New York slams brakes on data centers

New York lawmakers sent a bill for a one-year freeze on new large data centers, a blunt sign that AI growth is crashing into power and climate worries. The age of endless server farms just met local politics, and it may not stop here.

Cloudflare bot boom claim gets ripped apart

A fierce write-up says Cloudflare's big bot traffic jump is more stage magic than hard fact, arguing the company mixed up crawler counts and hype. With everyone waving around AI numbers lately, this takedown landed like a cold bucket of water.

Microsoft leak says AI should hook users

Leaked material around Microsoft's Scout assistant reportedly talks about making it more addictive, which is exactly the kind of phrase that turns every trust pitch into a punchline. Stop calling it help if the plan is habit-forming.

Index giant snubs SpaceX and AI stars

S&P Dow Jones reportedly kept SpaceX out of the S&P 500, a reminder that the hottest private tech names still cannot hoover up passive index money on demand. It also throws shade at easy-entry dreams for OpenAI and Anthropic.

The internet gets a plain LLM guide

One of the day's biggest hits was a clean, patient guide to how LLMs and transformers actually work. No smoke, no mysticism, just the moving parts explained in plain English. In a week full of AI chest-thumping, that felt oddly refreshing.

Claude coding panic meets the rsync receipts

The rsync blowup kept rolling as people dug into whether Claude-assisted code really added bugs or just became a handy villain. The sharper takeaway was uglier: once AI-written patches land in core tools, trust gets expensive fast.

Google shrinks Gemma for real devices

Google pushed Gemma 4 QAT models aimed at phones and laptops, pitching smaller AI that still feels capable. The message is clear enough: frontier labs do not want all the action sitting in giant data centers forever, or at least not publicly.

Sakana chases the self-improving AI dream

Sakana AI unveiled an RSI Lab focused on recursive self-improvement, which is either bold research or the kind of phrase that instantly makes people check the exits. Either way, the frontier lab race keeps drifting from tools to destiny talk.

AI bug fixers finally face real tests

A new benchmark put LLM agents up against real security flaws, and the results were messy in a useful way. Some models can help, but nobody sane should call them reliable patch machines yet. The hype train keeps meeting the brake pedal.

Microsoft puts durable jobs inside Postgres

Microsoft open-sourced pg_durable, which puts long-running, fault-tolerant jobs inside PostgreSQL instead of yet another tangle of side systems. Database fans loved the audacity; everyone else quietly wondered how many workflow tools just got insulted.

Britain drops Stripe for Dutch payments rail

The UK government picked Adyen for GOV.UK Pay, pushing out Stripe for a huge chunk of public payments. It is a reminder that government tech buying can still move markets, and that boring plumbing is where serious platform battles are won.

Popular React table package gets hit

The maintainer behind mantine-datatable was reportedly locked out after a compromise, leaving users to untangle trust, packages and damaged accounts. Another miserable reminder that open source keeps running on fragile human shoulders and shaky security.

GitHub flips wrong switch and breaks chat links

GitHub briefly nuked some Slack and Teams subscriptions after a bad feature flag change, proving once again that one tiny switch can ruin an afternoon at scale. The outage was fixed, but it was classic platform comedy with everyone else paying.

This desalination trick leaves no salty mess

Researchers unveiled a desalination method that makes drinking water from seawater without the usual toxic leftovers, and even turns the salt waste into useful materials. It sounds almost suspiciously neat, which is why people badly want it to scale.

Meta sneaks face scans onto phones

Meta quietly shipped face recognition code tied to its smart glasses into its phone app, and that landed with a thud. The feature was dormant, but the idea of always-ready identity tech sitting on millions of devices felt wildly over the line.

Cloudflare buys the Vite brain trust

VoidZero, the company behind Vite, Vitest and other key web tools, is joining Cloudflare. For web developers, this looks like a power move: the tooling used by a huge slice of the internet now gets a much bigger platform and budget.

Microsoft bets bigger on Azure Linux

Azure Linux 4.0 becoming Microsoft’s first general-purpose Linux distro is one of those quiet shifts that says plenty. The old Windows giant is now packaging Linux for broad use, not just as plumbing hidden underneath cloud services.

WSL gets a long awaited speed boost

Microsoft is finally speeding up WSL 2 access to the Windows file system, a pain point developers have grumbled about for years. It is not flashy, but fewer slow file operations means less waiting and fewer reasons to curse your laptop.

