorn companies found that more than half have never led a scientific paper or preprint.
- Influence is wildly concentrated. A tiny sliver of firms account for almost all the citations, and OpenAI alone accounts for close to forty percent of them.
- Even at OpenAI, output comes from a small circle of repeat authors despite the company employing thousands of people.
- Researchers say startups get little competitive benefit from publishing, since ideas can be copied freely and patents rarely apply.
- Chinese AI firms publish far more than their American counterparts and lean into open weights, while US labs increasingly keep models closed.
- Some researchers argue the real question isn't papers versus blog posts, but whether companies release enough code and data for others to verify their work.
What HN is saying
- Several commenters say this isn't surprising at all. Most AI companies are commercial products, not research shops, so why expect papers.
- A recurring argument is that publishing mainly helps you when you have no other way to reach smart people. Once you can hire anyone, sharing your results just helps competitors and rivals like OpenAI or Anthropic copy your work for free.
- One person pushes back that copying is supposedly the whole point of publishing research in the first place.
- Others note the article is oddly vague about which companies actually are skipping publication, since OpenAI, Anthropic, and Hugging Face are all named as ones that do publish.
- A physics student shares that writing up independent research, even unpublished, helped them meet other researchers and led directly to new collaborations.
A guy building a house used the Vision Pro to walk through the floor plan before a single wall existed.
What the article says
- A developer building his first house turned floor plan PDFs into a walkable 3D model since flat drawings never gave him a real feel for scale.
- He modeled the walls in free software, then pulled in furniture from IKEA and a 3D model community to make rooms feel properly furnished instead of empty.
- He and Codex vibe coded a custom viewer for the headset, adding controller movement, a skybox, terrain following, and a toggle to see the real room underneath.
- Walking through the finished house in headset felt more convincing than the polished walkthrough his architect's professional software produced.
- He put the janky but working project on GitHub for anyone who wants to try the same trick on their own house plans.
What HN is saying
- Several architects and builders said this is already standard practice, using Quest headsets with design software to walk clients through homes before construction and to catch layout mistakes early.
- The sharpest pushback was that none of this needs a Vision Pro specifically, cheaper headsets or even a phone with ARKit can do the same walkthrough.
- One person pointed out the irony that the author named his tool Prospector, unaware an existing VR design tool is literally called Prospect.
- A popular add on idea was simulating sunlight through the year to catch which rooms get too hot or too dark before building.
- Others shared their own version of the trick, using Meta Quest and Unreal Engine, and said it cured their decision paralysis during a renovation.
NSF is testing a four year PhD where a company, not just a university, helps train you.
What the article says
- The NSF is putting forty seven million dollars into a pilot with dozens of universities and companies to reshape the PhD.
- The idea is that most STEM PhDs now end up in industry jobs anyway, so training should reflect that instead of assuming everyone becomes a professor.
- Students get their first year funded by their university, then NSF and an industry partner cover the rest, including at least a year working on their thesis inside a company.
- Students are co-advised by an academic mentor and an industry mentor, and the whole thing is meant to fit into four years instead of the usual five or six.
- The first students would start in Fall 2026, with the plan to grow the program in later years if it works.
What HN is saying
- Commenters point out this already exists elsewhere, Germany, China, Japan and the UK all have versions of industry linked doctorates, so the US is catching up rather than inventing something new.
- The sharpest disagreement is about motive. Some see it as a genuine fix for a real mismatch between how PhDs are trained and where they actually work. Others think it is ideological, an attempt by the current administration to shift money and prestige away from universities and toward industry.
- A few worry this just means more research locked up as private company property instead of published openly.
- Some people who lived through industry research push back less cynically, saying real funding and a clear path to a job outweigh the downsides of academia's politics and low pay.
- Practical worry from someone who has done a similar setup, students need real independence in these placements or the year risks becoming free labor rather than research.
A solo dev got a 26 billion parameter AI model running on a laptop with barely any memory to spare.
What the article says
- A developer built a custom engine that runs a large Gemma language model using only about two gigabytes of memory, on any recent Mac.
- The trick is not loading the whole model at once. It keeps a small shared core in memory and streams the rest from the hard drive as needed, timed so the chip is never left waiting.
- It took over a hundred failed experiments to get the timing right between reading from disk and computing on the chip.
- Speed varies a lot by machine. Older, lower memory Macs get a slow but usable chat experience, while the newest chips run several times faster thanks to much quicker storage.
- The whole thing is open source and free to try, though the first run downloads about fifteen gigabytes of model weights.
What HN is saying
- Commenters were impressed that a fourteen gigabyte model can run in a sliver of that in RAM, and dug into why: only a fraction of the model's knowledge gets touched per word generated.
- Several people questioned why frontier AI labs do not do this kind of trimming already. The answer, others explained, is that nobody fully understands which specific parts of a model's weights matter for which answers, so you cannot safely leave pieces out.
- The sharpest exchange was technical. Someone asked why an existing tool that already supports minimal memory usage would be slower, and the author explained his version deliberately schedules disk reads to overlap with computation instead of leaving it to chance.
- A few readers ran the code themselves and reported real world numbers, generally confirming newer chips see a big jump thanks to much faster storage, not just a faster processor.
- Some commenters called out stilted phrasing in the writeup as a sign it was cleaned up by an AI writing assistant, which the author admitted to and apologized for since English is not his first language.
Ghostty's creator just launched a startup to replace tmux with something built for humans and AI agents together.
What the article says
- Mitchell Hashimoto, who built Vagrant and Terraform, has a new company called Superlogical.
- The pitch is that coding, running jobs, and production work all get scattered across different disconnected tools.
- Their answer is a single durable session that follows the work everywhere, whether a person or an agent is driving.
- They are starting with a terminal multiplexer, the tool that lets you organize and reconnect to running sessions from any device.
- Later they plan to add shared sessions, web and native apps, and eventually production grade tooling.
What HN is saying
- Commenters mostly debate whether the world actually needs another tmux replacement or whether current tools are fine.
- Several people compare it to similar projects already in this space, like herdr, pi web, and Orca.
- One sharp tangent compares the idea to old Microsoft technology for embedding live documents inside other documents, calling it a nice idea that is painful to build well.
- A few readers grumble that the launch title was vague and clickbaity rather than explaining what the product does.
- There is some side chatter about the company only hiring in person in LA, New York, and London, which some read as a lifestyle choice.