Senate blocks Sh88bn Nairobi deal, Alibaba 2.4T Qwen open-weight

The nonprofit, profiled by TechCrunch today, aims to build what it calls the World Wide Web of AI free for everyone. Its first device runs entirely on-device no cloud dependency, no per-query billing and supports languages including Swahili, Hindi, Bengali, Tamil, and Quechua. The hardware is open-source, letting local communities manufacture and customize it. The project addresses a real gap: most AI systems serve the top 20 languages well and ignore the rest. For developers, this is a rare opportunity to build on a genuinely open stack trained for non-English contexts. The device is already being tested in India and East Africa.

Built through Mozilla AI's Builders in Residence program, transcribe.cpp unifies 16 ASR model families and over 60 variants under one native C and C++ library with GPU acceleration. Unlike cloud transcription APIs, it runs fully on-device no data leaves the machine. The HN thread (602 points) highlighted its speed and simplicity compared to juggling whisper.cpp and ONNX separately. It supports streaming and batch transcription across Windows, Linux, macOS, and ARM. For anyone building voice interfaces, call recording analysis, or accessibility tools, this kills the friction of choosing and maintaining multiple inference backends.

The vulnerability, tracked as CVE-2026-58644, is a critical deserialization flaw allowing remote unauthenticated code execution on on-premises SharePoint Server. CISA added it to the Known Exploited Vulnerabilities catalog on July 16, giving federal civilian agencies until July 19 to apply fixes. Microsoft patched it in the record-breaking July 2026 Patch Tuesday that addressed 570 flaws. For any organization running on-premises SharePoint (Subscription Edition, 2019, or 2016), this is a patch-now situation the exploit allows attackers to deploy web shells, move laterally, and establish persistent access. If your team deferred the July updates, today's the deadline.

Described as Better and Cheaper Than IPTV, Castor uses Chrome headless for extraction, ffmpeg for transcoding, and serves content over your local network. It runs as a single Docker container with no external dependencies. The project hit 247 points on Hacker News, with the thread praising its simplicity compared to the complexity of Plex or Jellyfin setups for live content. For anyone tired of unreliable IPTV subscriptions or complicated media server stacks, Castor offers a straightforward self-hosted alternative. It scans for local or streaming media sources and makes them available through a web interface no monthly fees, no shady providers.

Weinberger, a career software developer, built Jamcorder a standalone device that constantly records MIDI data from digital pianos and synths, storing up to 25,000 hours. He sold 2,500 units at $99 each, and the business is now self-sustaining. His key insight: hardware is hard if you design custom PCBs and injection-molded enclosures. If you use off-the-shelf components, keep the design simple, and maintain healthy margins, it is actually more predictable than complex software. For developers who have considered physical products but were scared off by conventional wisdom, the post is a useful reality check. The device detail that sealed the product: it never asks you to hit record.