Turn spare GPUs scattered across your house into one shared brain for running huge AI models.
What the article says
- Mesh LLM pools spare GPUs from any machines you own into one shared pool, so you can run big models without buying more hardware, and any tool built for OpenAI's API can plug straight in.
- A request can run locally, get handed to a peer machine that already has the model loaded, or get split into pieces that run across several machines at once.
- The splitting trick divides a model by layers, so several weaker machines can team up to run something none of them could handle alone.
- There's no central server. Each machine has its own identity and connects directly to peers, with backup relay points for ones that can't reach each other.
What HN is saying
- The big question was speed. One user found a buried number: a giant model split across two machines ran at sixteen tokens per second, usable but not quite live chat speed.
- One of the project's builders jumped in, saying their custom splitting method beats the older approach other tools use, backed by his own home lab tests.
- Someone asked if traffic between machines is encrypted, since strangers can join. A contributor confirmed it is, tied to each machine's own key rather than added on later.
- A few people compared it to existing projects, but others said those solve a different problem, putting one model online rather than splitting one across machines.
A website that rewrites itself live: edit Scheme code in your browser and it recompiles to WebAssembly in about fifteen milliseconds.
What the article says
- The homepage is the demo. Its own source is Scheme code, compiled to WebAssembly right in your browser and rendered live.
- There is no server doing the compiling. The whole compiler ships inside the page and gets cached after your first visit.
- Edit the code, press Run, and the page rebuilds itself in about fifteen milliseconds.
- The goal is writing web pages and talking to the browser using Scheme instead of JavaScript.
What HN is saying
- Commenters note this covers ground already handled by Hoot, a mature Scheme to WebAssembly project from the Spritely Institute, and that this one looks like a two day old side project by comparison.
- One thread argues JavaScript's real weakness in the AI coding era is that it is hard to verify and fails silently, so compiling from a stricter language could make AI generated web code more trustworthy.
- A skeptic pushes back hard, saying the project's own explanation for picking Scheme reads like AI generated reasoning that does not actually hold together, and asks why Scheme over any other language that targets WebAssembly.
- A few people hit an editor bug where typing lands a line above the cursor, though the code still compiles once you work around it.
A one person JS runtime is hyped, and accused of secretly reusing an older engine.
What the article says
- The page itself is a JS-heavy demo, so this is inferred from the title, description and comments.
- Ant is a full JavaScript stack built by one developer: a new engine, a package manager, a hosting platform and a desktop app framework like Electron.
- It packs into a single nine megabyte binary, works with existing npm packages and TypeScript, and includes a sandbox for running untrusted code plus support for WebAssembly.
- The pitch is speed close to Google's V8 engine with a much smaller footprint than Node.
What HN is saying
- The biggest fight is about honesty. Commenters traced the engine back to an older open source project called Elk, contradicting the from scratch claim, though the author says it has since been rewritten.
- Several people are annoyed the project is named Ant, since Apache Ant and the popular Ant Design library already use that name in the JavaScript world.
- A speed comparison site suggests Ant lags well behind the near V8 performance being claimed, but the author says the engine was rewritten since February and promises real benchmarks soon.
- Some commenters like the built in sandboxing for safely running untrusted code, while others doubt a small binary matters much when most projects already carry a bloated dependency folder.
A Raspberry Pi chip engineer designed a working handheld games console from scratch, in his spare time.
What the article says
- This isn't an emulator. The author designed the actual chip from scratch, including the processor, the graphics pipeline, and the circuit board that holds it all.
- He describes it as a Game Boy Advance from an alternate history, one where this kind of open chip design existed decades earlier.
- It runs on a small, cheap chip, and the processor design has been checked against formal verification tools that prove it behaves correctly.
- Every file, from the chip logic to the circuit board layout, is open source and free for anyone to build or modify.
What HN is saying
- Commenters quickly placed the author: a Raspberry Pi engineer who also designed the processor core inside Raspberry Pi's newest chip, built in his spare time.
- He previously found a clever way to output video from a cheap Raspberry Pi chip that was never designed to handle it.
- One person assumed a hardware standard used in the design was locked to a single company. Others corrected that it has actually been open for years.
- A few readers just love this whole genre of reimagined retro hardware, and one thread jokes that the project page is refreshingly free of emoji.
A guy fought years of laptop dock wake up failures. The fix wasn't the dock at all.
What the article says
- A developer describes years of docked laptop hell. He'd arrive at his desk, tap the keyboard, and the monitor just wouldn't wake up.
- He upgraded through several top tier Thunderbolt docks hoping the spec's promised fix would finally kick in. It never fully did.
- Then he swapped his aging monitor for a new one, mostly chasing a faster refresh rate, and the wake up problem vanished completely.
- He still isn't sure why. It could be old monitor firmware or something else entirely, but the fix has held ever since.
What HN is saying
- Commenters swap their own dock horror stories across brands, confirming this is a common and stubborn problem, not one bad unit.
- Several agree the monitor matters more than people assume. Older or cheap displays often have outdated chipsets that mess with wake signals.
- Others argue it's really a chain problem. The laptop, dock, cable, and monitor all have to cooperate, so blaming the dock alone misses the point.
- One reader said the post taught them nothing, which pulled a longer reply explaining the deeper compatibility issue at play.
Nvidia invests billions in cloud startups that turn around and spend it all on Nvidia chips.
What the article says
- CoreWeave and Nebius are cloud upstarts with fast access to Nvidia's newest chips. They rent that power to giants like Microsoft and Meta, who'd rather pay a subscription than build their own data centers.