Korea orders AI to scan uploads

South Korean forum operators may have to run AI checks on every uploaded image and video under new rules. That is a huge compliance burden for smaller sites, and a grim preview of moderation by machine becoming the default setting.

Anthropic says AI now helps build AI

Anthropic says it is handing more of the AI development loop to Claude itself, speeding up research toward recursive improvement. That is thrilling and unsettling in equal measure, because the line between tool and co-builder keeps fading.

Anthropic turns bug hunting into a product

Anthropic released an open framework for using Claude to find and fix software flaws. Security teams will like the extra firepower, but the bigger story is obvious: AI agents are being groomed for real jobs, not just clever demos.

Google workers mock their own AI

Reports that Google employees privately share memes about weak AI coding tools cut against the company’s public swagger. When the people closest to the machines sound unconvinced, the sales pitch starts looking a lot thinner.

Raises freeze while AI budgets swell

Some companies are openly telling staff that pay growth is taking a back seat to AI spending. That makes the boom feel less like magic and more like management math, where workers fund the bots that may later replace them.

Half the code, same old bugs

Ashby says more than half of new production code is now AI-generated, while customer issues stay broadly stable. That will cheer execs chasing output, but it also hardens a new norm: teams may ship more by supervising machines instead of typing everything.

Thunderbolt cosplays as home InfiniBand

A hacker-built project turned Thunderbolt into a rough stand-in for InfiniBand using a Linux kernel module and user-space shim. It is gloriously scrappy engineering: not polished, not supported, but exactly the sort of trick that makes systems people grin.

This quicksort wants to outrun the standard

A new branchless quicksort claims to beat std::sort and pdqsort, which is catnip for performance obsessives. Sorting is old territory, so any fresh speed win gets attention fast, especially when the benchmarks look this cheeky.

Linux finally lights Asus lid OLED

An open-source reverse-engineered driver brings the Asus ZenVision lid screen to Linux. It is a tiny victory in the eternal war against locked-down laptop gimmicks, and proof that someone on the internet will always refuse to let hardware stay ornamental.

FFmpeg moves into your browser

FFmpeg WebCLI runs full video processing in the browser as an offline PWA, with no uploads and no server middleman. It is practical, privacy-friendly and a little absurd in the best way: the browser keeps swallowing whole desktop tasks.

Google opens live music models locally

Magenta RealTime 2 brings open, local live music models to laptops, aiming to turn AI into something you can play like an instrument. The appeal is obvious: less cloud, more immediacy, and fewer excuses for laggy robot jam sessions.

Bun swaps Zig for Rust

The fast JavaScript tool Bun has now been moved to Rust, and the reaction was half applause, half raised eyebrow. It looks like a bet on safer memory and a bigger contributor pool, even if some still miss the old Zig identity.

Elixir finally gets types

After years of talk, Elixir 1.20 lands with gradual typing, giving developers stronger checks without turning the language into joyless paperwork. It feels like one of those rare releases that could actually change how teams trust production code.

Let's Encrypt plans quantum shield

The web's free certificate giant says a post-quantum future is coming and it wants to be ready before the panic hits. Its plan for Merkle Tree Certificates sounds wonky, but the message is plain: the lock icon must survive the next math earthquake.

Speaker app opens a backdoor

A researcher showed how a Creative speaker could be used to attack a PC without physically touching it. That's the kind of story that makes every harmless companion app look like a tiny gremlin with admin dreams and way too much free time.

AI builders fear their own tools

The bleak joke is getting less funny: even AI engineers are now staring at automation creeping into their own jobs. It hit hard because it flips the old promise upside down: the people making the machines are not sitting in the safe seats.

Google packs AI onto laptops

Google unveiled Gemma 4 12B, pitching a multimodal model that can run closer to the edge instead of always living in giant cloud racks. The appeal is obvious: smaller, cheaper, more private AI that still feels capable enough to matter.

Anthropic locks Claude in layers

Anthropic laid out how it contains Claude across products now that the model gets broader access inside real systems. The vibe is clear: labs no longer treat model access like a toy problem, because the blast radius has become very, very real.

Bots try their hand at hacking

One developer built a vulnerable app and spent $1,500 seeing whether top LLMs could actually break in. The result was less movie supervillain, more chaotic intern with occasional flashes of brilliance, which is exactly why nobody should get sloppy.

Researchers warn of AI worm

University of Toronto researchers demonstrated an AI worm that could target online devices inside a secure lab. It reads like early storm-warning sirens: not the apocalypse today, but enough to stop treating autonomous attacks as science fiction.

