Published on August 23, 2026 at 06:01 CEST (UTC+2)
Scrap (2006) (336 points by tosh)
Moxie Marlinspike shares an old journal entry from around 2006 titled “Scrap.” He recalls moving from the West Coast to Pittsburgh with romanticized expectations of winter—snowballs, sledding, and cozy snow-covered streets. Instead, he finds himself in a house without working utilities, standing in a cold basement trying to install gas lines. The piece is a reflective, personal story about naivety meeting reality.
Why your local LLM feels dumber than it is (217 points by felineflock)
A Level1Techs forum post explains why locally-running LLMs often feel worse than official benchmarks suggest. The author argues that most users run quantized or poorly integrated versions of models on very different hardware and software from the reference environment. The post runs experiments to demonstrate how inference implementation choices—sampling, context handling, quantization—affect perceived quality. It concludes that local implementations are all slightly broken, but that’s an engineering challenge, not necessarily a model failure.
NanoGPT Speedrun Frontier (62 points by stared)
Prime Intellect’s “NanoGPT Speedrun Frontier” tracks AI coding agents as they attempt a NanoGPT training task. The dashboard shows models like Fable 5, Opus 5, and Kimi K3 racing to close the gap to the human record over days of “agent time.” Several agents are shown achieving large percentages of the human-record benchmark. It’s a demonstration that AI agents are becoming capable of doing serious, open-ended ML engineering work.
ElevenLabs, TwelveLabs, ThirteenLabs (335 points by jemoka)
The author, after hearing about ElevenLabs and TwelveLabs, jokingly searched for “ThirteenLabs” and found yet another AI startup. They kept going and produced an annotated chart mapping numbers 0–99 to real “Number+Labs” companies, marking which ones appear AI-related. The post is a commentary on how AI startups have converged on the same naming convention and how hard it is to tell companies apart. It’s a satirical look at AI industry hype and branding.
Hister – A private, full content search index that you control (263 points by auraham)
Hister is a privacy-focused tool that turns the pages you visit and files you keep into a private, full-text search index. It stores extracted content locally or on a server you control and shows readable previews alongside search results. The project advertises no telemetry, no external requests, free software under AGPLv3, and self-hosting. It’s pitched as a way to keep useful knowledge findable without relying on centralized search or cloud data collection.
I set a trap for a book-marketing scammer (2025) (16 points by rznicolet)
A traditionally published sci-fi author describes receiving more than fifty scam book-marketing pitches in thirty-three
typ.ing (206 points by bookofjoe)
Analysis generated by deepseek-reasoner