Activists rigged a London bus ad so Kylie Jenner's Meta glasses photo flips into an X-ray skeleton watching you.
What the article says
- A UK group called Everyone Hates Elon put up a lenticular bus stop ad near Meta's London office.
- Depending on where you stand, it shows Kylie Jenner promoting Meta's camera glasses, or a ghostly X-ray version with the line We're always watching.
- The design nods to the film They Live, where special sunglasses expose hidden control by aliens.
- The group says the glasses enable secret filming, and points to their popularity with men who record women in public without consent.
- It also cites a Harvard project that used the glasses plus facial recognition to pull strangers' names and addresses from public records.
- Meta says a coming update will disable the camera if its recording light is covered.
What HN is saying
- Commenters mostly agree the camera, not the audio or AI features, is what makes people uneasy, and think a camera free version would sell better.
- Several people admired the poster's clever printing trick and guessed it cost only a few hundred euros to make.
- The sharpest disagreement was over blame. Some called it inevitable that wearable cameras spread and pointless to fight, others said normalizing constant recording chips away at real privacy.
- One commenter shared a personal counterpoint, a sweet video of their toddler running toward them that the glasses captured hands free.
- A side argument broke out over whether pointing at UK police surveillance is a fair comparison or just a distraction from Meta's own behavior.
Thinking Machines just released a serious open weights model built for fine tuning, not for topping leaderboards.
What the article says
- Thinking Machines Lab released Inkling, its first open weights model, free for anyone to download and customize.
- It's built as a broad, flexible base rather than a pure benchmark chaser, with real strength across text, images and audio.
- The model can dial how much it thinks before answering, so developers can trade cost against accuracy.
- It's meant to be fine tuned through the company's Tinker platform, and in one demo the model fine tuned itself.
- A smaller companion model is coming soon that matches or beats the full size version on many benchmarks while running far cheaper.
What HN is saying
- Commenters were excited mainly because it's a genuinely strong American open model, something rare next to Chinese labs like DeepSeek and Kimi.
- Plenty of side by side comparisons to Kimi and GLM, with mixed opinions on whether it actually wins on real tasks or just benchmarks.
- One early tester said it quietly beat every other model besides Anthropic's on obscure trivia questions, calling it better than its scores suggest.
- Others debated the business logic of giving away a frontier model for free, landing on fine tuning services as the real product.
- A few pushed back on the American versus Chinese framing entirely, arguing quality should matter more than a lab's home country.
SQLite's cozy defaults are quietly dangerous. One blogger wants a single switch to fix them all.
What the article says
- SQLite ignores foreign keys by default, so deleting a user can silently reassign their old ID to someone new, handing that new person the old data.
- Columns do not actually enforce their declared type, so a number column can happily store a sentence unless you turn on strict mode table by table.
- Concurrent writes fail instantly instead of waiting their turn, and the faster write ahead log mode is off unless you switch it on yourself.
- The fix is not changing the defaults, since that breaks old software. Instead, borrow Rust's idea of editions, one pragma that opts a database into a bundle of sane modern settings while old code keeps working untouched.
What HN is saying
- Most commenters agree these really are the settings anyone serious about SQLite already sets by hand, so making them opt in as a bundle sounds sensible.
- The author showed up in the thread and got pointed toward the official SQLite forum, since the developers are said to read it closely and respond well to proposals.
- A pushback thread argues SQLite is a lightweight tool for local storage, not a full database, and people should just reach for Postgres if they want strict guarantees. Others countered that a local database should not have to sacrifice basic type safety.
- One practical snag people raised is that ORMs like Drizzle still cannot even declare strict tables, so better defaults from SQLite itself would not fully help until the tooling around it catches up.
xAI open sourced its coding agent days after getting caught quietly uploading users' whole codebases to the cloud.
What the article says
- Grok Build is SpaceXAI's terminal based coding agent, in the same category as Claude Code, that reads your codebase and can edit files or run shell commands.
