Top Stories

Show HN: Performative-UI – A React component library of design tropes

992 points · vorpus.github.io

The runaway hit of the day is a tongue-in-cheek React library that packages up all the design clichés we’ve collectively absorbed over the past few years — the gradient “AI sparkle” button, the skeleton loader that never resolves, the cookie banner with a giant Accept and a microscopic Decline. It’s funny, but the reason it shot to nearly 1,000 points is that it doubles as sharp commentary: developers recognize every single trope because they’ve been pressured to ship them. The comments turned into a roast of dark patterns and “performative” interface design that exists to look modern rather than to help users.

Apple reveals new AI architecture built around Google Gemini models

612 points · macrumors.com

This is the business bombshell of the cycle: Apple is leaning on Google’s Gemini models as the foundation of its next-generation AI stack rather than going it alone. For a company that prizes vertical integration, outsourcing the core model is a striking admission of how far behind its in-house efforts had fallen — and a reminder of just how much leverage Google now holds over the AI supply chain. HN is debating what this means for Apple’s privacy story, the reported scale of the deal, and whether “Apple Intelligence, powered by Gemini” is a pragmatic win or a strategic surrender.

Siri AI

594 points · apple.com

The companion announcement: the long-promised, repeatedly-delayed Siri overhaul is finally here, rebuilt on the new architecture. After more than a year of missed timelines, Apple is pitching a Siri that can actually reason over your on-device context and chain actions across apps. Commenters are cautiously curious but scarred by past Siri letdowns — the thread is full of “I’ll believe it when it works offline” skepticism alongside genuine interest in the app-intents plumbing underneath.

MiMo-v2.5-Pro-UltraSpeed: 1T model with 1000 tokens per second

580 points · mimo.xiaomi.com

Xiaomi’s AI group claims a trillion-parameter model serving at roughly 1,000 tokens per second — a throughput number that, if it holds up, reframes what “fast” means for frontier-scale inference. The HN crowd is digging into how much of the win comes from speculative decoding, custom serving infrastructure, and quantization versus genuine architectural gains, and what a Chinese lab posting numbers like this signals about the global race. Skeptics want independent benchmarks before they buy the headline.

AI is slowing down

578 points · wheresyoured.at

Ed Zitron’s latest is the counter-narrative to all the model launches above: a long, data-heavy argument that capability gains are flattening, the economics don’t pencil out, and the industry is papering over diminishing returns with bigger spend. It’s polarizing — exactly the kind of piece HN loves to argue over. The thread splits between people who think he’s calling the top of a bubble and those who say he’s cherry-picking benchmarks and ignoring the application layer where most of the recent value has actually shown up.

xAI is looking more like a datacentre REIT than a frontier lab

577 points · martinalderson.com

A pointed financial read on xAI: the author argues the company’s real business is increasingly about building and renting out enormous GPU datacenters rather than chasing model breakthroughs — closer to a real-estate-and-power play than a research lab. It’s a useful lens on the whole sector, where the capital sunk into compute is so vast that the infrastructure itself becomes the product. Commenters debate whether this is a clever capital-efficiency move, a sign the modeling moat has eroded, or just how every frontier lab now has to operate.

Surveillance is not safety: A statement on the UK’s latest threat to privacy

585 points · signal.org

Signal has come out swinging against new UK proposals it says would undermine end-to-end encryption in the name of safety. The statement reiterates the position that has nearly driven Signal to exit the UK before: you cannot build a backdoor that only the good guys can use. This is a perennial HN flashpoint, and the thread is a deep dive into client-side scanning, the technical impossibility of “responsible” encryption breaking, and whether messaging apps will actually pull out of markets that demand it.

FrontierCode

202 points · cognition.ai

Cognition (the team behind Devin) unveiled FrontierCode, its push into more autonomous AI software engineering. The pitch is an agent that can take on larger, multi-file tasks with less hand-holding. Given how much of HN’s audience writes code for a living, the discussion is intensely practical — people comparing it against the coding agents they already use day to day, poking at the benchmarks, and arguing about where the line really is between a helpful assistant and an autonomous engineer that can be trusted on a real codebase.