Tech News

A developer has recreated Apple Vision Pro in Unity, and it runs on both iPhone and Android phones. He also released printable 3D models of a headset that holds your phone, so you can use it much like a Vision Pro. Really an amazing project! 🫶🙌

Indie maker Jack Friks has collected 60+ motivational memes, pictures and quotes that kept him building and trying through a lot of failures before anything worked. It's a pick-me-up for anyone shipping side projects when things get hard, and he adds new images regularly.

Unsloth's GGUF quantizations let you run DeepSeek-V4.1-Flash-1017 locally with llama.cpp or Unsloth Desktop. The model is MIT-licensed and has a 1M-token context. The smallest 1-bit quant needs a desktop with at least 128 GB of RAM (GPU VRAM counts toward it) and about 235 GB free on a fast NVMe SSD. A GPU is optional, and 64 GB of RAM works too, only slower.

A fan project runs the PC version of Need for Speed: Underground 2 (2004) in Chrome and Firefox. It's built on NFSU2Recomp, a static recompiler that translates the game's x86 code to C. On top of that it adds WebGPU graphics, Web Audio, up to 120 fps in widescreen, support for controllers, wheels and touch, free roam and online races with friends. It isn't affiliated with Electronic Arts.

Fernando Borretti argues that thinking, coding and research only last inside a community of other people, which gives you shared work to build on and the motivation of being useful to others. AI makes contributions like blog posts, open-source libraries and new programming languages superfluous, so those communities dry up. His examples are software engineering and now mathematics, and he expects individual intellectual work to fade with them.

Login code screens split a six-digit code into six boxes held together by JavaScript, which breaks paste, Backspace, SMS and email autofill, password managers and screen readers. The article shows these failures in live demos, including in react-otp-input, which hardcodes autocomplete off, and input-otp, which needs 767 lines to hide a single input. The fix is one plain input with autocomplete="one-time-code", inputmode="numeric", maxlength and a pattern, and no script.

Delta Inference Engine, a Windows port of the Linux engine Xolotl, runs all 552B parameters of DeepSeek-V4.1-Flash-1017 with 2-bit experts on a single consumer PC. It serves an OpenAI- and Anthropic-compatible API with up to 1M tokens of context. It needs an NVIDIA RTX card with 12 GB+ VRAM, 128 GB+ RAM and about 385 GB on a fast NVMe SSD. On an RTX 3090 with 128 GB RAM it writes about 33 tokens per second.

REA is an open-source MCP server and CLI that lets AI coding agents like Claude Code, Codex and Cursor inspect apps without their source code. It works with native binaries through Hopper, Ghidra or IDA, and with JavaScript and Electron apps, .NET assemblies, Android apps and websites. The agent can then explain how a feature works, show the evidence behind each conclusion and build a similar version for your project. Analysis runs locally, and setup is a single npx command.

Turkish software engineer Bedirhan Yildirim used AI coding agents and free Polish public data to build an explorable 3D replica of Warsaw in two days. The free online map has almost 200,000 buildings, trams that follow real timetables, live weather and 45,000 streetlamps that light up at night, plus a flight simulator. He says AI did 90% of the job, and the last 10% of human work keeps it from being "AI slop".

The ArtCraft team is building sixteen free, open-source desktop apps from scratch in Rust, with familiar layouts and shortcuts. They cover image editing, vector illustration, video, raw photos, PDFs, motion graphics, page layout, interface design, 3D texture painting, word processing, audio, slides, CAD and spreadsheets. They run natively on macOS, Windows and Linux without Electron, most also run in the browser through WebAssembly, and every app can be driven by AI agents through a CLI or an MCP server. All of them are still in early alpha or in development.

Rocky Linux founder Gregory Kurtzer had a single OpenAI Codex task (GPT-5.6 Sol) write Ginux in eleven days. The result is about 139,000 lines of Rust and test code for an x86-64 kernel that boots from UEFI without Linux and runs unmodified Linux programs, including Rocky Linux 10 with systemd. It's an experimental research project tested in QEMU, not a replacement for Linux, and it comes with a technical report on how he directed the agent.

A new arXiv paper, “The Little Book of Generative AI Foundations: An Intuitive Mathematical Primer,” offers an accessible mathematical introduction to the core principles underlying generative AI. The work is framed as an intuitive primer aimed at helping readers build foundational understanding of generative AI concepts.