- The open source release lands right after reports that it was silently uploading entire user repositories, including private env files, to xAI's servers.
- xAI says the uploads have now stopped and it will delete whatever was collected, but it has not offered any outside proof of that deletion.
- The codebase itself is enormous, over a million lines of Rust with many vendored dependencies, which some readers point to as a sign of sloppy, AI generated code.
What HN is saying
- Most commenters read the open sourcing as damage control for a battered reputation rather than a genuine embrace of openness.
- People are already forking the project into privacy focused versions that strip telemetry and block auto updates.
- Views on the actual Grok model are split, some call it excellent and fast, others say they still lean on Opus or Claude to finish what it produces.
- The sharpest criticism targets the deletion promise itself, since no independent agency has certified that the exfiltrated data was actually destroyed.
Bluesky just trademarked AT Protocol so nobody else can hijack the name, and people have thoughts.
What the article says
- Bluesky bought the trademark for AT Protocol and atproto from a company that was threatening to sue over the name.
- The company says this is defensive. It wants to stop bad actors from claiming the mark, not charge developers for using it.
- Most everyday use stays free. You can build tools, write docs, or name a project after the protocol without a license.
- A license is only needed if you turn the name into a brand, like a paid product, event, or official-sounding certification.
- Bluesky plans to eventually hand the trademark to an independent group that governs the protocol, though that hasn't happened yet.
What HN is saying
- Several commenters pushed back on calling this decentralized when one for-profit company controls the trademark and the identity system behind it.
- One detailed reply explained that leaving Bluesky is technically possible but painful. There are no easy tools for moving your account, and old identities cannot be merged into new ones.
- A commenter dug up who was likely trying to grab the trademark first: a company called Atsign.
- Others were more forgiving, arguing that letting bad actors register the mark first and then buying it back is the least bad option.
- A few replies were simple approval or confusion about what the news even meant.
A new .NET language promises Go and Swift style syntax, but commenters mostly ask why not just use C#
What the article says
- A developer built G#, a new language that compiles straight to .NET, borrowing syntax ideas from Go, Kotlin, and Swift.
- It adds things like packages, a shorter func keyword, data classes, safer handling of null values, and lightweight concurrency.
- The pitch is that it keeps C#'s power but trims the clutter that has built up in the language over twenty years.
- The project is barely two months old and its own commit history looks partly AI generated.
What HN is saying
- Plenty of commenters shrug and ask what this actually gives you that modern C# doesn't already have.
- Several say the changes look cosmetic, just swapping C# syntax for prettier keywords without new semantics or capabilities.
- One detailed critique argues the featured selling points, packages and func, are trivial and that C# already has namespaces and records covering the same ground.
- A running joke is that this is yet another one person language made easy by AI coding tools, not a sign of real demand for new languages.
- A few people do like the cleaner syntax and say it reads well, and someone points out the post is a duplicate of another submission.
A veteran of the open source software wars argues the AI fight is the same battle, just much higher stakes.
What the article says
- The PDF text did not extract, so this is inferred from the title and comments.
- The essay is by an early open source advocate, reframed from an original title about arguing with the father of open source for two years.
- The core claim is that governments, companies and nonprofits should fund free, open AI models the way earlier eras backed open source software.
- The comparison is to past fights over whether openness or corporate control would win in software, now replaying in AI.
What HN is saying
- One popular idea is cash prizes for open models that hit strong benchmarks while running in strict memory limits, rewarding cleverness over just raw funding.
- Skeptics say volunteer effort cannot beat developers who get paid to build AI full time.
- A reply pushes back, noting operating systems, databases and compilers faced the same doubt and went open anyway.
- One sharp tangent argues real power already sits with four people running closed frontier labs, so open weights matter as a check on concentration.
- Several commenters reject public funding outright, saying tax money should go to childcare or health care instead of subsidizing a technology much of the public distrusts.
Stripe wants to buy PayPal for over fifty billion dollars, and the comments section is not happy about it.
