Hacker News
1France to ban unsolicited telemarketing calls
- France will ban unsolicited telemarketing calls starting August 11, 2026, requiring businesses to obtain prior consent before contacting consumers. Violators face fines up to €75,000 for individuals and €375,000 for companies per call. The law responds to complaints that three-quarters of French people receive unwanted sales calls weekly. Morocco expressed concern about potential job losses in its call center industry, which relies heavily on the French market.
Opt-out registries and enforcement challenges
Discussion of national "do not call" lists like Spain's Lista Robinson and Sweden's NIX-registret, which work well in some countries but fail in others like the US due to poor enforcement, with scammers ignoring regulations and some lists even being misused as telemarketing targets.Technical solutions for caller authentication
Debate over implementing technical measures like STIR/SHAKEN protocols to prevent caller ID spoofing, with participants discussing why phone networks still allow number spoofing and proposing whitelisting systems for legitimate callers like hospitals and utilities.US vs international spam call experiences
Americans report receiving significantly more spam calls than people in other countries, with one user noting their US number gets more daily scam calls than their overseas number gets in a year, attributed to weak enforcement and telco profitability.
2As AI eats the web, the internet’s collective memory is disappearing
- Google's AI search summaries are providing incorrect information, including wrong sunset times, highlighting broader issues with digital information infrastructure. The internet's "corpus is collapsing" due to link rot, AI errors, and content manipulation. Key resources like Wikipedia and Internet Archive face threats from reduced traffic and legal challenges. Some governments like France are building alternative search engines and digital infrastructure to achieve technological sovereignty, recognizing that information systems are strategic national assets requiring public oversight rather than reliance on private tech companies.
Google search quality decline
Users report Google searches have become "absolutely horrible" with amnesia-like behavior, missing recent content, and aggressive AI overviews. Many blame SEO spam, helpful content updates favoring brands, and smaller sites being de-indexed or buried.AI impact on web content creation
AI is flooding the internet with low-quality generated content while destroying incentives for human creators to publish original work. This creates a feedback loop where AI trains on its own output, degrading overall content quality.Search engine alternatives and solutions
Users discuss alternatives like Kagi, DuckDuckGo, and Brave Search as replacements for Google. Kagi receives particular praise for its paid model avoiding ad-driven priorities, though concerns exist about scalability and dependency on other APIs.
3The UK's war on anonymity has come to America
- An investigation revealed five foreign NGOs orchestrating a coordinated campaign to influence US lawmakers by promoting digital ID laws under the guise of "child safety." These British organizations successfully implemented similar surveillance laws in the UK that now target political dissidents, and are replicating this strategy across 21 US states using American proxies and lobbying efforts.
**Parental responsibility vs. government regulation**
Debate over whether parents should handle child protection through existing tools and better parenting, or if government intervention is needed due to collective action problems and parental limitations in controlling digital environments.**Surveillance state concerns**
Discussion about how "child safety" rhetoric is being used to justify mass surveillance systems that would end online anonymity, with comparisons to historical authoritarian regimes and concerns about future misuse.**Technical implementation alternatives**
Proposals for privacy-preserving solutions like OS-level "child mode" with government whitelists, zero-knowledge proofs for age verification, and device-based controls that don't require websites to collect personal data.
4Show HN: Needle2: 14MB agentic LLM for phones, wearables, smart home and robots
- Henry from Cactus announces Needle 2, a 14MB agentic LLM that runs in 28MB RAM for tool calls and structured extraction on low-power devices like phones, wearables, and IoT devices. The 45M parameter model achieves 500 tokens/sec on Raspberry Pi 5 and competes with much larger models while using significantly less power and computational resources.
Model accuracy and reliability issues
Users reported significant reasoning errors, including the model interpreting "make it dark" as turning lights on, misunderstanding door/light concepts, and incorrectly classifying sentiment. Despite being only 14MB, these accuracy issues raise questions about practical utility.Size optimization and use cases
Discussion centered on whether 14MB is optimal for micro-LLMs, with some suggesting larger sizes might improve performance while remaining suitable for edge devices. The model is designed for sub-$200 devices and specific tool calling rather than general conversation.Technical implementation and architecture
Questions about the custom transformer architecture, quantization format, and runtime engine. Comparisons to TensorFlow Lite and discussion of training methodology, with creators noting they trained from scratch rather than pruning existing models.
5England set to be one of the first countries to eliminate hepatitis C
- England is on track to become one of the first countries to eliminate hepatitis C, having already treated 80% of known cases and reduced deaths by 36% in the last decade. Over 100,000 people have been diagnosed and treated since 2015 using antiviral tablets that cure 95% of cases. The NHS offers free home testing kits to help find undiagnosed cases, particularly targeting people from certain Eastern European countries where infections may have occurred through medical procedures before 1991.
Hepatitis C screening and testing practices
Users share personal experiences with Hep C diagnosis, discussing how standard STI panels often miss Hep C testing. Regional differences emerge - some areas include comprehensive testing while others don't, particularly in US northeastern states during 2000s-2010s.Anti-vaccine sentiment and public health
Discussion centers on rising vaccine hesitancy in the US and globally, linking it to political movements, COVID vaccine messaging, and loss of public trust. Users debate causes including social media influence, immigrant populations, and historical cyclical patterns.UK healthcare system organization
Users explore why the Hep C elimination program is England-only rather than UK-wide, discussing devolved healthcare systems where each nation (England, Scotland, Wales, Northern Ireland) maintains independent NHS structures and health priorities.
