Top Stories

I believe there are entire companies right now under AI psychosis

1990 points · twitter.com

HashiCorp founder Mitchell Hashimoto lit a fuse with a short post arguing that whole organizations have lost the plot on what AI can actually do — boards greenlighting headcount cuts on the assumption agents will “just take over,” products rebuilt around demos that never survive real customers, and engineers being asked to ship things their tools can’t yet deliver. The thread blew up with replies from founders, VCs, and ICs all swapping war stories, and the HN discussion is one of those rare moments where the comment section is essentially required reading.

It’s notable because the pushback isn’t coming from skeptics — it’s coming from the people who’ve been shipping with LLMs the longest. The vibe shift, if there is one, looks less like “AI is overhyped” and more like “the people writing the checks have stopped listening to the people writing the code.”

Zerostack – A Unix-inspired coding agent written in pure Rust

415 points · crates.io

Zerostack hit 1.0 and immediately landed at the top of HN. It’s a coding agent built around small composable Unix-style tools — read, grep, edit, shell — but the whole runtime is pure Rust with no Python, no Node, no Electron, and no cloud dependency once you’ve pointed it at a model. The pitch is that you get the ergonomics of the current crop of agents with the install footprint and startup time of a single static binary.

The comment thread is a who’s-who of people who’ve gotten burned by the dependency sprawl of existing agent CLIs, and the “Unix-inspired” framing seems to be landing — several commenters are pointing out that the agent loop really is just a shell with better autocomplete, and a Rust-native implementation makes that lineage obvious.

Moving away from Tailwind, and learning to structure my CSS

560 points · jvns.ca

Julia Evans does what Julia Evans does best — works through a thing in public until it makes sense. After years on Tailwind she’s moved back to hand-written CSS, not because Tailwind is bad but because she wanted to understand cascade, specificity, and component boundaries directly instead of through a utility-class abstraction. The post is part diary, part beginner-friendly explainer on how to actually organize a stylesheet in 2026.

The HN thread is the usual Tailwind tribal warfare, but with an unusually high signal-to-noise ratio because Julia’s framing is so non-combative. Worth reading even if you have no intention of switching either way.

Codex is now in the ChatGPT mobile app

433 points · openai.com

OpenAI has folded Codex — the cloud-based coding agent — directly into the ChatGPT mobile app. You can hand it a task, walk away from your laptop, and check back from your phone to see PRs it’s drafted against your repos. The pitch is “asynchronous coding from anywhere,” and it’s a clear shot at making agentic coding a background activity rather than something you babysit in a terminal.

The HN reaction is split between “this is the form factor agents have been waiting for” and “I am not ready for my pocket to ship to production.” Either way, mobile agentic coding now has a flagship implementation from the company most people are already paying.

Frontier AI has broken the open CTF format

386 points · kabir.au

A long, mournful post from a longtime capture-the-flag organizer arguing that frontier models — specifically the latest Claude, GPT, and Gemini agents — can now solve a meaningful fraction of the qualifier-round challenges that used to take experienced teams hours. The result is that open online CTFs are getting flooded with model-assisted solves and the signal that made them valuable for hiring and learning is collapsing.

The discussion is doing the work of figuring out what comes next: physically-attended events, problems that require novel hardware or domain knowledge, or accepting that “did you use a model” is the new “did you use the internet” — i.e., a question nobody bothers asking anymore.

SANA-WM, a 2.6B open-source world model for 1-minute 720p video

347 points · nvlabs.github.io

NVIDIA Labs released SANA-WM, a 2.6-billion-parameter world model that generates a full minute of coherent 720p video from a prompt — and the weights are actually open. The demos show sustained physics, persistent characters across long shots, and a meaningful drop in the “morphing” artifacts that have haunted long-form video models.

What has people excited isn’t just the quality, it’s the size. A 2.6B model that does this implies the recipe has gotten dramatically more efficient, which means downstream researchers and small teams now have a credible base to build on rather than being locked out of the frontier by compute.

OpenAI and Government of Malta partner to roll out ChatGPT Plus to all citizens

204 points · openai.com

Malta has signed a deal with OpenAI to provide ChatGPT Plus to every Maltese citizen as a public service, becoming the first national government to do so. The announcement frames it as digital infrastructure on par with broadband, with a stated goal of leveling access to AI tools across education, small business, and the civil service.

HN commenters are split between “this is the kind of bet small countries should be making” and questions about data sovereignty, dependency on a US private company for civic infrastructure, and what happens at contract renewal. It’s a small country making a big move, and it’s going to be watched closely as a model — or a cautionary tale — by everyone else.

Mozilla to UK regulators: VPNs are essential privacy and security tools

245 points · blog.mozilla.org

Mozilla has filed a public response to the UK regulator’s consultation on whether VPN usage should be restricted in the context of the Online Safety Act, and the answer is a polite-but-firm no. The post lays out the case that VPNs are critical infrastructure for journalists, abuse survivors, remote workers, and ordinary people doing ordinary things, and that any rule restricting their use would harm the people it claims to protect.

It’s the latest entry in a recurring pattern — a government proposes something well-intentioned, browser vendors and privacy orgs explain why it would backfire, the conversation continues. HN is reading it as a barometer of how seriously UK regulators are taking the technical pushback.

δ-mem: Efficient Online Memory for Large Language Models

220 points · arxiv.org

A new paper proposes δ-mem, an online memory mechanism that lets LLMs accumulate and retrieve information during a long-running session without the cost of growing a full context window. The approach uses a small learned update step over a compressed memory state — hence the delta — and the experiments suggest competitive performance against long-context baselines at a fraction of the inference cost.

If it holds up, this is a meaningful piece of the puzzle for agent runs that need to last hours or days without forgetting earlier decisions. The HN thread is doing the work of comparing it to prior memory papers and arguing about whether “online” really means what the authors say it means.

MCP Hello Page

102 points · hybridlogic.co.uk

A small, deliberately minimal example of an MCP (Model Context Protocol) server that does exactly one thing — serve a hello page — written as a teaching tool. The post walks through what MCP actually is in concrete terms, why the protocol matters, and what the smallest possible useful implementation looks like.

It’s getting upvoted because the broader MCP ecosystem has been hard to enter — there are now real specs, several SDKs, and a lot of marketing, but very few short examples that fit on one page. This one does, which is exactly what the protocol needs to graduate from “interesting” to “boring infrastructure.”