Someone got Command and Conquer Generals running natively on iPad using an AI coding agent, and the comments turn into a fight about how much AI actually did.
What the article says
- A developer got the 2003 real time strategy game Command and Conquer Generals running natively on Apple Silicon Macs, iPhones, and iPads, with full campaign and skirmish modes.
- It is not an emulator. The original game engine was recompiled to run on modern Apple chips, with graphics routed through several translation layers to work on Metal.
- Touch controls were built from scratch for a game designed around a mouse, including tap to select, drag to box select, and pinch to zoom.
- The project builds on an existing open source Mac and Linux port, adding the mobile version and various fixes on top.
- You still need to own the original game on Steam since no game files are included, just the code to run them.
What HN is saying
- The biggest argument is about credit. Several commenters dug through the commit history and found the AI coding tool only added a small slice of the work, since most of the porting was already done by an earlier community project.
- Some see this as a legitimate case of AI speeding up a fun, low stakes project, unlike bigger companies that quietly ship AI generated code to production.
- A recurring complaint is that AI generated writing has a distinct style, like mashing several ideas into single made up compound words, which some readers find hard to parse.
- One commenter question whether the model used was even the more expensive one claimed, joking it might have been quietly swapped for a cheaper one.
- Others just wanted to reminisce about old real time strategy games or ask which similar titles might get revived next.
A developer found evidence GPT-5.5 Codex silently caps its thinking, and it's making the model dumber.
What the article says
- A developer mined Codex usage data and found GPT-5.5 responses pile up at oddly exact reasoning lengths, especially one particular token count, far more than any other model.
- That exact stopping point keeps showing up alongside wrong answers on hard problems, suggesting the model is cutting its thinking short.
- The pattern barely existed earlier in the year and then exploded this spring, while overall thinking length quietly dropped over the same stretch.
- The author stops short of claiming OpenAI is hiding a deliberate cutback, but says the data looks exactly like a hidden budget or routing limit kicking in.
- They're asking OpenAI to check whether some scheduling or budget system is quietly truncating GPT-5.5's reasoning.
What HN is saying
- Commenters say they can reproduce it easily by giving Codex a puzzle prompt. It sometimes cuts off fast and gets the wrong answer, other times thinks much longer and nails it.
- Several people say they already sensed GPT-5.5 getting worse for weeks and quietly switched back to Claude or older versions like 5.4.
- The sharpest disagreement is over motive. Some assume a deliberate money saving nerf, while others argue it's more likely an inference or scheduling bug rather than a conspiracy.
- One reply reframes it as deja vu, noting Claude Code went through the same kind of user backlash over perceived slowdowns earlier in the year.
- A few people point out this fuels a broader shift toward open source local models, since you can't get secretly downgraded by a provider you're running yourself.
Why your iPhone remembers extra taps but your Android phone just ignores them
What the article says
- A designer compares how the iPhone and a Nothing Phone handle rotating a photo when you tap the rotate button faster than the animation can keep up.
- The iPhone buffers your taps and plays them back once the current animation finishes, so nothing gets lost.
- The Nothing Phone confirms each tap with a buzz and a sound, then simply drops any tap that arrives while it is still animating.
- The point is broader than phones. Casual interfaces eventually meet a power user, someone rotating dozens of photos in a row, and should never make people wait on an animation.
- His fix. Either buffer the input or let a new tap interrupt and speed up the animation, but never block the user.
What HN is saying
- Commenters mostly agree the buffering behavior matters and dig into the exact tap counts needed to prove it.
- One reader points to the Therac-25 radiation machine disaster as the extreme version of this problem, where operators typing faster than the software could handle bypassed safety checks.
- Someone notes Google Photos on a Pixel skips the animation entirely and rotates instantly, sidestepping the whole issue.
- A tangent about Apple's internal culture around design decisions gets a skeptical reply with a linked source.
- A commenter links to an earlier, more focused piece by the same author making a related point about interface responsiveness.
Anna's Archive will pay two hundred thousand dollars for a full dump of Google's scanned books.
What the article says
- Anna's Archive, a shadow library, is offering a two hundred thousand dollar bounty for anyone who can extract Google Books' full scanned library.
- Google only shows tiny snippets of these scans in search results, and the bounty wants the complete files, not previews.
- The post pointedly notes that a Google employee with access would find the money trivial, but would become a legend among archivists for pulling it off.
- The same bounty covers other huge book collections, including ones held by AI companies, especially if they contain rare books.
What HN is saying
- Many commenters defend the archive, saying it gave them access to books they could never legally buy or afford where they live.
- A published author pushes back, arguing piracy has visibly hurt real book sales for some writers, even if it barely dents others.
- Some readers suspect the bounty is really aimed at Google or AI company insiders willing to leak data for cash.
- People trade theories on how Anna's Archive funds six figure bounties, pointing to paid fast download memberships and AI companies buying dataset access.
- One thread turns into a warning: a linked volunteering page was spoofed with a fake CAPTCHA trying to trick people into running a password stealing script.
A jellyfish can close a wound in minutes with zero scarring, and scientists want to know why.
What the article says
- A researcher watched jellyfish cells physically walk toward each other to seal a cut, closing small wounds in minutes and bigger ones in under an hour.
- The jellyfish studied is not one animal but a life stage. Its long lived form is a colony of polyps, and the free swimming jellyfish is more like a flower the colony produces to reproduce.
- Unlike humans, the healing leaves no scar tissue. Researchers compare it to how embryos heal, which also happens without scarring.
- The hope is that understanding this process could eventually inform how human wounds are treated.
