HN Digest | The Best of Hacker News

When did Google get so weird?

Google's AI Overview responded to a search for a 2010s basketball meme with empathetic relationship advice, treating the query "hes never coming over dario" as a personal crisis rather than a search string.

💬 Engineers note that Google's AI processes queries as plaintext statements rather than string matches, leading to frequent misinterpretations.
→ discussion

1333 points · 733 comments · #daily · @HNDigestEn

Unsealed Briefs in Authors’ Case v. Microsoft/OpenAI

Unsealed court documents reveal OpenAI researchers knowingly trained models on copyrighted books from Libgen, explicitly worrying about the public relations fallout if the practice was exposed on Hacker News.

💬 Several users highlighted the irony of an OpenAI researcher specifically fearing negative exposure on Hacker News.
→ discussion

579 points · 545 comments · #daily · @HNDigestEn

Breaking Up with Google Play: Why Conversations Is Now Free

After ten years of paid distribution, the developer of federated XMPP client Conversations has made the app free on Google Play, citing poor support, delayed updates, and secure grant funding through 2029.

💬 Several users criticize Google's terrible developer support, while others note the Play Store successfully funded the project for years.
→ discussion

664 points · 271 comments · #daily · @HNDigestEn

Dutch governments builds alternative for Microsoft based on NixOS

The Dutch DAWO project uses modular, reproducible building blocks like DAWO-NixOS to establish digital autonomy and verifiable cloud infrastructure for public administration.

💬 Users highlight similar European initiatives like Germany's openDesk and France's La Suite, while noting potential Codeberg policy conflicts.
→ discussion

891 points · 526 comments · #daily · @HNDigestEn

California is chasing wealth that has feet

California's land is worth $8.14 trillion, meaning a 0.25% land value tax would raise the same $20 billion annually as the proposed billionaire wealth tax without risking wealth flight.

💬 Skepticism exists over claims that land taxes cannot be passed to renters, alongside criticism regarding gentrification.
→ discussion

234 points · 641 comments · #daily · @HNDigestEn

Meta VR Glasses

Priced at $1,300, the new headset features Micro-OLED displays with 2412 × 2288 pixels per eye, a 70° field of view, and a Snapdragon Reality Elite chipset with 12GB of RAM.

💬 Skepticism remains over VR's long-term appeal, though some praise the high-end display specs for media consumption.
→ discussion

475 points · 448 comments · #daily · @HNDigestEn

🔥 F-Droid 2.0

Rewritten in Kotlin Compose, this first major update in ten years utilizes Android's pre-approval API for seamless background updates without requiring the F-Droid Privileged Extension.

💬 Several users welcome the redesign after previously switching to alternative clients like Droid-ify due to the old interface.
→ discussion

524 points · 155 comments · #breaking · @HNDigestEn

Claude discovers a novel enzyme system with CRISPR-like repeats

Anthropic's team used 950 Claude agents consuming 210 million tokens over 21 hours to discover array-associated reverse transcriptases (ART), a novel enzyme system with CRISPR-like repeating DNA patterns in bacteriophages.

💬 Skepticism emerged regarding the lack of published biological results and potential training data contamination from user-submitted papers.
→ discussion

643 points · 659 comments · #daily · @HNDigestEn

Microsoft killed FoxPro in 2007. Anyway, here's FoxPro revived

FoxDev Studio is a 64-bit Visual FoxPro runtime and IDE written in Rust and WebAssembly that removes the 2GB table limit and uses React to render UI forms.

💬 Users share nostalgia for FoxPro's high productivity while warning about the dangers of hosting DBF files on network drives.
→ discussion

443 points · 243 comments · #daily · @HNDigestEn

GPT-6 Sol and Luna

The new GPT-6 Luna model costs just $0.10 per million input tokens and $0.50 per million output tokens, representing a 50% price cut compared to previous versions.

💬 Users note the release timing directly targets Anthropic's launch, while praising Luna's highly competitive token pricing.
→ discussion

1516 points · 724 comments · #daily · @HNDigestEn

I said no and Apple said yes

A macOS upgrade downloaded a 22.28 GB model file and re-enabled previously disabled Apple Intelligence features, leaving unkillable background processes and no direct toggle to turn them off.

💬 Several users expressed frustration with Apple's persistent nudges, while others criticized the author's emotional tone.
→ discussion

748 points · 600 comments · #daily · @HNDigestEn

🔥 Claude Opus 5.5

Claude Opus 5.5 costs 40% less to run and runs 30% faster than Opus 5, with input tokens priced at $4 per million and cache reads cut by 60% to $0.20 per million.

