A Model on a $500 GPU Beats Sonnet on a Coding Benchmark
A model running on a $500 consumer GPU beat Claude Sonnet on a coding benchmark. The headline is explosive, but on closer look the conclusion needs a few qualifiers.
A model running on a $500 consumer GPU beat Claude Sonnet on a coding benchmark. The headline is explosive, but on closer look the conclusion needs a few qualifiers.
New York City's hospital system decided to drop its collaboration with Palantir, while at the same time this controversy-plagued company keeps expanding within the UK's health system—the two stories are only complete read against each other.
A post titled "Flying Airports" gained traction on HN, digging up an old romance of aviation history—from airborne aircraft carriers to airship hubs—to play with anew.
TurboQuant headlines extreme quantization compression: squeezing a model to a fraction of its original size while preserving usable accuracy, targeting the two hard needs of inference cost and on-device deployment.
A farewell post to Sora spread through the community, with regret, griping, and reflection each taking a share as people take stock of where this once-viral video-generation product ended up.
A demo showed an iPhone 17 Pro running a 400B-parameter large model. Behind the gimmick of a phone running a flagship-class model, the technical details are more interesting than the conclusion.
A community-compiled Claude Code cheat sheet was pushed to the front page: shortcuts, slash commands, and config tricks all on one page, and even veteran users can pick up something they'd missed.
A practice log: the author trained Claude as a QA tester for their own mobile app, having it operate the UI, run test cases, and write bug reports—the pitfalls and gains along the way are documented in detail.
More and more sites, to guard against AI scraping, are blocking the Internet Archive along with it, and this piece argues why that's a lose-lose choice: AI can't be stopped, and the history goes first.
Le Monde's investigative reporters used public data from a fitness app to track the position of France's aircraft carrier in real time—crew members' running records drew a military secret onto the map.
The open source coding agent OpenCode drew attention on HN: runs in the terminal, swappable models, auditable code—another option for those who don't want to bet their workflow on closed-source tools.
Astral, the makers of ruff and uv, announced it's joining OpenAI. The most popular team in the Python toolchain going into a large-model lab's embrace—the community's feelings are honestly complicated.
Showing 12 of 12 stories