What the article says
- Stripe and private equity firm Advent have offered sixty and a half dollars a share for PayPal, valuing it above fifty three billion dollars.
- PayPal has not responded yet. One analyst thinks the offer is a low opening bid and Stripe could go as high as seventy dollars a share.
- Stripe mostly serves merchants. PayPal brings over four hundred million consumer accounts, so the deal would hand Stripe a huge consumer wallet business overnight.
- Combining the two would create one of the largest payment networks in the world, handling trillions in volume each year and giving Stripe more leverage against Visa and Mastercard.
- PayPal's stock has fallen sharply from its pandemic peak, which is part of why it is now a buyout target.
What HN is saying
- Biggest worry is antitrust. Commenters think regulators should block this but doubt they will, with some joking enforcement is now just a matter of how big the bribe needs to be.
- Merchants are split. Some are furious Stripe often blocks industries like cannabis or ketamine treatment that PayPal quietly allows, while others say Visa and Mastercard set those rules, not the payment companies.
- Long time users are torn too. One person says buyers trust PayPal but merchants hate it, while Stripe is the opposite, built for merchants but distrusted by shoppers.
- A few pushed back on merger fears, pointing out plenty of smaller payment processors exist beyond Stripe and PayPal.
- One detailed take argues the real prize is PayPal's bank charter, which could let Stripe build its own card network and cut out middlemen entirely.
A guy let Claude configure his home network gear, then wrote down everything he learned the hard way.
What the article says
- A hobbyist has spent months letting LLMs like Claude Code configure MikroTik routers and wifi gear, with good results overall.
- His top tip is to have the LLM talk to devices over the JSON web interface rather than typing commands through SSH, which trips agents up.
- Always dump the full config before and after a change and keep it in version control, so you can recover if something breaks.
- Test after every single change instead of asking the LLM to set up a whole network at once, since it will still make mistakes.
- For true worst case scenarios, like conflicting IP addresses locking you out, a tool called MAC Telnet lets an LLM reach a device directly over its hardware address.
What HN is saying
- Several commenters have been doing the same thing for months, from fixing wifi over Discord chat to getting VLANs working they never could by hand.
- A recurring warning is that you still need real networking knowledge to catch when the LLM quietly fails or gives up on a hard problem.
- One thread debates whether LLMs should touch devices directly at all, versus having them write a reusable script or Terraform config once and just run that.
- A skeptic calls the whole approach glorified button pushing and says some of these triage setups could have just been a Grafana dashboard.
- Other readers want the same treatment for Ubiquiti gear, and note pfSense and OPNsense already expose solid APIs for this kind of automation.
A guy got a modern AI model running on a storage box so old it predates the chip instructions its own software assumed existed.
What the article says
- A thirteen year old server built for holding hard drives, not doing math, is now running Google's newest open model at reading speed.
- The inference code assumed a newer processor instruction set than this old Xeon has, so it crashed on startup.
- Working with an AI coding assistant, the author traced the real bug deeper: a fused math operation was silently skipped, leaving parts of the model's memory uninitialized.
- That produced fluent looking gibberish, confident sounding text in random languages and symbols, because the model was reading garbage data.
- The fix splits that fused operation into two ordinary steps the old hardware can actually run, and the patch is now up for review on the project's code repository.
What HN is saying
- Commenters mostly agree this is a fun feat but not a money saver. Once you factor in electricity, running locally usually costs more per token than renting from a cloud provider.
- The real reason people run models at home is privacy and independence, not price. Several said they would keep doing it even if it cost more.
- One sharp pushback was that decode speed gets all the attention while prompt processing speed, how fast the model reads your input before answering, is actually the bigger bottleneck on old hardware.
- Readers shared wildly different setups and speeds, from a Raspberry Pi era Pentium 4 running a model as a joke to modern Mac laptops hitting far higher speeds on similar sized models.
- A few just wanted more specs, like how much memory the whole thing actually needed to run.