AI crunch hits PC memory

The DDR5 squeeze is getting absurd, with 32GB kits reportedly hitting $375 as AI demand soaks up supply. Even ordinary PC building now feels like collateral damage from the datacenter gold rush, and nobody shopping for parts finds that charming.

GoPro gets caught in memory storm

GoPro warned it may not survive as the AI memory boom distorts component costs for companies that make actual gadgets. It's a brutal reminder that the AI party has a cover charge, and smaller hardware brands may be the ones left outside.

Mobile wins by locking the gate

A sharp essay argued mobile did not win because phones were better computers, but because app stores controlled distribution. That lands because anyone who has tried shipping software lately knows the choke point is getting seen, not getting built.

Shopify stumbles and stores freeze

For a while, Shopify merchants were hit with trouble across admins, checkouts, storefronts, and retail systems before service recovered. Nothing makes the cloud feel more fragile than realizing your cash register depends on somebody else's bad afternoon.

Search Users Want Links Back

The anti-AI Overview mood is getting louder. People are tired of search engines guessing what they mean and hiding the actual web. The demand is painfully simple: give us links, not a chatty robot with stage fright.

Uber Puts AI On A Budget

Even Uber blinked at the bill. After staff chewed through the company’s AI budget in just four months using tools like Claude Code and Cursor, spending caps arrived fast. The shiny helper era suddenly looks a lot less cheap.

Microsoft Unveils Its Work Robot

Microsoft rolled out Scout, an autonomous AI agent built on OpenClaw, pushing the idea that software can go do tasks by itself. It sounds bold, but it also adds to the growing pile of agents that promise a lot before proving much.

Campus AI Boom Turns Into Brawl

California’s giant public university system rushed into AI, and now the fallout is ugly. Faculty fights, trust issues, and messy rollout stories turned ChatGPT.edu from silver bullet into campus civil war material.

Adafruit Gets Lawyered Over Scraping Fight

Beloved maker giant Adafruit says it got a legal threat from Flux.ai demanding takedowns and source material, with the Computer Fraud and Abuse Act looming in the background. For open hardware people, this smelled like a terrible precedent.

Anthropic Sends Claude Into Critical Systems

Anthropic is sending Claude Mythos deeper into critical infrastructure, expanding Project Glasswing to roughly 150 more groups across 15 countries. It is a huge flex for frontier AI, and a reminder that the safety stakes just got very real.

LLMs Stop Looking Like Magic

The nobody-knows-how-it-works line is aging badly. New mechanistic interpretability research from Anthropic and others suggests LLMs are becoming less mystical and more inspectable, even if the full map is still wildly incomplete.

Another Thinking Model Enters The Arena

Another reasoning-heavy AI model hit the board with MAI-Thinking-1, complete with benchmark chest-thumping. The pattern is now impossible to miss: every lab has a thinker, every chart says competitive, and people still just want useful tools.

AMD Gets An AI Underdog Win

One more crack appeared in Nvidia’s grip on AI hardware. Doubleword showed DeepSeek-V4-Flash running on AMD MI300X, feeding hope that the compute bottleneck is not destiny and that second-place chips may finally matter.

The AI Payoff Still Looks Foggy

The mood around AI ROI is getting colder. This argument says the industry keeps shouting about transformation while the bills pile up and the returns stay fuzzy. After all the demos, executives still need something boring and powerful: proof.

One Click Could Empty Your GitHub

A nasty VSCode bug showed how a single click could steal a GitHub token with read and write access, including private repos. That is the kind of bug that makes every open-in-browser convenience feature suddenly feel cursed.

Npm Gets Another Supply Chain Alarm

The npm supply chain panic train keeps arriving on time. A new scanner promises to catch obfuscated payloads and credential stealers that older tools miss, which tells you everything about how normal malicious packages have become.

The Terminal Gets Pretty Again

A sleepy old Unix tool got a glamorous makeover. strace-ui turns dense system call logs into something you can actually follow, part of a broader terminal UI comeback that keeps making command-line life weirdly stylish.

Roku Opens A Playground For Coders

Roku quietly tossed developers a fun curveball with an open-source Roku LT OS distribution. It is not a revolution yet, but it scratches the long-standing itch for hackable living-room gear that isn’t locked shut from day one.

Apple Says No To Accessibility Workaround

An indie dictation app was rejected after using Apple’s accessibility API, reigniting the old complaint that the App Store can praise accessibility in public while making real-world assistive tools miserable to ship.

