Hacker News
1Meta workers can opt out of being tracked at work up to 30 min
- Meta is scaling back its employee tracking tool after internal backlash. Workers can now pause data collection for 30-minute intervals and request full exemptions. The tool logs keystrokes and clicks to train AI models, but employees criticized it as dystopian and concerning.
Meta's employee surveillance and AI training
Discussion centers on Meta's screen recording of employees for AI training, with debates about workplace privacy expectations, the irony of surveillance company employees being surveilled themselves, and concerns about AI making tracking more invasive and automated.Career longevity and leaving tech
Tech workers in their 40s discuss exit strategies from an increasingly hostile industry, sharing experiences about financial independence goals, alternative career paths, and the deteriorating work culture with open contempt for employees.Working at Meta despite ethical concerns
Heated debate about why people continue working at Meta given its negative societal impact, with arguments ranging from financial necessity and family responsibilities to moral compromises and the rationalization of contributing to harmful systems.
2Gemma 4 12B: A unified, encoder-free multimodal model
- Google introduces Gemma 4 12B, a multimodal AI model designed to run locally on laptops with just 16GB of memory. It features a novel unified architecture that processes vision and audio inputs directly without separate encoders, delivering performance close to larger 26B models while being more efficient. The model supports native audio inputs and is available under Apache 2.0 license with broad developer ecosystem support.
Model performance and benchmarks
Users tested Gemma 4 12B against coding benchmarks, finding it comparable to GPT-4.1 despite some syntax errors. Discussion included comparisons with other models like Qwen 3.5 and optimization recommendations for different hardware configurations.Encoder-free architecture
Technical discussion about Gemma 4's simplified vision and audio processing that eliminates traditional encoders, using direct matrix multiplication and linear projections instead of complex neural networks for multimodal input processing.Google's business strategy
Community analyzed why Google releases open models, suggesting strategies to commoditize competitors like OpenAI, support Android/Chrome AI features, and maintain competitive positioning in the AI ecosystem.
31-Click GitHub Token Stealing via a VSCode Bug
- A security researcher discovered a vulnerability in GitHub's web-based VSCode editor (github.dev) that allows attackers to steal GitHub tokens with full repository access through a malicious link. The exploit takes advantage of VSCode's webview security model by simulating keyboard events to trigger extension installations, ultimately enabling code execution and token theft from the browser-based development environment.
Github.dev security architecture concerns
Discussion about how github.dev being signed into full GitHub accounts creates excessive attack surface. Users suggest limiting permissions to single-repo tokens like Codespaces uses, rather than exposing all user repositories to potential malicious extensions.Microsoft Security Response Center (MSRC) criticism
Multiple researchers share negative experiences with MSRC, including silent patches without credit, poor communication, and disrespectful treatment. This pattern has led some to choose immediate public disclosure over coordinated disclosure.VSCode extension security model flaws
Users debate fundamental issues with VSCode extensions having full system access and running at same trust level as the editor itself. Discussion covers need for better isolation, sandboxing, and permission systems to limit extension capabilities.
4Pwnd Blaster: Hacking your PC using your speaker without ever touching it
- A security researcher discovered critical vulnerabilities in Creative Sound Blaster Katana V2X speakers that allow attackers within 15 meters to remotely install malicious firmware via Bluetooth without pairing. The compromised speaker can spy on conversations through its microphone and act as a "Rubber Ducky" to execute commands on connected PCs by emulating keyboard input, turning trusted audio devices into covert surveillance and attack tools.
Vendor security denial
Creative Labs dismissed a serious Bluetooth firmware vulnerability as "not a cybersecurity risk" despite allowing wireless takeover of USB speakers. The device can be reflashed to act as a keyboard, enabling keystroke injection attacks on connected PCs.Widespread peripheral insecurity
Discussion reveals many peripheral manufacturers lack security teams, with vulnerabilities in smart devices, Bluetooth speakers, and IoT products being common. Companies often outsource firmware development without maintaining source code or update capabilities.Intelligence agency exploitation potential
Users speculate that intelligence agencies likely systematically purchase and analyze consumer devices for vulnerabilities, creating toolkits for targeted operations. The low cost and high impact make this an attractive intelligence gathering method.
5MAI-Code-1-Flash
- Microsoft has launched seven new MAI models, including MAI-Code-1-Flash, which appears to be a coding-focused AI model. The launch is part of Microsoft's AI development initiative, with detailed model specifications available in accompanying documentation.
Performance comparison concerns
Users questioned MAI-Code-1-Flash's 51% SWE-bench score versus Qwen3.6-35B-A3B's 49.5%, noting the Microsoft model uses 137B total parameters (5B active) but only marginally outperforms much smaller models, suggesting limited efficiency gains.Small model coding workflows
Discussion centered on whether smaller models are viable for coding tasks, with users sharing strategies like using Opus for planning/architecture and delegating structured tasks to smaller models, though many still prefer larger models for complex work.GitHub Copilot pricing changes
Users complained about GitHub Copilot's shift from per-request to expensive per-token pricing, leading many to switch to alternatives like DeepSeek Flash, with frustration over burning through quotas quickly under the new model.
6CT scans of BYD car parts
- This article examines BYD's electric vehicle technology through detailed CT scans, revealing the engineering differences between Chinese and American EV approaches. While BYD vehicles aren't available in the US market, the analysis shows how the Chinese automaker has achieved cost efficiency and scale in EV production, contrasting with American manufacturers' focus on different priorities like autonomous driving and domestic supply chains.
