Top of Hacker News

Citizen Lab traced an Israeli ad platform, Discover-Tech, that won 442 million of 3.77 billion impressions it bid on in four months while spending only about $109,000. Advertiser conversions were essentially zero: the ads carried script tags that can hand code to a phone.

An adtech engineer with 13 years in the industry told the thread that vetting that traffic is "virtually impossible". A commenter who once sat in telco-versus-police meetings had heard that line before, and asked why billing is not impossible too.

OpenAI has published hundreds of results in number theory, theoretical computer science and logic, many on problems that resisted the best mathematicians for decades. Derek Thompson's essay says the mathematicians weep because a shared conversation across generations just got cut short.

On HN, one commenter was "kind of disgusted by the tears" and blamed decades of chasing results and money. Another checked the pay at their own university: two associate professors on $80-85K, after postdocs that push tenure-track starts into the mid-30s.

Meta released Muse Spark Open on October 15 as downloadable weights: 816 billion parameters, 42 billion active per token, a 1M-token context window, under a Muse Model License with limited commercial terms. The benchmark section is empty, and the full evaluation report is "available soon".

Epic Games Launcher V2 preview, which Epic built with AI assistance, starts in about 5.47 seconds against Steam's 5.23 after Valve's quick patch, though Epic still restores from the tray faster (0.36s vs 0.42s).

The stopwatch did not survive HN. A commenter who timed both launchers got 10-15 seconds for a cold Steam boot, and others replied that the faster client had already been sitting in Steam's beta channel.

Schaller used Claude Code to build drivers for a Plustek film scanner and fix a ThinkPad webcam, using Windows traces as the answer key. In low light, the webcam's frame brightness rose from 31.1 to 90.9.

In the thread, a guitar-pedal maker said they decompiled three rivals' award-winning effects in 72 hours and that their industry's moat had vanished. A hardware engineer replied that the moat was never there: if the firmware is readable, it was always open.

Current Affairs adopted one of publishing's strictest AI policies as part of its union contract, arguing that tools built to replace workers must be resisted. The article's example: 100 AI posts in four minutes at almost no cost versus 20 human posts for $5,000.

On HN, a commenter posted a parable of fifty diggers with shovels versus five with machines to argue no one will care who wrote what. The reply that stuck: dirt is paid for by the hole, but writing is only worth anything if someone chooses to read it.

Goodfire's Nick Cammarata revisits the famous "cat and car" neuron and finds it responds to pointed versus round shapes, not to a list of objects. His claim: much of the polysemanticity (one neuron firing for several things) came from fixable problems, not from superposition.

Withings has moved automatic syncing to Apple Health and Google Health Connect behind its Withings+ subscription. The rule keys off the account creation date: accounts created on or after October 13, 2026 need the plan, while accounts created on or before October 12 keep syncing for free, even if the owner buys new hardware later.

A Withings reply marked as official confirmed it: after October 13 «automatic sync to third-party apps is part of Withings+». The company says the subscription «funds the services» built around user data.

Free accounts keep full access inside the Withings app and can still export their history manually as CSV or PDF. What they lose is the automatic handoff to everything else.

Contracts let a C++ function state what callers must guarantee and what it promises back. A violation can be ignored, observed, enforced or quick-enforced, chosen per build with no portable per-assertion switch, and the guide warns they are not input validation.

Walter Bright, who says he added contracts to Zortech C++ in the 1990s and to D later, wrote that they "just turned out to not be very useful". An embedded developer wants the opposite trade: checks on at the workstation, off on weak production CPUs.

The file was a Windows .exe, not video. Only its first 118 KB was a real program; the rest was padding, so scanners would give up. It decrypted a hidden payload and ran it in memory: an infostealer after browser passwords, crypto wallets, SSH and cloud credentials.

The author's fix for *arr users: exclude .exe, .scr, .lnk, .bat and .cmd in qBittorrent.

Gambhir, a theoretical particle physicist, writes as someone who thinks it is inefficient to explain or check a result when the same resources can generate another, and calls the output a repo of potentially true facts nobody will read. Tao hosts it.

On HN, the sharpest hit ran the Library of Babel simulator: random pages are gibberish, while AI proofs are credible and frequently correct. A reply pointed out that the essay says exactly that, and that "look correct" is the joke.

DEsk is a beta Wayland desktop in the style of iOS 6 skeuomorphism: apps are filed in binders on a bookshelf, notifications arrive in a post box, and the lock screen is a front door you unlock with a key. The room is ray-traced and lit by one desk lamp.

Zach Landes tested "decision models", small cheap models that output a choice plus a probability, against OpenAI's Decisions API model GPT-6 Luna, which never chose its server over people. Reordering the track descriptions alone flipped Jev's answer in 370 of 2,620 pairs.