Top Stories

”I Believe There Are Entire Companies Right Now Under AI Psychosis”

1414 points · twitter.com

Mitchell Hashimoto’s drive-by tweet — that whole companies are operating under a kind of “AI psychosis,” chasing model-confidence over reality — has become the most-discussed post on HN today. The thread caught fire because Hashimoto isn’t a doomer: he runs Ghostty, helped build HashiCorp, and has been broadly bullish on coding agents. Coming from him, “psychosis” is a strong word.

The comments split along familiar lines. Some readers see it as a long-overdue gut-check on enterprises shipping LLM features they can’t evaluate; others argue the frame is too sweeping and dismisses real productivity gains. Either way, “AI psychosis” looks like it’s about to become the meme of the week in tech management circles.

Project Gutenberg – Keeps Getting Better

963 points · gutenberg.org

A nostalgic appreciation post for Project Gutenberg, the 50+ year old free e-book archive, rocketed to near the top of HN with 963 points. The submission highlights how Gutenberg has steadily improved its catalog browser, search, and metadata over the last few years — quietly, without fanfare, without an LLM in sight.

The comment thread reads like a love letter. People share favorite obscure 19th-century novels, talk about how the site’s clean HTML beats every modern e-reader, and reflect on the contrast between Gutenberg’s volunteer-driven persistence and the venture-funded book platforms that come and go. It’s the kind of post HN occasionally needs as ballast.

Bill to Block Publishers From Killing Online Games Advances in California

496 points · arstechnica.com

California’s “Stop Killing Games”-aligned bill has cleared a key committee, taking it one step closer to forcing publishers to keep online games playable — or hand over server software — after they pull the plug. The movement, originally galvanized by Ubisoft’s shutdown of The Crew, has been gaining real legislative teeth.

The HN reaction is mostly cheering, with a healthy minority worried about the unintended consequences for MMO economics, anti-cheat infrastructure, and small studios. But the bigger story is that “you bought it, you should still be able to play it” is becoming actual law in a major market — and California’s reach means publishers serving the US will have to comply.

A 0-Click Exploit Chain for the Pixel 10

386 points · projectzero.google

Google’s own Project Zero published a deep dive into a zero-click exploit chain against the Pixel 10 — meaning the target’s phone could be compromised with no user interaction at all. The post walks through the bug primitives, the chain assembly, and the patches that landed alongside disclosure.

What makes this notable beyond the usual P0 write-up: it’s Google researchers picking apart Google’s flagship phone, and the chain leans on weaknesses in the new Tensor G5 secure enclave. The HN thread is a mix of security pros admiring the bug-chaining craft and Pixel owners checking their patch level. Update your phone.

SQL Patterns I Use to Catch Transaction Fraud

248 points · analytics.fixelsmith.com

A working analyst’s guide to spotting fraud with nothing fancier than SQL — window functions, percentile filters, velocity checks, and a handful of well-chosen joins. The post is grounded in real card-not-present patterns the author has seen, and reads like the kind of internal doc most fraud teams have but rarely publish.

The comments are full of practitioners adding their own tricks: “look at the gap between account creation and first transaction,” “watch BIN-country mismatches,” “the most useful column is the one you don’t have yet.” It’s a nice reminder that you don’t need a graph database and an ML team to get the first 80% of fraud detection right.

The Sigmoids Won’t Save You

212 points · astralcodexten.com

Scott Alexander pushes back on a comfortable argument inside AI-safety-skeptic circles: that progress curves are sigmoidal, so we’ll naturally hit a plateau before anything dangerous happens. His point — that “sigmoid” is a vibe, not a forecast, and that no one has shown why this particular technology should flatten on a humanly comfortable schedule — is making the rounds.

HN’s reaction is split between people who think Alexander is making a much-needed epistemic correction and those who think he’s reverse-arguing toward a conclusion he already holds. Either way, the post is going to be cited a lot the next time someone says “don’t worry, scaling will hit a wall.”

Image-Blaster: 3D Environments, SFX, and Meshes From a Single Image

164 points · github.com

A weekend-project-flavored repo that takes a single 2D image and synthesizes a navigable 3D environment, sound effects, and meshes — combining depth estimation, generative texture work, and a small physics layer. It’s not production-grade, but the demos are striking, and it shows how far you can get by chaining a handful of open models together with good glue code.

The HN thread is half “this is the future of game asset pipelines” and half “the seams are obvious if you look.” Both can be true. What’s interesting is how much of the credit goes to the integration — the individual models have been around for months; the trick was wiring them into something a single developer could ship.

Show HN: Watch a Neural Net Learn to Play Snake

161 points · ppo.gradexp.xyz

A browser-based visualization that trains a PPO (Proximal Policy Optimization) agent to play Snake live in the tab, showing the policy network’s evolving decisions overlaid on the game grid. It’s the kind of “look how learning actually happens” demo that’s hard to make with a static blog post.

Education-minded folks in the comments are bookmarking it for students; ML practitioners are quibbling about hyperparameter defaults and reward shaping. The author has clearly thought about pacing — the agent goes from random flailing to competent in a couple of minutes, which is the right length for a curious visitor to actually watch.

Frontier AI Has Broken the Open CTF Format

156 points · kabir.au

A former competitive CTF player argues that modern frontier models have effectively destroyed the open Capture-the-Flag scene. Challenges that used to take teams of grad students an evening can now be one-shot by a sufficiently capable LLM, and organizers can’t keep up with task design.

The HN comments are mostly grim agreement, with a side discussion about whether private/closed CTFs and live in-person events can preserve what made the scene fun. It’s another data point in the slow, uncomfortable conversation about which “human-only” technical disciplines are next on the AI menu.