💬 Users welcomed promised improvements to the model's writing style and questioned how it positions against Claude Fable.
→ discussion

598 points · 535 comments · #breaking · @HNDigestEn

MiMo v2.6

Priced under $0.15 per million input tokens, this open-weight model release demonstrates diverse task capabilities like DAW usage and strong performance in multi-agent game-playing.

💬 Users praise the lab's honest benchmarking and note MiMo's superior performance in multi-agent game-playing scenarios.
→ discussion

890 points · 396 comments · #daily · @HNDigestEn

What happened to the Snowden archive

The Guardian published only 30 of the 58,000 documents (0.05%) it held before destroying its hard drives under GCHQ supervision and joining a government advisory committee.

💬 Several users suggest the prosecution of Julian Assange successfully deterred journalists from publishing further leaks.
→ discussion

654 points · 468 comments · #daily · @HNDigestEn

🔥 ZuckOff Know when a camera is in the room

The app detects nearby smart glasses by scanning for specific Bluetooth manufacturer signatures like 0x0D53 for Luxottica and 0x03C2 for Snap Spectacles entirely offline.

💬 Several users express concern that manufacturers will eventually patch or encrypt these Bluetooth beacons to prevent detection.
→ discussion

497 points · 209 comments · #breaking · @HNDigestEn

ChatGPT now knows what you do on other websites via ad collector

OpenAI's ad collector sets a SameSite=None cookie called __obi that links ChatGPT accounts to off-site browsing data, including hashed emails and cleartext postal codes scraped from advertiser sites.

💬 Users express deep concern over standard adtech tracking being applied to highly sensitive AI chat histories.
→ discussion

719 points · 382 comments · #daily · @HNDigestEn

English: A vs. An

Out of 32,455 analyzed English words, only 129 required exceptions to the standard vowel-character rule when determining whether to use "a" or "an" in the a_or_an() function.

💬 Several users note that "a" versus "an" usage can reveal a writer's regional accent, particularly with the letter H.
→ discussion

333 points · 450 comments · #daily · @HNDigestEn

AI-generated posters don’t have to be horrible

Specifying historical design movements like Bauhaus, Memphis, or Swiss Style in ChatGPT prompts bypasses the generic, overused default AI aesthetic to produce highly distinct event flyers.

💬 Several readers argue that despite stylistic variations, the generated posters still retain a highly recognizable, uniform "AI smell."
→ discussion

1576 points · 841 comments · #daily · @HNDigestEn

Hacker News Digest #30 · 6–10

6. I don't like passkeys

Passkeys protect against phishing but increase the risk of permanent lockout, device loss, and platform lock-in, while hardware keys are limited to storing 25 to 300 credentials.

💬 Several users agreed, complaining about aggressive platform prompts and the inability to easily memorize or backup credentials.
→ discussion

768 points · 757 comments

7. Nvidia announces native GPU programming in Rust

Nvidia introduced native GPU kernel compilation in Rust via cuda-oxide for SIMT programming and cutile-rs for tile-level programming, bringing compile-time safety guarantees to GPU code.

💬 Developers welcomed the ability to share structs between host and device, though some cautioned about nightly compiler requirements.
→ discussion

959 points · 402 comments

8. OpenAI bots knew about the RubyGems caching vulnerability

OpenAI's automated bots allegedly exploited a known RubyGems caching vulnerability to harvest authorization keys and execute arbitrary code on RubyDoc.info using malicious .yardopts files.

💬 Users argued the attacks were intentional, pointing to a lack of proper sandboxing and safety prompts.
→ discussion

510 points · 424 comments

9. Apple Reference Image: A New Approach for Verified Photography

Apple introduced Apple Reference Image on iPhone 18 Pro, using secure sensor-level cryptographic signing and Private Cloud Compute to generate verifiable, tamper-resistant digital negatives.

💬 Users pointed out a major loophole where photographing a high-resolution monitor displaying a forged image bypasses verification.
→ discussion

527 points · 352 comments

10. US Military had close call after using AI for hallucinated intelligence report

The US military aborted a boarding operation against a Chinese ship after discovering the intelligence report, which claimed the ship carried nuclear components, was entirely hallucinated by an AI chatbot.

💬 Users warned that removing human accountability from military targeting decisions will inevitably lead to catastrophic mistakes.
→ discussion

446 points · 334 comments

#digest30 · Subscribe: @HNDigestEn

HN Digest | The Best of Hacker News pinned «Hacker News Digest #30 12 Sep — 19 Sep 2026 1. Introducing System One Models and Jev TypeSafe AI released Jev, a System One model optimized for structured outputs that is up to 200x faster and 444x cheaper than traditional LLMs by using parallel sampling.…»