A mathematician says OpenAI raced to copy his unpublished proof, then threatened him to keep quiet about it.
What the article says
- Two researchers announced blowup results for several fluid equations, built with heavy help from language models, and credited the underlying idea to other mathematicians who came before them.
- One of the pair says that after rumors spread about their progress, OpenAI told him its own model had cracked a related million dollar problem using the same approach, just days later.
- He says OpenAI wanted his research partner left off the paper simply because that partner works at a rival lab, and would not say whether his private chat logs had trained their model.
- When he said he would go public, he says an OpenAI researcher warned him it could ruin his career.
What HN is saying
- Most people find OpenAI's mealy mouthed denial about training on his private chats more damning than reassuring, since a flat no would have been easy to say.
- Some read this as plain industrial espionage, someone hearing a promising lead and throwing huge computing power at it before the original researchers could finish.
- Others push back, pointing out that employees rarely have access to that kind of user data and the whole training claim is still just speculation.
- Several note both sides leaned on language models to extend an older idea that wasn't originally theirs, which muddies who actually deserves credit here.
- The alleged line about ruining his career gets singled out repeatedly as the part people most want OpenAI to explain.
OpenAI says it cracked a century old fluid dynamics puzzle, then a mathematician accused it of lifting his unpublished work.
What the article says
- OpenAI claims an unreleased internal model solved the Navier Stokes problem, a famous open question about whether smooth fluid flow can suddenly blow up into a singularity.
- The proof shows a spinning vortex of fluid can spiral inward and speed up without limit in finite time, while staying within the laws of physics.
- Thousands of coordinating agents worked the problem for roughly four days, first cracking a related simpler case before turning to the harder one.
- The company says it did not access anyone's private work, though it admits it cannot fully rule out that anonymized user data shaped the model.
What HN is saying
- The loudest thread is a dispute over credit. A rival researcher says he had been quietly working the same problem for a year and felt OpenAI rushed out a competing solution after hearing rumors of his progress.
- Screenshots shared in the comments show tense messages where an OpenAI researcher allegedly pressured him about going public, which many read as bad behavior regardless of who solved what first.
- Commenters split on the science itself. Some call the result a genuine milestone on par with famous AI breakthroughs in games, others say the two proofs differ enough that comparisons are unfair.
- Terence Tao's warning gets a lot of traction: if labs race to bulldoze any research rumor with massive compute, mathematicians may stop sharing ideas early, which would hurt open science.
- A side argument breaks out over whether training on customer chat data, even anonymized, crosses an ethical line the company hasn't been straight about.
Europe's AI hope just raised a huge round, and the comments argue about whether that even matters anymore
What the article says
- Mistral raised three billion euros, the biggest funding round any European tech company has ever pulled off.
- Samsung led the round, alongside a European public investment fund and existing backers.
- The pitch is not the smartest model but a sovereign one, where the data, the model, and the servers all stay under a customer's own control.
- It already runs in twenty countries for big clients like Airbus and HSBC.
- The company frames this as a different question than the one other labs are chasing, control instead of raw power.
What HN is saying
- Most commenters like the idea of a European lab existing but think the actual models are mediocre next to the top Chinese and American ones.
- Several people say Mistral's real strength is boring and useful stuff like OCR and voice models, not chasing the frontier.
- A sharp pushback point is that the money mostly comes from non European investors, which undercuts the sovereignty story.
- One commenter calls the funding a rounding error next to what OpenAI has raised, and doubts it changes much.
- A few people argue this doesn't matter because cheap fine tuning of open Chinese models already beats needing your own frontier lab.
LibreOffice just had its best week ever, right after telling users it refuses to add AI features.
What the article says
- LibreOffice's newest version had the biggest first week of downloads in its history, over a million installs.
- The foundation behind it says the timing isn't a coincidence. It published a piece arguing that skipping AI is itself a selling point.
- Their reasoning: any AI has to run under the user's control, never send data off the computer, skip telemetry, avoid vendor lock in, and stay fully optional.
- They're not against AI outright. They just think it isn't ready to meet those rules, so for now they point people to community plugins instead.
- The writer suggests rivals use AI features partly to justify subscription price hikes, which is the opposite bet LibreOffice is making.
What HN is saying
- Biggest pushback is that download counts always trend upward, so calling this a record is a bit of a cheap headline trick.
- Several people point out a more likely cause. Coding tools like Codex and Claude now bundle or call LibreOffice to edit documents, which could be quietly inflating the numbers.
- Others credit people ditching Windows for Linux, since installing Linux often means installing LibreOffice too.
- A few say they don't want built in AI cluttering the app, but still end up asking a separate AI chatbot how to do basic things like changing paper size.
- One person actually built a small analysis to test the causal claim and concluded it's speculative, not proven.
Jellyfin drops a big free update, and the comments turn into a real Plex versus Jellyfin referendum.
What the article says
- Jellyfin just released version 12, and they renamed it from 10.12 to signal this is a genuinely major update.
- The database got rebuilt so playlists, collections, and your watch history load without freezing on big libraries.
- Books and comics are now properly supported. Covers, authors, and page counts get pulled in automatically.
- You can now keep multiple versions of a single episode, like a broadcast cut and an extended cut, the way movies already worked.
- Backups are safer, subtitle settings moved to a per library setting, and there is a stack of small transcoding and security fixes.
What HN is saying
- Plenty of longtime Plex users say this release is what finally convinces them to switch, calling it a solid escape hatch if Plex turns hostile.
- Subtitles are the most common complaint. Several people describe fiddly setups that only work after routing through separate tools like Bazarr.
- Upgraders mostly report smooth, fast migrations, though a couple hit failed database migrations and had to start their library fresh.
- A running gripe is that Jellyfin still leans on a single file based database, which struggles under big libraries and blocks running it across multiple machines.
- One sharp aside, someone points out the release notes read like they were written with heavy AI help, and the team's own contributor policy already bans copy pasting AI answers into pull requests without understanding them.