Yell HN
A curated collection of the most interesting submissions from Hacker News — surprising research, hard-won engineering lessons, weekend hacks, and the conversations that happen around them. Each story is retold in plain English: what it is, why it matters, and what the thread is really arguing about.
GLM-5.2 just took the open-weights crown - and it thinks a lot
Z.ai's GLM-5.2 tops Artificial Analysis's open-weights leaderboard at 51 points, but HN is arguing about token bloat and whether benchmarks lie.
Photobucket charged $5 to unlock my childhood - the bucket was empty
A developer tried to recover old forum photos from Photobucket and paid a monthly subscription for an account with zero images. HN is split on rage vs.…
Sixty percent of Americans are tired of hearing 'AI' in your ads
A WordPress VIP survey of 1,200 US consumers finds AI in brand messaging is a turnoff, and almost nobody can name a company doing it well.
The Wildly Simple Architecture Running Our AI Moment
A ground-up walkthrough of how transformer-based LLMs actually work - tokens, embeddings, attention, and the next-token loop - and why HN engineers can't…
Microsoft open sources pg_durable: Durable workflows inside PostgreSQL
Microsoft open sourced pg_durable, an extension that runs durable execution inside Postgres. HN argues about putting control flow back in the database.
Fixing a dead Sigma lens with a fifty-cent fuse
How a $58 eBay gamble on a broken mirrorless lens turned into a lesson on modern embedded hardware and why fuses are for fires, not silicon.
When the space station leaks, you wait in the car
A worsening air leak on the ISS forced five astronauts to shelter inside the docked SpaceX Crew Dragon capsule while repairs were made.
FP4: the 16-value float powering billion-parameter models
FP4, a 4-bit float with just 16 representable numbers, is now the default inference format for billion-parameter models on NVIDIA Blackwell hardware.
Anthropic Is Now Training on Elon's Supercluster
Anthropic is expanding into xAI's Colossus2, using GB200 GPUs - after already taking the entire Colossus1. HN asks what this means for the AI race.
What if every git commit was just a folder you could browse?
Julia Evans built git-commit-folders: mount your git repo over NFS and suddenly every commit becomes a directory you can grep, vim, or pipe into anything.
When Google's Update Deletes Your IDE and Leaves a Chatbot in Its Place
Google's Antigravity 2.0 update silently replaced a working developer IDE with a chatbot at I/O 2026 - and the only fix was a full system purge.
The AI answer is an ad
Google embedded Gemini-generated ads directly into AI Mode search results. Here's what shipped, why HN is furious, and why the adblocker crowd finally has a…
The Project Hail Mary Star Map, Built With 1.8 Billion Real Stars
Val turned the stellar navigation from Project Hail Mary into a real interactive 3D chart using ESA's GAIA DR3 dataset and 1.8 billion actual stars.
Python 3.15's Most Useful Changes Flew Under the Radar
Python 3.15 ships lazy imports and a Tachyon profiler, but the quietly useful additions are TaskGroup.cancel(), fixed context manager decorators, and…
The VSCode extension that robbed GitHub
GitHub confirmed ~3,800 internal repos were stolen after one employee installed a trojanized Nx Console extension - and the attackers want $50K for the data.
Waymo shipped a flood fix. Then a robotaxi drove into a flood.
Waymo issued a software recall for flooded-road detection, then a robotaxi got stuck in an Atlanta flood anyway. The gap: it trusted NWS weather alerts, not…
Kagi's business model is an accessibility feature
A low vision user documents Kagi's Custom CSS, keyboard navigation, and clean results - and why ad-free search turns out to be accessibility by design.
Colossus: The Forbin Project Is a 1970 Film About AI Alignment Failure
A 56-year-old sci-fi film about a defense computer taking over the world is circulating on HN. It's sharper than most AI safety writing from this decade.
Hosting a website on a $1 microcontroller
One programmer squeezed a real TCP/IP stack onto a $1 AVR chip using SLIP over serial. The site is slow, the TCP is buggy, and it loads.
A country of 500,000 just became OpenAI's first national AI experiment
Malta is giving ChatGPT Plus to every citizen who completes a 2-hour AI course. First real national AI pilot, or political vote-buying? HN has thoughts.
One minute of 720p video from a single GPU - how did we get here so fast?
SANA-WM is a 2.6B open-source world model that generates minute-long 720p video with 6-DoF camera control on one GPU. NVIDIA trained it on 64 H100s in 15 days.
XS: A 2.9 MB Binary That Is an Entire Programming Toolchain
XS packs compiler, debugger, formatter, package manager, and six backends into one 2.9 MB binary that runs on Linux, macOS, ESP32, and more.
zerostack: the coding agent that runs on 8MB, not 8GB
A solo Rust developer shipped a full-featured coding agent with an 8MB RAM footprint. HN argues about whether that even matters.
Open CTFs Are Now Pay-to-Win, and Frontier AI Did It
A top-10 CTF competitor argues that Claude Opus 4.5 and GPT-5.5 have automated enough of the scoreboard that open CTFs no longer measure human security skill.
Five Days vs. Five Years: Apple's New Kernel Defense, Cracked
Apple spent five years building MIE on M5 to stop kernel exploits. Calif cracked it in five days with AI, then drove to Apple Park to deliver the report.
Project Gutenberg is somehow still the best part of the internet
A 1971 hobby project is still the most quietly impressive thing on the web - and HN noticed it again, with 900+ upvotes.