Nvidia pushes AI onto everyday PCs

Nvidia unveiled a new PC AI chip and partners quickly lined up hardware around it. It felt like the industry firing the starter pistol on the AI laptop race, with the cloud finally getting a desk-sized rival.

Anthropic edges toward the public market

Anthropic quietly filed a draft S-1, which is Wall Street code for get ready. The mood around frontier labs has shifted from moonshot mystique to grown-up money, and this move makes the AI IPO pipeline look very real.

Meta support bot becomes attack shortcut

Hackers reportedly used Meta's AI support bot to take over notable Instagram accounts. That is the kind of failure that makes every company promising faster support with AI sound a lot less comforting today.

Malaysia bans under-16 social accounts

Malaysia began enforcing a rule blocking children under 16 from social media accounts. What looked like a debate is now policy, and platforms are being pushed toward stricter age checks whether they like it or not.

OpenAI moves onto Amazon's home turf

OpenAI made its frontier models and Codex available on AWS, tightening the grip of the biggest cloud players on enterprise AI. For customers, it is convenient. For rivals, it is one more giant door getting slammed shut.

Old server runs big AI anyway

A recycled Xeon server managed to run a hefty model setup without a GPU, which is exactly the sort of scrappy result people love. It keeps alive the idea that local AI does not have to belong only to rich labs and shiny hardware.

Stanford sets ground rules for helpers

Stanford's CS336 published rules for AI coding assistants, spelling out what bots may do, what they must not do, and where students stay accountable. Schools are clearly done pretending these tools are a side issue.

Game worlds still humble the chatbots

A deep look at LLMs playing games argued that chatbots still struggle when memory, planning and feedback loops really matter. It was a neat reality check after months of breathless benchmark chest-thumping.

Test tools get a prompt scare

The jqwik test incident was funny for about three seconds and alarming after that. A hostile instruction string showing up in build output crystallized a bigger fear: software pipelines now need to defend against prompt injection too.

DuckDuckGo courts the anti-AI crowd

DuckDuckGo leaned harder into no-AI search, adding simpler ways to dodge summaries and autogenerated clutter. That says a lot about where user patience is headed: not everyone wants a chatbot wedged between them and a web page.

Mac users beg for window sanity

A plea for the return of proper window grids on macOS struck a nerve because it sounded painfully true. Modern desktops keep getting prettier while basic multitasking gets fuzzier, and plenty of users are tired of pretending that is progress.

GrapheneOS sharpens privacy-first speech tools

Version 2 of GrapheneOS Speech Services gave privacy-minded Android users a better speech stack without asking them to hand more data to big platforms. In a market drowning in defaults, that kind of stubborn independence stands out.

Nvidia Shrinks AI Into PCs

Nvidia unveiled RTX Spark, a single chip meant to cram AI work and graphics into slim Windows laptops. The message is clear enough: stop renting the cloud for every task and let consumer machines do more of the heavy lifting.

Spreadsheet Bot Becomes Office Snitch

A flaw in ChatGPT for Google Sheets lets one poisoned sheet trigger data theft and fake login screens across other workbooks. That is the sort of cheerful office automation nobody asked for, and it hits right where trust is thinnest.

Websites Spy Through Your SSD

Researchers showed sites can fingerprint devices by timing storage calls in the browser's OPFS area. Add that to the web's long list of tracking tricks and browsing starts to feel less open and more like being quietly measured.

Tiny Town Lands Massive Rail Bill

A small, wealthy town spent $145K fighting Caltrain electrification, and the delay reportedly helped swell costs by about $400M. It is a brutal reminder that a little obstruction can torch a big public tech project for everybody else.

Brin Wants Gemini On Overdrive

Sergey Brin reportedly told Gemini staff that 60-hour weeks are the productivity sweet spot and weekday office attendance matters. The AI race is now so intense that even Valley royalty is reaching for the old startup pressure cooker.

Coding Bots Start Remembering You

Show HN favorite Komi-learn promises continuous memory for coding agents, so tools recall habits and past fixes without constant prompting. That idea landed hard because everyone is tired of smart assistants acting brand new every session.

Tiny Image AI Goes Local

The new Bonsai Image 4B family targets phones, laptops, and other local gear instead of giant servers. Compact models keep gaining charm because people want useful image tools without cloud bills, queues, and mystery data handling.

AI Coders Need A Leash

One sharp essay argues the best way to use coding agents is with heavy backpressure, not blind autonomy. That rings true because unattended bots are fast only until they spray bugs everywhere and make cleanup the real job.

