Archive entry
HN Weekly — 2026-08-16
- Firefox is now the last major browser that still supports uBlock Origin
Firefox announced it will keep supporting the uBlock Origin ad blocker even as Microsoft Edge and other Chromium-based browsers discontinue support under Manifest V3 [1.1]. The new rules prevent extensions from inspecting and blocking ads effectively. This leaves Firefox as the only major non-Chromium browser providing full support for the extension.
HN reaction
Commenters appreciate Firefox for curating and vetting trusted extensions like uBlock Origin. While some debate the trade-offs of Google's security changes, many note that modern web browsing without strong ad blockers has become nearly unusable.
(Source) - Qwen 3.8 27B
The Qwen team released Qwen 3.8 27B, a dense vision-language model available in an FP8-quantized format. It offers improved capabilities in coding, research, and long-horizon tasks, with native support for images and long videos. The model runs locally on consumer hardware and supports extended context lengths.
HN reaction
Users report solid results on private benchmarks and local coding tasks, praising its detailed reasoning steps. However, several note high memory usage during long tasks, and others share playful tests rendering complex graphics.
(Source) - Tracking down the 16-year-old WAL-reset SQLite bug
Tailscale investigated sporadic database corruption across its control plane and discovered a rare data race in SQLite. The flaw occurred during aggressive checkpointing in write-ahead logging mode and had existed for sixteen years. Working with SQLite maintainers, the team built custom diagnostic tools to isolate and resolve the issue.
HN reaction
Commenters praise Tailscale for funding open-source tooling and taking software correctness seriously. Several readers discuss the difficulty of finding rare race conditions and express surprise at a long-standing flaw in such a well-tested database.
(Source) - Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
Meta introduced Muse Glimmer, an open-weight 30-billion parameter model designed for local agent workflows. It uses quantization and speculative decoding to run efficiently on single consumer GPUs. The model handles multi-step reasoning, tool execution, and multimodal inputs without needing continuous cloud access.
HN reaction
Readers discuss how smaller, local models might challenge expensive cloud data centers. Many welcome Meta's open-weights strategy, though some debate whether dense 30-billion parameter models can outpace competing models on complex tasks.
(Source) - GLM-5.3: Frontier coding with emergent cyber capabilities
Z.ai announced GLM-5.3, a coding model updated through large-scale reinforcement learning and post-training. The new release shows significant gains on software engineering benchmarks and autonomous vulnerability discovery. During real-world testing, it identified thousands of security flaws across numerous open-source software repositories.
HN reaction
Commenters express excitement over the model's strong red-teaming skills and its ability to uncover real security vulnerabilities. Others highlight the rapid progress of open-weight Chinese models compared to proprietary American offerings.
(Source) - France to ban unsolicited telemarketing calls
According to its title, this article reports that France plans to ban unsolicited telemarketing calls starting in August 2026. The linked page could not be accessed directly to provide additional details beyond this planned ban on unwanted commercial calls.
HN reaction
Commenters strongly support the ban, sharing personal frustrations with constant spam calls. Many discuss how other countries handle unwanted calls through national opt-out lists, while others argue for technical caller authentication to stop number spoofing.
(Source) - DeepSeek V4 Pro 0813
OpenRouter published pricing, performance metrics, and access details for DeepSeek V4 Pro 0813. The model is a large mixture-of-experts architecture offering a one-million token context window. It supports tool calling and structured outputs at competitive developer rates.
HN reaction
Several users feel this release offers only modest improvements over previous versions. Others test its practical coding and deployment capabilities on real tasks, noting mixed results compared to competing options.
(Source) - Why does Opus 5 feel worse to work with?
The author argues that Opus 5 feels worse for coding than earlier versions despite higher benchmark scores. Because benchmarks reward aggressive guessing, the model makes unsupported assumptions rather than pausing for clarification. This behavior creates extra friction for developers who need reliable communication over autonomous guesses.
HN reaction
Many commenters agree, noting that the model generates overly verbose, stylized prose and unnecessary code comments. Others share that they switched back to earlier releases or alternative tools to avoid constant supervision.
(Source) - Every Fucking Website (2020)
The webpage presents a satirical, interactive recreation of modern web design annoyances. It bombards visitors with cookie consent banners, newsletter pop-ups, promotional discounts, and unwanted chatbot widgets. The parody highlights how cluttered and hostile the everyday browsing experience has become.
HN reaction
Readers jokingly suggest adding more intrusive elements, such as autoplaying videos and slow loaders. Others reflect on how aggressive dark patterns and annoying pop-ups persist because they reliably boost conversions and revenue.
(Source) - How I use LLMs to learn complex topics
The author describes using language models to understand difficult technical subjects by generating interactive, game-like simulations. Instead of reading dry text, the workflow builds and verifies a knowledge base before rendering animated processes. The visual approach helps users follow complex steps and retain information better.
HN reaction
Some commenters find AI-generated prose too exhausting and prefer traditional books for serious learning. Others question whether self-checking models truly eliminate errors, while a few describe success using targeted prompts alongside primary texts.
(Source)