HN reading notes — 8 October 2026: AI mathematics, durable software, and small delightful tools
AI mathematics took up most of my HN reading today. I kept coming back to a question: what happens after a problem is solved, when the answer arrives faster than anyone can understand it?
Proofs, errors, and what comes next
OpenAI’s mathematics update describes a batch of results and work on formalising proofs. Then there’s the revision history: a sign error led to three dependent manuscripts being withdrawn. Those two pages belong together. A checked proof is valuable, but checking part of a project doesn’t establish the rest of it.
Terence Tao’s Math 2.0 thread gave me more to think about than the announcements. A first proof usually sets other work in motion: people explain it, simplify it, teach it, and find new questions. If solving the problem becomes cheap, those contributions still need time and recognition. Scott Aaronson’s post and the Association for Humanistic Mathematics statement show how much disagreement there is about where this goes. I’d rather have a proof I can learn from than another entry on a leaderboard.
The software I want to keep using
Dan Cohen’s essay on durable software was a welcome change of pace. His account of Zotero’s development is about years of requirements and collaboration. Getting something to run is one thing; making it useful enough that people keep relying on it takes longer.
Hundred Rabbits is always a good rabbit hole: Uxn, small tools, preservation, and experiments with older languages. I also liked this history of DVD menus. There’s a lot of imagination in those awkward little interfaces. Constraints aren’t always something to escape.
A few programming bookmarks
Push Ifs Up and Fors Down is one I’ll return to. Move decisions toward the boundaries and think about collections as collections. Useful advice for organising code; I’d still want measurements before calling it an optimisation.
The floating-point essay makes a distinction that’s easy to blur: detecting a non-finite result and knowing whether a finite result is accurate are different jobs.
GitHub’s pull-request controls let maintainers restrict who opens PRs, or disable them altogether. I can see why some projects would want that. Docker Agent also caught my eye—YAML configuration, multiple model providers, and tools through MCP. I’d want a close look at the permissions and telemetry before putting it to work.
Away from the announcements
MIT’s obituary for Margaret Hamilton was the day’s most memorable read. Apollo software puts a lot of today’s talk about reliability into perspective.
The search for Salvage was another favourite: a music mystery, an unhelpful name, and eventually the right people to ask. I won’t spoil the route to the answer.
For something lighter, there’s Bigwords and a collection of ASCII animations. A browser can still be a place for a clever little toy. After all the arguments about the future of research, that was a pleasant place to finish.