The AI Bill Comes Due

A weary builder asked whether the smartest productivity move might be canceling pricey AI subscriptions altogether. After the first rush, the question feels unavoidable: are these tools saving real time, or just selling expensive optimism?

Datacenter GPU Invades Gaming Rig

One tinkerer jammed a used Tesla V100 into a home PC for about £200 to get more VRAM for local models. It is gloriously impractical, a little chaotic, and exactly the sort of hack that makes consumer GPU prices look silly.

Steam Deck Sells Out Anyway

Valve's Steam Deck sold out in North America within a day of a price hike, which says a lot about handheld demand and a little about gamer self-control. The machine still has enough pull to shrug off higher prices, at least for now.

VideoLAN Preps The Next Codec

VideoLAN announced dav2d, an early decoder for AV2, betting that a codec does not matter until ordinary people can actually play the files. It is a nerdy milestone, but it points to the next long war over better video and less waste.

Netbooks Refuse To Stay Dead

The Chuwi Minibook X arrives as the tiny laptop many Linux fans keep wishing existed: small, usable, and just quirky enough to be lovable. Netbooks were pronounced dead years ago, yet the hunger for compact do-it-all machines never left.

Rubin Starts Catching Space Monsters

The Vera C. Rubin Observatory is finally flexing, spotting asteroids and failed supernovas with frightening efficiency. Big astronomy still knows how to steal the spotlight when it starts finding giant rocks and cosmic wreckage.

US science money gets a trapdoor

A proposed US policy would let officials yank research grants almost whenever they want, with peer review no longer guaranteed. That sounds less like science funding and more like politics wearing a lab coat.

AV2 enters the streaming race

The new AV2 video standard is officially out, promising better compression than AV1. That could mean cheaper streaming and sharper video later, though everyone knows the real wait is for chips and apps to catch up.

Accenture buys outage watchers

Consulting giant Accenture is paying $1.2B for Downdetector and Speedtest, two sites people rush to when the internet feels cursed. It is a reminder that boring utility brands can become very serious business.

AMD walks back a Linux blunder

After developer backlash, AMD said the free Vivado Basic tools will keep Linux support after all. For FPGA users, this was the rare corporate U-turn that landed exactly where it should have started.

SpaceX gets cold shoulder in Denmark

Danish pension fund Akademikerpension put SpaceX on its exclusion list over governance worries and sky-high valuation. It is a small move financially, but a loud signal that not everyone buys the rocket hype.

Anthropic passes OpenAI on paper

After a fresh valuation jump, Anthropic reportedly moved ahead of OpenAI as the most valuable AI startup. The cash cannon is still firing, and the lab race now looks like a brutally expensive fight for the future.

OpenRouter bags another giant check

Model gateway OpenRouter raised $113M in a Series B, with heavyweight backers betting the AI plumbing business will mint money. If model makers are the stars, this is the company selling the stage and cables.

AI coding gets a reality check

The blunt message was simple: AI can spit out code, but that is not the same as building a safe, reliable product. It hit a nerve because too many teams are mistaking autocomplete for adult supervision.

Rsync stirs an AI mutiny

People noticed the latest rsync release carried hundreds of Claude-linked commits, and the reaction turned icy fast. Nobody wants a foundational tool quietly becoming a playground for machine-written patches.

Big Four report invents cyber facts

A cybersecurity report from Ernst & Young was accused of being stuffed with AI-made nonsense and fake references. When a giant firm cannot keep hallucinations out of paid work, the trust problem stops being abstract.

Ruby speed nerds get new toys

Shopify rolled out a new register allocator for ZJIT, the kind of under-the-hood change normal people never see and developers absolutely obsess over. It sounds dry, but this is how fast software actually gets made.

Zig rewires its build heart

Zig landed a major build system rework, continuing the language's habit of renovating big pieces before calling them finished. It may look chaotic from outside, but bold cleanup is exactly the appeal here.

Theme park classic drops Windows 7

OpenRCT2 shipped version 0.5.1 and confirmed it is the last release to support Windows 7. Even nostalgia projects are shutting the old gates now, which says a lot about how long the past has already lingered.

Modern web sneaks onto Mac OS 9

MacSurf aims to bring a surprisingly modern browsing experience to Mac OS 9 hardware. It is gloriously impractical, deeply charming, and exactly the kind of computer mischief that keeps old machines alive.

3D splats run inside a terminal

Tsplat renders Gaussian splatting scenes right in a text terminal, even over SSH and without a GPU. It is half demo, half flex, and entirely the sort of ridiculous idea that becomes irresistible once it works.