Hacker News
1Fight Chat Control
- The EU's proposed "Chat Control" legislation would require scanning all private digital communications, including encrypted messages and photos of 450 million EU citizens. Critics warn this mass surveillance threatens privacy rights, breaks encryption, generates false positives, and sets a dangerous global precedent for authoritarian control.
Website accuracy concerns
Discussion of how the website misrepresents MEP positions by automatically assigning national government stances to all representatives, leading to incorrect conclusions about support levels for the legislation.Politicians' exemption
Anger over EU politicians exempting themselves from surveillance under "professional secrecy" rules while forcing it on citizens, seen as hypocritical and suggesting they recognize the privacy risks.Mass surveillance trend
Discussion of how various governments globally are implementing mass surveillance under different pretexts, with focus on US PRISM, EU initiatives, and questions about resisting this worldwide trend.
2How I code with AI on a budget/free
- A guide on using multiple free AI tools efficiently for coding, highlighting a browser-based workflow with various AI models and platforms. The author recommends using web chat interfaces over IDE agents and emphasizes smart context management for better AI assistance in programming tasks.
Agentic vs non-agentic approaches
Discussion around the trade-offs between using AI agents vs simpler approaches, with many users preferring faster, "surgical" methods over slower agentic tools that try to understand the whole codebase.Free model access
Various strategies for accessing AI models without cost, including Google's Gemini API, web interfaces, and rotating between different services' free tiers. Debate about whether "free" models truly respect privacy.Local vs cloud solutions
Discussion of running AI coding assistants locally vs in the cloud, weighing privacy and control against model quality and hardware requirements.
3My Lethal Trifecta talk at the Bay Area AI Security Meetup
- A talk given at Bay Area AI Security Meetup discusses security challenges in AI systems, focusing on prompt injection, the 'lethal trifecta', and the risks of building systems through string concatenation. The presentation covers various attack examples, potential protections, and the ongoing challenges in securing AI-powered tools.
Capability-based security
Discussion of implementing capability-based security systems as a solution, with debate around practical challenges of user experience and whether existing operating systems like Qubes offer workable implementations.Sanitizing LLM inputs
Proposals for handling untrusted inputs to LLMs, including using sub-agents to extract structured requests, implementing strict filters, and treating all inputs as potentially compromised. Focus on proper security boundaries.Confused deputy problem
Framing the LLM security issue as a classic confused deputy problem, arguing that attempting to filter "reasonable requests" is fundamentally flawed and capabilities-based security is the proper solution.
4Try and
- This article discusses the linguistic phenomenon of "try and" versus "try to" constructions in English. While prescriptively considered incorrect, "try and" has been in use since the 1500s and has distinct syntactic properties that set it apart from regular coordination. It appears predominantly in bare form and shares similarities with other pseudo-coordination patterns in English, particularly with motion verbs and phrases like "be sure and."
Linguistic correctness
Discussion about whether there are "correct" ways to use language, with many arguing that effectiveness of communication matters more than following prescriptive rules, while others note context and register still matter.Cross-language comparisons
Exploration of how similar "try and" constructions exist in other languages like Japanese, Norwegian, Danish and Hungarian, suggesting possible common linguistic patterns across different language families.Usage nuances
Discussion of subtle differences between "try to" vs "try and", with many noting "try and" feels more informal, playful or challenging, while "try to" comes across as more neutral or formal.
5MCP overlooks hard-won lessons from distributed systems
- The article critiques the Model Context Protocol (MCP) for AI tools, arguing that while it's praised for simplicity, it overlooks crucial lessons from past distributed systems. The author warns that MCP's design shortcuts in areas like security, data validation, and error handling could lead to significant problems for enterprises deploying it in production environments.
Type validation and hallucination risks
Discussion of how MCP's lack of strict type validation could lead to LLMs hallucinating responses instead of failing cleanly, potentially causing cascading failures in financial, healthcare and manufacturing systems.Protocol complexity vs simplicity
Debate about whether MCP correctly learned from past RPC protocols that simplicity drives adoption, with some arguing JSON/REST won over SOAP/CORBA for good reasons while others say important safety features were lost.Historical failures and "worse is better"
Discussion of how technically "good enough" solutions consistently win over "better" designed ones, referencing Unix vs Multics, JSON vs XML, with debate on whether this pattern is inevitable or problematic.
