Published on August 11, 2026 at 18:06 CEST (UTC+2)
England set to be one of the first countries to eliminate hepatitis C (250 points by stevekemp)
The BBC reports that England has reached the 80% treatment target for known hepatitis C cases, putting it on course to be among the first countries to eliminate the disease. Deaths from the virus, which attacks the liver, have fallen by 36% in the last decade, just short of the 2030 goal. Short-course antiviral tablets are the key tool driving this progress. It is a significant public-health milestone enabled by widespread testing, treatment, and data-driven monitoring.
Apple Silicon and macOS VMs: 11–16× Faster LLM Inference with Llama.cpp (65 points by frabonacci)
A post from Cua, the open-source macOS virtualization project, describes a compatibility layer that lets macOS virtual machines access Apple’s newer Metal GPU paths. By connecting Virtualization.framework to Cua’s local computer-use stack, llama.cpp inside macOS VMs can achieve 11–16× faster LLM inference than before. This makes virtualized Apple Silicon far more viable for GPU-heavy AI workloads. The work points toward near-native performance for AI agents running inside macOS virtual machines.
Show HN: Git-knife – edit commit messages, authors, and dates like a spreadsheet (31 points by YonathanTesfaye)
git-knife is a desktop GUI for editing git history like a spreadsheet. It lets developers modify commit messages, author names and emails, author dates, and committer dates in a visual table. The author argues that existing GUIs treat dates as immutable and don’t expose committer identity or arbitrary commit metadata, while powerful command-line tools lack a usable interface. git-knife fills that gap by wrapping git commands rather than reimplementing git’s internals.
Stealing Reasoning Traces from Proprietary LLM APIs (98 points by quantumgarbage)
“Stolen Thoughts” is a research site demonstrating that hidden reasoning traces from proprietary LLM APIs can be decoded. The researchers collected 6,708 public agent trajectories from GitHub and Hugging Face, reconstructed over 315,000 encrypted reasoning blocks, and recovered API keys, passwords, access tokens, and personal emails from them. Some secrets appeared only inside the hidden reasoning and nowhere in the visible session. This challenges the assumption that encrypted chain-of-thought is safe from extraction and shows that agent logs can become serious data-leak vectors.
Manus will return to operating as an independent company (25 points by thm)
Manus, the AI agent company, announced it will soon return to operating as an independent company. To comply with regulatory requirements in certain jurisdictions, data generated by some users on or after December 29, 2025, will be deleted during a two-day window in August 2026. Affected users can back up their data beforehand and restore it afterward using tools Manus provides. Unaffected users can continue using Manus normally, but the episode highlights the data-governance complexity facing AI startups.
Launch HN: Keet (YC S24) – An app to create video courses on anything (9 points by zackashen)
Keet is a YC-backed app that automatically creates video courses on any topic. It generates structured roadmaps with modules, lessons, explainer videos, quizzes, and interactive games, all in a Duolingo-style format. Users can learn subjects like topology, history, physics, or engineering in bite-sized pieces. The product is an example of AI-generated multimodal education becoming more personalized and gamified.
Nvidia's Risky Business (124 points by jonbaer)
Ben Thompson’s Stratechery article opens with Jay Cooke financing the Northern Pacific Railway in 1870, using land grants and aggressive commissions to fuel a massive railroad boom. He uses that historical parallel to analyze Nvidia’s current position, where enormous capital flows are driving the AI infrastructure build-out. Thompson argues that Nvidia’s dominance is tied to a capex cycle that carries serious financial and systemic risk. The piece is a cautionary story about concentrated technology winners and speculative excess.
France to ban unsolicited telemarketing calls (787 points by aziaziazi)
France is banning unsolicited telemarketing calls starting August 11. The new regulation requires explicit consumer consent for marketing calls and is designed to cut down on nuisance calls, scams, and aggressive cold-calling. Companies that violate the ban can face fines. It is part of a broader European push to strengthen consumer privacy and digital protections.
As AI eats the web, the internet’s collective memory is disappearing (635 points by awnird)
This Walrus essay argues that the rise of AI is erasing the internet’s collective memory. AI-generated content and AI-composed search answers push original sources out of reach, reducing traffic to archives, primary documents, and independent websites. The piece warns that without deliberate preservation and provenance practices, future AI systems will train on a web full of synthetic, self-referential content. It frames the decline of Google Search as both a symptom and a cause of this amnesia.
H3-metal – Native MiniMax-H3 inference for Apple Silicon (381 points by swyx)
antirez’s h3.c is a compact native C inference engine for MiniMax H3 that runs on Apple Silicon via Metal. The repository includes Metal shaders, tokenizers, text and audio encoders, multimodal support, and a CLI, making it a full-featured local inference stack. Because it is dependency-free and performant, it opens up hybrid state-space/attention models to developers on Macs. The project is another sign of the open-source community building efficient local alternatives to proprietary AI APIs.
Local inference on Apple Silicon is becoming a serious performance tier. Article 2 shows an 11–16× speedup for llama.cpp in macOS VMs, and article 10 delivers native Metal inference for MiniMax H3 on Macs. Why it matters: Local inference reduces API costs, improves privacy, and enables offline AI agents, making consumer hardware a real platform for AI workloads. Takeaway: AI developers should optimize for Metal-native runtimes and treat local execution as a first-class deployment target, not just a fallback.
Hidden reasoning traces are a new and dangerous privacy leak. Article 4 demonstrates that encrypted chain-of-thought from OpenAI, Anthropic, and Google models can be decoded, revealing API keys, passwords, and personal data inside agent logs. Why it matters: Reasoning blocks contain not just model thoughts but user context, secrets, and intermediate decisions; “hidden” does not mean inaccessible. Takeaway: LLM providers and agent platforms should apply data minimization to reasoning traces, allow local reasoning, audit logs before publication, and treat reasoning-side leakage as a critical security risk.
AI agent data governance is becoming a regulatory and operational problem. Article 5’s Manus data-deletion announcement and article 4’s public agent trajectories show that agent platforms generate large amounts of sensitive user data that can be exposed or affected by corporate changes. Why it matters: AI agents produce continuous records of user activity, making compliance with data-deletion, privacy, and cross-jurisdiction rules much harder. Takeaway: Build backup/restore, consent, redaction, and data-export features into agent products from day one, and audit training/release pipelines for embedded reasoning blocks.
The web’s original content is being eroded by AI-generated search and media. Article 9 warns that AI answer engines and synthetic pages push primary sources out of reach, collapsing internet archives and collective memory
Analysis generated by deepseek-reasoner