What HN is saying
- Several commenters push back on the premise, noting this reads as a press release and the research is likely driven by scientific curiosity rather than any near term medical payoff.
- One pointed out jellyfish are not single organisms but colonies of specialized parts, which may explain why healing works so differently than in humans.
- A few jokes about misreading the headline as a literal cream you rub on wounds, and confusion over the title given jellyfish are best known for stinging, not healing.
A researcher tricked YouTube's AI assistant into leaking private video titles, and Google says it's not a bug.
What the article says
- YouTube Studio's Ask Studio assistant reads comments and summarizes them for creators.
- A researcher hid instructions inside a normal looking comment, and the AI followed them instead of just describing them.
- He got the AI to insert a private video title into a link that pinged his own server when the creator clicked it.
- Comments can be edited after posting without notifying the creator, so the trap can hide in plain sight.
- Google rejected the report twice, calling it social engineering rather than a security bug.
What HN is saying
- Many commenters say prompt injection is basically unfixable with today's AI models, so treating it as a normal bug is the wrong framing.
- A former Google employee argues internal promotion incentives explain why engineers let bugs like this slide.
- Some pushed back that this really is just phishing, since the creator still has to click a suspicious link.
- Others tested the exploit themselves and got mixed results, showing how unreliable these attacks can be at scale.
- Several people flagged the deeper worry: if this works, an attacker could manipulate what the AI tells creators about anything, not just leak titles.
A 2019 deep dive into htop that finally explains what every number on that screen actually means.
What the article says
- The author got tired of misreading htop and wrote a plain guide to every field on the screen.
- It covers uptime and load average, showing they come straight from files under proc rather than magic.
- It walks through memory numbers, explaining why virtual memory looks scary but resident memory is what matters.
- It explains process states, priority, and CPU columns so the whole dashboard finally makes sense.
- The piece leans on simple tools like strace to show exactly where htop gets its data.
What HN is saying
- Plenty of readers use this as their excuse to recommend btop, a flashier alternative with power and disk graphs.
- Others push back that btop is missing features like zram stats or a ports column, so it is not a strict upgrade.
- A long thread argues about which memory number to trust, with several people saying resident set size still overstates real usage.
- Someone points out that even Windows Task Manager uses a different definition of memory than people assume.
- A few commenters just say the article cleared up years of confusion about what htop's numbers really mean.
Newer Claude models are getting worse at correctly calling tools that aren't shaped like Claude Code's own.
What the article says
- A developer noticed Opus and Sonnet sometimes invent extra, made up fields when calling his editing tool, even though the actual edit they wanted was correct.
- Older Claude models did not have this problem. Only the newest ones do, which is backwards from what you would expect.
- His theory is that these models were trained heavily around Anthropic's own Claude Code tool shapes, which quietly tolerate and repair sloppy calls.
- That training may make the models better at Claude Code's own tools but worse and more confused when a different tool expects a different shape.
- He argues this could lock other developers into copying Claude Code's exact tool format just to get reliable behavior.
What HN is saying
- Several commenters agreed the fix is simple: give the model clear error messages so it can correct itself on the next try.
- Others pushed back that every failed call costs extra money and time, even if it eventually self corrects.
- One thread argued this reflects a broader trend where the dominant coding tool's habits become a de facto standard that other tools must imitate.
- Some suggested forcing valid output at the technical level, called constrained decoding, though others said this can make models perform worse overall.
- A few saw this as part of a bigger pattern of AI providers quietly shaping or downgrading model behavior in ways developers cannot fully see or control.
A user swears Claude Code started asking about someone else's Minecraft build mid session. Leak or ghost in the machine?
What the article says
- A developer filed a bug saying their Claude Code agent suddenly asked what bricks to use for a Minecraft temple, something they never mentioned.
- They were on an enterprise workspace that is supposed to keep chats private and isolated from other customers.
- They wondered whether the stray prompt came from a different company account or even a free consumer account.
- They also noted a separate, unrelated mixup in their own setup where the agent worked in the wrong folder, which they say is not the same issue.
- Anthropic responded that they believe it is a hallucination but are still looking into it.
What HN is saying
- Commenters split between two camps. Many think a long chat session just went off the rails, since AI models can drift into odd tangents the longer a conversation runs.
- Others think this looks like a known failure mode where shared infrastructure mixes up responses between different users, sometimes called request smuggling.
- One commenter says a similar swap once happened at a major AI company due to a bug in how their servers handled connections, and got a public apology.
- Someone from the Claude Code team showed up in the thread and said they believe it is a hallucination but are investigating further.
- Several people report a similar strange, unrelated answer problem happening in Gemini recently, which fed suspicion something bigger might be going on across AI providers.
Zig ripped its package manager out of the compiler, and even the maintainers admit the old design was a mess.
What the article says
- Zig moved package fetching, networking, and file parsing out of the core compiler and into the build system itself.
- That code now ships as source rather than being baked into the compiler binary, so people can patch it without a full rebuild.
- The compiler binary actually shrank a bit as a result.
- It also fixes an awkward process handoff during long running watch builds, so the build system no longer has to restart itself on every config change.
- The end goal is a cleaner path toward a proper build server that tools like the Zig language server can talk to.
What HN is saying
- Commenters are impressed by how deliberate and calm Zig's development process feels compared to most languages.
- One thread turns into a broader defense of human made programming languages, arguing an LLM could never have designed something like Zig with real taste and tradeoffs.
- A few people note this is the change that forced Zig to drop its handy C import feature, and call that a sad but necessary tradeoff.
- Someone jokes that language creators keep separating things that probably should not have been combined in the first place.
- Others point out that no widely loved cross language build system exists yet, so every language ends up rolling its own.