A developer blamed Claude for cloning a rival app down to its name and a bug, and nobody's buying it.
What the article says
- A developer launched a stargazing web app called Dark Hours, built with Claude.
- Another developer pointed out it was nearly identical to his own open source project of the same name, bug for bug.
- The blogger redirected his domain to the original creator, scrapped his iOS plans, and posted an apology.
- He says he never saw the other project before, and blames carelessness in trusting AI output without checking it.
What HN is saying
- Commenters largely think the apology is a dodge and that Claude would not clone a project's exact name and bugs unprompted.
- Many tie this to a bigger story where the developer had earlier told John Gruber his app was mistaken for astrology software, a claim that started unraveling once the copying came out.
- One person coined it a limited hangout, admitting a small piece of a scandal to bury the worse parts.
- A recurring complaint is that blaming the AI is becoming a universal excuse for bad behavior, and that people should own what they publish regardless of how it was made.
- A few commenters push back gently, noting the developer's underlying idea for how RSS readers should work is genuinely interesting even if this episode damaged his credibility.
A guy learned chip manufacturing by having an LLM build him a tiny 3D factory game instead of an essay.
What the article says
- An engineer researching AI data center bottlenecks realized he didn't understand how chips get made, so he had an LLM teach him by building something playable instead of writing him an explanation.
- His method has four steps. Ask a model to build a knowledge base on the topic, have it fact check itself, turn that into a small low poly animated simulation like an old theme park game, then publish it as a website.
- He built one for chip manufacturing called ChipTycoon, where you follow a cart of sand turning into a finished chip through each stage of the process.
- He suggests going further by turning the low poly shapes into more realistic 3D objects and adding quiz style challenges to make the learning stick.
What HN is saying
- The loudest pushback targets his claim that the result is completely accurate and free of hallucinations. Commenters ask how he could possibly know that, since the checking step is just the same AI reviewing its own work.
- Several people say they are tired of dense, over-explained LLM writing and have drifted back to books, using the model only to answer specific questions instead of generate whole lessons.
- One reader argues the real risk is that you cannot catch a hallucination in a subject you are still learning, since you have no prior knowledge to catch it against.
- A recurring theme is that LLMs work best as a guide or study partner rather than a replacement for doing the hard, boring work yourself. One teacher found students got better results asking the model to point them to source material rather than just answer questions directly.
- A few commenters enjoyed the playful idea itself, with one calling the pivot to a theme park style game genuinely delightful.
Fastmail now lets you keep your inbox in Amsterdam, but reads the fine print itself.
What the article says
- Fastmail built its own servers in Amsterdam so customers can choose to make Europe the primary home for their email instead of the US.
- The backup copy still lives in the US for now, since Amsterdam is the company's only European location so far.
- Switching regions is a simple settings toggle and your account keeps working during the move, though moving into Europe takes longer because everything has to sync across the ocean.
- Fastmail is upfront that this is not a guarantee your data never touches the US. It is an Australian company and says it will comply with lawful government requests no matter where your data sits.
What HN is saying
- The biggest pushback is that the US CLOUD Act can still compel Fastmail to hand over data as long as it does any business in the US, so a European server does not really solve the legal problem.
- Several commenters point out Fastmail is actually Australian, not American, which brings its own surveillance law into play alongside any US exposure.
- Plenty of people still welcome the move, calling it a genuine first step and praising Fastmail for being honest about the limits instead of overselling privacy.
- Others suggest fully European email providers like Mailbox.org, Posteo, or Tuta, though some warn that a couple of those drop standard IMAP support, which locks you into their own apps.
A guy turned his old smartphone into a real server, and the mess of a title took over the comments.
What the article says
- Instead of paying for a cloud server, the author dug out an old CMF Phone and turned it into his home server.
- First he tried wiping Android for Linux, and it nearly bricked the phone since drivers for Wi-Fi and other hardware were missing.
- So he kept Android and ran everything through Termux, a Linux environment layered on top, with a supervisor keeping services alive through reboots.
- He rooted the phone so his heaviest app, a remote browser, could talk to the hardware directly instead of through a slow compatibility layer.
- Traffic reaches the phone through Cloudflare and Tailscale tunnels, so it keeps working even if he unplugs it and takes it elsewhere.
What HN is saying
- Most of the thread argued about the title itself, debating whether it should have said the phone is now a server rather than the reverse.
- The author, a native Turkish speaker, said the phrasing came from translating sentence structure between the two languages.
- Several commenters warned that phones without unlockable bootloaders cannot bind to network ports, which blocks this trick entirely.
- People traded notes on cheaper alternatives, like old office desktops or fanless mini routers doubling as home servers.
- One reader mentioned doing the exact same project the week before, saying an AI coding assistant got it working in about two hours.
DeepMind's new weather AI gives forecasters a full extra day of warning before a cyclone hits, and it's free.
What the article says
- DeepMind built a model called WeatherNext that predicts a cyclone's path, strength, and wind pattern more accurately than existing systems.
- Its three day forecasts are as good as what older models could only manage two days out, which researchers compare to a decade of normal progress.
- It already proved itself this year, helping forecasters warn Jamaica ahead of Hurricane Melissa's sudden intensification.
- The model can produce a full fifteen day forecast in under a minute, and DeepMind is now releasing it publicly for anyone to use.
- Even the team admits they don't fully understand why the model is this accurate at this level of detail.
What HN is saying
- Commenters are excited that DeepMind is doing serious science instead of chasing another coding assistant.
- Several point out this leans on graph neural networks, a less hyped architecture than the LLMs everyone talks about.
- A sharp pushback: these models still depend entirely on decades of government weather data collected by agencies like NOAA, so claims that AI beats the government miss the point.
- Another thread worries the model is confident without being able to explain its own uncertainty, which matters when an evacuation order is on the line.
- People also swapped tips on other storm tracking sites, and joked this breakthrough might be why Google's AI chief got promoted.