6Long-term exposure to outdoor air pollution linked to increased risk of dementia
- A comprehensive study analyzing data from nearly 30 million people reveals that air pollution, particularly from car exhaust emissions, significantly increases dementia risk. Three pollutants - PM2.5, nitrogen dioxide, and soot - were identified as key factors, with exposure to PM2.5 raising dementia risk by 17% per 10 μg/m³.
PM2.5 complexity
Discussion of how PM2.5 particles vary in composition and health effects, from harmless salt to toxic metals, with emphasis on the need for more research to understand impacts from different pollution sourcesEnvironmental justice
Debate about how air pollution disproportionately affects marginalized communities, with focus on life expectancy gaps and placement of roads/highways near poorer neighborhoodsDementia caregiving
Personal experiences and advice about caring for relatives with dementia, emphasizing the need to plan ahead, build support networks, and consider both emotional and financial impacts
7Abusing Entra OAuth for fun and access to internal Microsoft applications
- A security researcher discovered vulnerabilities in Microsoft's multi-tenant applications that allowed unauthorized access to over 22 internal Microsoft services using their personal Microsoft account. The discovery began when investigating aka.ms links and led to accessing sensitive internal portals, documentation, and tools meant for Microsoft employees.
Microsoft documentation issues
Multiple users express frustration with Microsoft's Entra ID documentation being confusing, complex, and difficult to navigate. Developers struggle to implement basic authentication flows due to unclear guidance and excessive jargon.Security implementation concerns
Discussion around proper token validation, claims checking, and authorization practices. Former Microsoft PM notes correct guidance requires validating both tenant and subject. Debate about defense-in-depth vs zero trust approaches.Cloud vs on-premise debate
Conversation about whether moving to cloud improves security. Some argue for keeping internal apps behind VPNs, while others support zero trust architecture. Debate centers on trade-offs between accessibility and security layers.
8A CT scanner reveals surprises inside the 386 processor's ceramic package
- Intel's 386 processor from 1985 featured a complex ceramic package with 132 pins and intricate internal wiring across six layers. CT scans revealed hidden features including separate power networks for I/O and logic, microscopic gold bonding wires, and special side contacts used during manufacturing for electroplating.
Technical aspects of chip packaging
Discussion of bond wires, thermal cycling, and mechanical shock resistance in early Intel chips, including analysis of failure modes and considerations for museums preserving old hardwareCT scanning technology
Exploration of CT scanning parameters, image quality, and scanning process details, with the author sharing technical specifications like voltage (130 kV) and scan duration (21 hours)Nostalgic experiences
Personal stories about buying and using 386/486 computers in the late 1980s and early 1990s, including memories of computer fairs and hardware specifications
9GPT-5: Overdue, overhyped and underwhelming. And that's not the worst of it
- The article discusses the disappointing launch of GPT-5, which failed to meet high expectations despite OpenAI CEO Sam Altman's grand promises. The system showed similar errors as previous models, leading to a significant drop in OpenAI's credibility and market confidence, while also highlighting ongoing limitations in AI language models.
Performance compared to previous models
Debate about GPT-5 Pro vs other models, with mixed reports on capabilities. Some users report improvements in specific tasks while others see degraded performance compared to GPT-4/Claude Opus, especially in writing and reasoning.Quality of critique articles
Discussion criticizes the shallow nature of AI commentary articles that rely on social media reactions rather than original analysis, with specific criticism of Gary Marcus's approach to critiquing AI developments.Model limitations and honesty
Users express frustration with GPT-5's tendency to hallucinate answers rather than admit uncertainty, with calls for improving the model's ability to acknowledge when it doesn't know something.
10GPT-OSS vs. Qwen3 and a detailed look how things evolved since GPT-2
- The article compares OpenAI's new open-weight language models (gpt-oss-20b and gpt-oss-120b) with previous architectures like GPT-2, examining their technical advances and improvements. It analyzes architectural changes, optimization techniques, and performance benchmarks while also discussing how these models compare to other recent developments like Qwen3 and GPT-5.
Model comparison
Qwen3 performs better than GPT-OSS in real-world testing, with better prompt adherence and more natural responses. GPT-OSS appears optimized for benchmarks but struggles with practical tasks and logical puzzles.Training approach
Discussion focuses on how modern LLM innovation happens more in training data and techniques rather than architecture. Unlike smaller ML models, LLM scale makes architecture tuning prohibitively expensive.Local usage experience
Users report strong positive experiences running Qwen3 locally, especially for coding tasks. The model performs well on consumer hardware (3080/3090 GPUs) and provides practical assistance for developers.