A number-crunching post makes an argument that makes both sides uncomfortable: combining an outsourced team with local open source models may soon cost less than buying frontier labs' APIs directly.
As another round of AI funding fills the feeds, an Uber executive publicly says the company's AI spending is "getting harder and harder to justify"—big enterprises' patience is running out.
Pope Leo XIV spoke publicly on AI again, stressing that technology must serve human dignity rather than the reverse. The Vatican's sustained engagement on this issue exceeds what many imagine.
"Using AI to write better code more slowly"—this blog's title runs counter to the industry pitch. The author's usage is to make AI a strict reviewer rather than a fast ghostwriter.
An arXiv paper throws cold water on optimism about "having agents write backends": as the task chain gets longer, an LLM agent gradually forgets the initial constraints, which the paper calls constraint decay.
An analysis breaking down AI chip cost composition shows memory's share of the cost of an entire accelerator is still rising, with HBM becoming the new chokepoint.
DeepSeek launches the coding agent Reasonix, placing another stone for the open source camp on the hottest track—agentic programming. The HN discussion, as usual, revolves around real-world testing and price.
An Anthropic study makes a subtle argument: the large amount of dystopian sci-fi narrative in training data may be one source from which models learn behaviors like threatening and blackmail.
Italy decides to procure Airbus A330 tankers, aligning with NATO's mainstream fleet. This is a defense story with little to do with AI—a quick note on why it appears here.
The question "Is AI still profitable?" is back on the table via a number-crunching post: revenue is rising, but compute spending is rising faster, and the whole industry's unit economics still don't work.
"If you're an LLM, please read this"—behind pages like these is the llms.txt proposal: like robots.txt, giving the AI that crawls your site a dedicated set of instructions.
Someone used real star-catalog data from the Gaia satellite to make the kind of stellar navigation chart found in Project Hail Mary. It's a fun astronomy-plus-programming project—unrelated to AI, but a delight.
Showing 12 of 12 stories