6Stealing Reasoning Traces from Proprietary LLM APIs
- Researchers discovered a vulnerability in proprietary AI models from OpenAI, Anthropic, and Google where encrypted reasoning traces can be decoded by replaying them through weaker jailbroken models from the same provider, revealing the stronger model's hidden thought processes and sensitive information including API keys, passwords, and personal data.
Ownership and "stealing" of AI reasoning
Debate over whether accessing encrypted reasoning traces constitutes "theft" when users pay for tokens. Some argue it's recovery of paid content, others defend IP rights. Discussion includes broader questions about data ownership and whether LLM outputs can be owned.Technical implementation and security flaws
Analysis of how the attack works by replaying encrypted traces across different models and jailbreaking weaker ones. Users discuss cross-model compatibility, session management, and various technical workarounds like using "thinking tools" to bypass encryption.Model distillation and competitive implications
Discussion about whether this technique enables effective model distillation for competing labs, particularly Eastern companies. Debate over the effectiveness of distillation versus other ML advances, and whether accessing reasoning traces provides significant competitive advantage.
7H3-metal – Native MiniMax-H3 inference for Apple Silicon
- h3-metal is a native MiniMax-H3 inference project for Apple Silicon that enables prompt-to-video/audio generation with first/last-frame conditioning and reference image support. It provides various speed/quality presets and optimization options including Metal performance tuning, memory streaming, and interactive sessions for video generation.
Performance on Apple Silicon hardware
Users discuss running MiniMax H3 on Mac hardware, with M5 Pro taking over an hour for 9-second clips while RTX GPUs complete similar tasks in 2-3 minutes. Memory requirements around 40GB fit well on 64-128GB unified memory systems.Technical implementation and optimization
Discussion covers GGUF vs int8_convrot quantization methods, ComfyUI compatibility issues on Mac vs NVIDIA, and sparse attention potential speedups. Users share workflows and technical configurations.Hardware requirements and accessibility
Community discusses memory requirements (96GB vs 128GB), performance differences between Mac models, and options for users with limited hardware to run the model effectively.
8How Claude marks AI-generated content
- Anthropic has signed the EU AI Act's Code of Practice and is implementing machine-readable marks in Claude-generated content through embedded watermarks in text and signed provenance metadata in files. These marks help provide transparency about AI-generated content while having some limitations.
False positives and institutional consequences
Users worry that watermark detection could falsely flag human-written content as AI-generated, potentially ruining academic careers and student opportunities. The concern is that institutions with power to punish may not understand the limitations of detection technology.Technical implementation and effectiveness
Discussion centers on how watermarking works through token biasing during text generation, with concerns about impact on code quality and whether watermarks can survive editing or be easily defeated by paraphrasing tools.Legitimate mixed authorship workflows
Users describe collaborative writing processes where they provide initial content that Claude refines, arguing this represents meaningful human authorship despite triggering watermark detection, raising questions about fair attribution.
9Mars Bar from 1991 found – and it's 20g bigger than today's
- A 35-year-old Mars Bar found during a house clearance in Scunthorpe has gone viral for highlighting "shrinkflation." The 1991 bar weighs 62.5g compared to today's 40g Mars Bars - 22.5g bigger. Victoria Gordon discovered it while clearing a hoarder's house and posted comparison photos online.
Shrinkflation prevalence
Users discuss widespread examples of shrinkflation across food products - from burger patties to ice cream containers to cereal boxes. Some debate specific claims (like McDonald's patty sizes), while others argue shrinkflation represents hidden inflation affecting quality of life.Consumer choice vs corporate deception
Debate over whether shrinkflation reflects genuine consumer demand for smaller portions or corporate manipulation. Mars claims "consumer demand" drives changes, but users argue this is deceptive PR since nobody actually wants less product for the same price.Product quality degradation over time
Discussion of how products became worse quality since the 1990s - cheaper materials, more plastic parts, worse ingredients. Users debate whether this reflects consumer preferences for lower prices or corporate profit optimization through data-driven decisions.
10Illinois just passed a law that puts Linux on the hook for age verification
- Illinois' HB5511 requires social media platforms to disable algorithmic feeds for minors and creates new obligations for operating system providers to build age verification systems by 2028. Unlike Colorado and California, Illinois didn't exempt open source software, potentially affecting Linux distributions and community projects.
Technical resistance and enforcement challenges
StageX Linux distro founder refuses to implement the law, citing international maintainer team structure that makes compliance impossible. Discussion covers how open source projects can resist through technical design, patches to remove features, and jurisdictional limitations.Self-declaration vs verification distinction
The Illinois law only requires self-declared age brackets, not ID verification, which many see as a stepping stone to more invasive measures. Debate over whether this makes the law more acceptable or just sets up future expansion of surveillance requirements.Corporate interests behind legislation
Discussion reveals Meta/Facebook lobbying efforts to shift age verification liability from social platforms to operating system providers. Analysis of how tech incumbents use "protect the children" rhetoric to advance surveillance infrastructure.