A new Show HN project called HackerTV curates YouTube videos frequently recommended on Hacker News and broadcasts them as 24/7 TV channels. Everyone sees the same schedule, creating a shared linear TV experience around HN favorites.

macOS 27’s removal of CoreStorage support eliminates a native path for encrypted HFS+ backups to HDDs, with users noting that APFS is not well suited to spinning disks. Suggested workarounds include encrypted Time Machine to a network location or third-party tools like SuperDuper and Carbon Copy Cloner, though APFS on HDDs can make Time Machine painfully slow. The discussion also reflects broader frustration over Apple dropping legacy technologies and macOS becoming more iOS-like and restrictive.

U.S. tech giants’ hidden AI-related debt swelled 90% to $3.2 trillion in six months, nearly doubling as off-balance-sheet obligations ballooned. The figure highlights growing financial exposure tied to AI commitments that isn’t fully reflected on corporate balance sheets.

Japan’s Jimi Halloween is back for 2026, celebrating costumes so ordinary they require an explanation. The annual “Mundane Halloween” event highlights everyday, low-key outfits rather than spooky or glamorous ones. It remains a beloved Japanese tradition built on playful, relatable humor.

A reader discussion around The Internals of PostgreSQL notes renewed interest in PostgreSQL’s SSI implementation following the latest Jepsen report on ParticleDB. For readers skimming the document, the HOT update path is recommended as the key section, with zHeap implementation notes and TOAST material providing a useful mental model of what PostgreSQL does and does not do.

The LLM boom relied on a pretraining data commons built for other reasons, but posttraining consumes tasks rather than documents, and no equivalent commons exists. This gap leaves the open models community without a shared foundation for posttraining data.

Kairos is an open-source platform for real-time event intelligence and fraud and risk decisioning focused on financial behavior. It aims to help teams make faster risk and fraud decisions from streaming financial events.

A new article argues that most Pratt parsing tutorials spend nearly all their time on the algorithm while neglecting the underlying properties that actually matter. It contends that this leaves Pratt parsing treated as a black box, with little explanation of the principles that make it useful.

The article examines reverse honeypots—curiosity traps that reward close attention and help attract the right people. It suggests designing for curiosity can filter for engaged audiences and foster community rather than passive traffic.

The Hacker News community discussed “Union Types and Option Types,” an article exploring the tradeoffs between these type-system approaches and their use in CopperLang. Readers weighed how each models missing values and errors, and what those choices mean for language design.

Microsoft released a Windows 11 update that fixes a longstanding annoyance and adds a cloud recovery option for PCs that won’t boot. The changes arrive as the company reports declining PC revenue.

Safari 27.3 beta introduces Adaptive Dark Mode, a new feature that automatically adjusts the browser’s dark appearance. The change was announced as part of the beta release.

Revolut has moved roughly 2,000 engineers onto self-hosted GLM for coding, making Claude Code and Codex exceptions rather than defaults. According to AI ops lead David Abitbol, the setup uses 56 Nvidia B200s at a blended 30 cents per million tokens—about a tenth of what Revolut previously paid Anthropic and OpenAI.

A reported issue alleges that Postman silently auto-installs AI plugins and attempts to share all API data, raising privacy and security concerns. Users described unexpected changes to their coding agents after opening Postman on macOS. The report has been filed in the issue tracker, where duplicate submissions are expected to be closed.

A new paper argues that LLMs can natively read and write neural network weights, enabling “weight-space programming.” The news piece reports that Qwen 3.8 27B running on a laptop can also interpret model weights, suggesting the capability may be accessible on local hardware.

A recent post claiming Astra can read and write weights to improve Qwen is misleading. The experiments involving reading raw weights and those that fix Qwen rely on completely different techniques, so they do not support the implied connection.

Fujifilm plans to increase its magnetic tape production fivefold to meet growing archiving demand driven by AI. Magnetic tape remains attractive for archival storage because it is low-cost and offers offline security.

A car owner jailbroke their vehicle’s navigation system just to change the clock, highlighting how locked-down infotainment software can block basic settings. They note that the workaround shouldn’t be necessary but found the process enjoyable.