Top Stories

An OpenAI model has disproved a central conjecture in discrete geometry

1205 points · openai.com

OpenAI is claiming a serious mathematical milestone: one of its models produced a counterexample that disproves a long-standing conjecture in discrete geometry. This is the kind of result that lands somewhere between “neat AI demo” and “actual contribution to mathematics” — and HN is, predictably, dissecting which one it really is. The big question for readers is whether the model genuinely reasoned its way to a novel construction or whether it stumbled into a known-but-obscure result that wasn’t in the original conjecture’s evidence base. Either way, it’s another data point in the steadily growing pile suggesting that frontier models are starting to be useful research collaborators rather than just chatty interns.

GitHub confirms breach of 3,800 repos via malicious VSCode extension

875 points · bleepingcomputer.com

A malicious VSCode extension harvested credentials from developers and used them to access nearly 4,000 GitHub repositories — the kind of supply-chain horror story that gets bookmarked, screenshotted, and forwarded to every security team. The discussion has the usual two camps: the “this is why you audit your dependencies” crowd and the more reasonable “the marketplace UX makes auditing effectively impossible for working devs” camp. Practical takeaway for anyone reading: this is a good week to review which VSCode extensions you have installed and rotate any tokens that ever touched a machine running an unfamiliar one.

How fast is N tokens per second really?

424 points · mikeveerman.github.io

A delightfully simple interactive page that maps LLM token throughput onto reading speed and word-per-minute comparisons, so you can finally answer “is 80 tok/s fast?” in human terms. The visualization is what makes it land — you watch text stream at the speed your model would produce it, and suddenly the gap between a local Llama on a laptop and a hosted frontier model becomes viscerally obvious. Useful as a sanity check next time someone tries to sell you on a small open model “matching” a frontier one.

Flipper One Tech Specs

395 points · docs.flipper.net

The Flipper One is here — successor to the cult-favorite Flipper Zero — and the published specs confirm a significant jump: a more capable SoC, an integrated SDR, and improved support for sub-GHz and NFC work. The comment thread is the usual lively mix of pentesters salivating, hobbyists complaining about price, and at least one person worrying about regulators noticing. For anyone in security research or hardware hacking, this is the device to know about for the next year or two.

Saying goodbye to asm.js

386 points · spidermonkey.dev

SpiderMonkey is retiring its dedicated asm.js optimization path now that WebAssembly has fully eclipsed it. asm.js was a fascinating bridge technology — a strict, typed subset of JavaScript that could be compiled almost directly to machine code — and a reminder of how the web platform sometimes evolves through clever subsets rather than entirely new specs. Engineering nostalgia aside, the post is also a clean read on why “remove a thing” is often more painful than “ship a thing,” with notes on backward compatibility and the long tail of code still depending on it.

Google’s AI is being manipulated. The search giant is quietly fighting back

318 points · bbc.com

A BBC Future piece on the ongoing arms race between SEO operators (now “AEO” — answer engine optimization) and Google’s AI Mode. The interesting stuff isn’t the headline framing but the specifics: prompt-injection-style attacks embedded in content, structured-data manipulation, and entire content farms now tuned for what gets quoted by AI summaries rather than what ranks in classic search. If you build for the web, this is the latest chapter of a story you need to be paying attention to.

Show HN: I reverse engineered Apple’s video wallpapers

303 points · github.com

A nicely-written reverse engineering writeup that dissects how Apple’s aerial-style video wallpapers are encoded, streamed, and rendered on macOS — complete with an open-source tool that lets you play them on non-Apple hardware. The kind of weekend project that makes HN happy: a single curious person, a closed system, and a clean walkthrough of how they pulled it apart. Bonus: the comments include a small library of similar “Apple internals” writeups people have collected over the years.

OpenAI Is Preparing to File for an IPO Soon

106 points · wsj.com

The Wall Street Journal reports that OpenAI is moving toward a public offering “very soon,” which would make it one of the largest tech IPOs in years and would force the company into a level of financial disclosure that frankly nobody has seen before. Comments are split between “this was always the plan post-restructure” and “good luck explaining the unit economics to public-market investors.” Either way, an OpenAI S-1 will be one of the most-read documents in tech this decade.

Haskell Foundation 2026 Update

125 points · discourse.haskell.org

Annual state-of-the-ecosystem from the Haskell Foundation: progress on tooling (GHC, HLS, Cabal), funded community work, and where the language is heading. Haskell isn’t going mainstream anytime soon, but it’s quietly maintained a remarkably stable identity as the language a generation of programmers reach for when they want to think harder about types. The thread is worth a read if you write Haskell professionally — or if you just want a window into how a small, fiercely opinionated language community keeps itself alive.

Google officially announces that ads will be included in AI Mode search results

72 points · blog.google

Google has formally confirmed what everyone assumed was coming: ads inside AI Mode answers. The framing is “shopping intent” and “useful results,” but the practical reality is that the AI answer panel — the thing replacing the classic blue links for many queries — now monetizes the same way the blue links did. The interesting question is whether AI-mediated ads end up more or less obtrusive than the pre-AI experience, and whether the inevitable optimization games (see the BBC story above) accelerate as a result.