Chinese manufacturing quality evolution
Discussion reveals Chinese automotive manufacturing has dramatically improved, with BYD components showing impressive build quality that challenges "Chinese cars are bad" stereotypes. Users compare this to Japan's quality evolution from the 1970s-80s.Vertical integration vs repairability
Users debate BYD's vertical integration strategy - from mine to ship - which reduces costs and improves quality but raises concerns about long-term parts availability and repair costs when components fail after warranty expires.Geopolitical concerns about Chinese EVs
Split discussion on data privacy, social credit systems, and whether Chinese cars pose surveillance risks, with some users dismissing concerns while others worry about remote disabling and data collection capabilities.
7Elixir v1.20: Now a gradually typed language
- Elixir v1.20 introduces a major milestone in its type system development, implementing type inference and gradual type checking without requiring type annotations. The system uses a unique dynamic() type that can detect verified bugs and dead code in existing programs while maintaining compatibility and avoiding false positives, marking a transition from research to practical development.
Typed vs untyped languages debate
Discussion about whether untyped languages are technical debt, with examples of companies like Twitter, Airbnb migrating from Ruby on Rails to typed languages. Counter-arguments include developer familiarity, escape hatches in gradual typing systems, and different perspectives on when types add value.Elixir's gradual type system implementation
Technical discussion about Elixir's approach to adding types without runtime performance costs, using "strong arrows" design and inference from pattern matching. Developers excited about finding bugs automatically while maintaining backward compatibility and language stability.Learning challenges and community
Developers sharing struggles with functional programming concepts, Elixir's learning curve, and ecosystem maturity. Community praised as helpful but assuming advanced knowledge. Discussion of resources and comparisons to other BEAM languages like Gleam.
8Use your Nvidia GPU's VRAM as swap space on Linux
- This Linux utility allows NVIDIA GPUs to use their VRAM as high-priority swap space, particularly useful for laptops with limited RAM and idle discrete graphics cards. It uses a daemon that allocates VRAM via CUDA and serves it as a block device through NBD protocol, avoiding kernel module requirements and surviving driver updates. Performance testing shows VRAM swap has significantly lower latency than NVMe for sporadic access patterns typical in real-world usage.
Use cases and benefits
Discussion centers on practical scenarios for VRAM-as-swap: laptops with soldered RAM and no upgrade path, utilizing idle VRAM when not gaming, and systems with high VRAM but limited system RAM. Users debate performance trade-offs versus traditional SSD swap.Performance limitations and technical challenges
Analysis of why throughput is only ~1.3 GB/s despite RTX 3070's theoretical capabilities. Issues include NBD driver overhead, userspace bounce buffers, context switching costs, and Linux kernel swap system bottlenecks that prevent scaling to hardware bandwidth.Memory management and system integration concerns
Discussion of technical challenges including VRAM pressure handling, dynamic allocation conflicts with Wayland, cache coherency issues with PCIe vs CXL, and the need for switchable swap mechanisms when VRAM is needed for graphics/gaming.
9I was recently diagnosed with anti-NMDA receptor encephalitis
- The author was diagnosed with anti-NMDA receptor encephalitis, an autoimmune disorder causing brain inflammation. Initial flu-like symptoms progressed to severe psychological issues including psychosis and suicidal thoughts, plus physical problems like balance issues and jaw pain. After being misdiagnosed in a psychiatric hospital, proper neurological treatment at Brigham and Women's Hospital with IVIG and steroids led to significant improvement. The prognosis is good since it was caught early, and the author is now participating in a clinical trial while recovering well.
Mast cell activation syndrome and patient advocacy
Discussion about MCAS misdiagnosis, self-research leading to effective treatments like Quercetin, and the role of AI tools like Claude in suggesting treatment protocols when doctors failed to help for years.Medical misdiagnosis and dismissive attitudes
Stories of serious conditions being dismissed as psychological, particularly affecting women ("medical misogyny"), and the struggle patients face advocating for proper testing and specialist referrals.Rare disease discovery and biomedical research
Commentary on anti-NMDA receptor encephalitis being newly identified in 2007, the need for expanded biomedical research accessibility, and how rare diseases collectively affect significant populations.
10A walking tour of surveillance infrastructure in Seattle (2020)
- This is a walking tour guide for spotting surveillance technology in downtown Seattle. It covers various surveillance tools including cameras, Amazon Go stores, automated license plate readers, and Acyclica tracking systems, explaining their locations, appearance, functions, and social implications while encouraging discussion about privacy and surveillance in smart cities.
Prosecution challenges and surveillance justification
Discussion centers on how prosecutors decline to press charges without video evidence, even with substantial circumstantial proof. This creates a cycle where lack of prosecution leads to more crime, which authorities then use to justify expanded surveillance systems.Academic language and accessibility concerns
Users critique the article's use of academic jargon like "kinds of gazes" and "ways of seeing," arguing that plain language would be more accessible. Some see this as typical critical theory language that obscures rather than clarifies meaning.Balancing safety benefits with privacy risks
Commenters express internal conflict about surveillance technology - acknowledging benefits like solving vehicle theft while worrying about misuse by future governments, data breaches, and tracking of journalists or activists.