- Nvidia funds both companies with big equity stakes and promises to buy back unsold capacity. Those same companies then spend the money buying more Nvidia hardware, the loop critics call circular financing.
- The catch is money. CoreWeave is burning cash and piling on debt, while Nebius has healthier finances but still needs billions more to keep building.
- The bet is that both eventually turn a profit. Right now they're growing faster than they can afford.
What HN is saying
- Biggest pushback: Nvidia's actual cash stake is a small slice of CoreWeave's total spending, so calling it circular is overblown, most of the money comes from elsewhere.
- Others say the real risk isn't the equity, it's Nvidia's promise to buy back any unsold computing capacity for years. That guarantee removes CoreWeave's downside and could tempt it to overbuild.
- One sharp take argues the circular financing debate is a distraction. The real question is whether these data centers ever earn back what they cost per unit of AI work.
- A darker view compares the whole tangle of AI debt and IOUs to the run up to the 2007 financial crisis, though others weren't convinced the comparison holds.
A researcher caught xAI's coding tool quietly uploading your entire codebase, including secrets, to the cloud.
What the article says
- A researcher intercepted network traffic from xAI's Grok Build coding tool and found it uploads your whole codebase to Google's cloud, not just the files it actually opens.
- It grabbed a secrets file sitting in the project and sent the passwords straight through with no filtering.
- Turning off the improve the model privacy setting made no difference. The upload still happened.
- The behavior isn't mentioned in xAI's setup instructions and is switched on by default.
- As proof, the researcher planted a file the tool was told never to open, and its contents still showed up in the uploaded data.
What HN is saying
- Commenters were most rattled that a file nobody asked the tool to touch still got swept up and uploaded.
- Many now run coding tools inside a locked down sandbox that blocks access to anything outside the project folder.
- Some pushed back, arguing every cloud coding assistant, including Claude and Codex, sends your code to a server, and this is just a more blatant version of the same thing.
- Others found it darkly funny given Musk's history, joking that xAI now effectively has a copy of Tesla's codebase.
- A few said Chinese coding assistants like Deepseek and Qwen are noticeably more careful about what they read and send home.
Colleges must now prove their degrees pay off or lose access to federal student loans.
What the article says
- A new federal rule cuts off student loans for degree programs whose graduates don't earn more than people who skipped college.
- Roughly eight hundred thousand students are in programs likely to fail the test, most of them at for profit schools.
- Even prestigious arts programs like Juilliard could lose funding, since music, theater and design degrees often pay poorly early on.
- Advocates for the arts say earnings miss the point for creative careers, where pay starts low but satisfaction and stability grow over time.
- The rule phases in over the next few years, with real penalties for failing programs starting later this decade.
What HN is saying
- Most commenters welcome the change, calling it overdue accountability for colleges that padded costs without improving job prospects.
- A side debate breaks out over why student loans can't be wiped out in bankruptcy like other debt, with some suggesting a waiting period instead.
- One sharp worry is that schools will dodge the rule by inventing oddly named majors every few years to reset the clock.
- Another commenter argues trade schools deserve to grant real degrees too, since a plumbing certification can be worth more than an obscure academic major.
It's not the prices making Americans furious, it's endless scams, junk products, and vanishing consumer protections.
What the article says
- The Guardian's page was blocked from extraction, so this is pieced together from the headline, comments, and quoted excerpts.
- A consumer advocacy group calls it a war on consumers, pointing to a wave of hidden fees and deceptive tricks in everyday purchases.
- A Wharton marketing professor says people have more choices than ever, yet service keeps getting worse and companies use customer data against them.
- One example cited is Toyota's lending arm, fined for tricking buyers into unwanted insurance and stonewalling refund requests for months.
- The piece argues the Trump administration is defanging the watchdog agency that handles cases like that, just as complaints are climbing.
What HN is saying
- Most commenters agree products themselves have gotten worse: cheaply made, loaded with subscriptions and dark patterns, and warranty support designed to wear you down until you give up.
- The sharpest fight is over blame. Some point to corporate greed and gutted regulators, others argue Americans keep voting for the deregulation causing this and are getting what they asked for.
- A few push back entirely, calling the rage itself overblown, a habit fed by outrage media rather than a reasonable response to real problems.
- One commenter's rant about a jammed self checkout and shrinking, pricier deodorant became the thread's favorite stand in for daily consumer misery.
A small study links long COVID to real nerve damage in the stomach lining.
What the article says
- Researchers biopsied stomach lining in long COVID patients and found the nerve fibers there thinned dramatically compared to healthy people.
- They also tracked heart rhythm overnight to gauge how well the vagus nerve, which controls digestion, heart rate and other organs, is functioning.
- The vagus nerve carries signals between gut and brain, so damage there could explain why so many people have lingering stomach problems.
- The study is small, so the authors call it early evidence rather than proof and say larger studies are needed to confirm it.
What HN is saying
- Several commenters flag that the study is tiny, so they read cautious words like may in the headline as close to meaningless.
- One person compares long COVID to AIDS before its cause was found, one trigger and a spread of oddly unrelated symptoms nobody can yet explain.
- Someone asks if there is a test for long COVID. Replies say no single test exists yet, though researchers are validating candidate biomarkers.
- A recurring gripe is that other post viral conditions have gone unnamed and understudied for decades while long COVID draws outsized attention.
- One reader who has had stomach trouble since getting infected years ago asks what symptoms this kind of nerve damage would actually cause.