Tech news, summarized and translated.
A Claude Code plugin recovers Kindle highlights that Amazon's undocumented export limit truncates or hides entirely, extracting 2,432 highlights from four real books with 815 previously blocked.
HBF promises cheap, high-capacity flash on-package, but its block-oriented access model forces a software rewrite that may not be worth it.
Nvidia wants CUDA on RISC-V, but the requirements are steep: RVA23, ACPI, PCIe coherency, and more. Most existing RISC-V boards need not apply.
A look at why XMPP's open standards approach matters for digital sovereignty, and why open-source walled gardens like Signal and Matrix fall short.
A developer built a real-time AI gaming companion for Skyrim that takes complex commands, fights alongside you, and evolves its personality — all with local-first processing and minimal latency.
AlpinDale's gpt2.cmake implements a GPT-2 model entirely in CMake, using Q16.16 fixed-point arithmetic. It's a wild proof-of-concept that pushes the build tool far beyond its intended use.
A developer discovers Google Workspace's sign-up page rejects valid domains like web.example.com because of a hardcoded regex list of 'email providers' — and the fix is a frontend-only bypass.
An agent harness is the software shell that turns a raw AI model into a tool you can actually control — and it's the key to breaking free from AI lab lock-in.
A deep-dive experiment shows that swapping attention backends in vLLM can flip the model's top-1 token choices on real agentic workloads, even with identical weights and hardware. The culprit: implementation-specific math, not the model.
The AT Protocol's biggest update yet introduces Spaces, a new primitive for gated, non-public data with access control (not encryption). Alpha code, SDKs, and a hosted PDS are live for experimentation.
Nari Labs claims a Qwen3-TTS 1.7B CustomVoice serving stack hits 10 RPS with sub-50 ms p95 time-to-first-audio on a single H100, beating vLLM-Omni, SGLang-Omni, VoxServe, and M* after tuning. The key: a unified scheduler for Talker, Code Predictor, and Codec.
A developer runs Photoshop on a Raspberry Pi RP2350 chip that costs 60 pence, emulating a 1990s Mac. It's a reminder that modern computing power is absurdly cheap, and a nudge to question the bloat we accept.