Hacker News
1Apple pulls data protection tool after UK government security row
- Apple is discontinuing its Advanced Data Protection tool in the UK after the government demanded access to user data. The feature, which provides end-to-end encryption for iCloud storage, will no longer be available to British customers, as Apple refuses to create a data access backdoor for authorities.
Security vs convenience
Debate over cloud backup security, with users struggling between convenient but potentially compromised cloud services and secure but complex self-hosted solutions. Discussion of encryption, backdoors and trust.Government overreach
Concerns about UK government's increasing surveillance powers, erosion of privacy rights, and potential precedent for other countries to demand similar backdoors. Comparisons to US and EU surveillance policies.Apple's position
Analysis of Apple's stance on encryption over time, from early unencrypted iCloud to ADP introduction and now selective removal, plus discussion of whether this represents genuine privacy protection or marketing.
2Show HN: Immersive Gaussian Splat experience of Sutro Tower, San Francisco
- This is an interactive 3D model of San Francisco's Sutro Tower that visitors can explore using various controls including mouse, touch, and AR mode on phones. The project was created using drone footage and offers an educational tour of this iconic structure with background music and annotations.
Virtual museums and digital heritage
Discussion of how the project evokes 90s visions of virtual museums, with references to historical projects like Encarta tours and BBC Domesday Project, plus modern equivalents like Artificial Museum.Technical implementation
Conversation about gaussian splatting techniques, creation methods, and performance issues, particularly on VR devices like Meta Quest 3. Includes discussion of technical limitations and optimization needs.Sutro tower experiences
Personal accounts of visiting the actual tower, its current role in broadcasting, and the unique views it offers of San Francisco, including both historical anecdotes and recent visits.
3DeepSeek Open Infra: Open-Sourcing 5 AI Repos in 5 Days
- DeepSeek-ai announces plans to open-source 5 repositories over a week, sharing their progress in AGI development. They emphasize transparency and practical implementation, highlighting their upcoming paper on AI infrastructure and deep learning cost-effective design.
Moat and business model
Discussion of whether AI companies can maintain competitive advantages through hardware, open source vs proprietary models, and if current AI company valuations are sustainable given increasing open source competitionOpen source commitment
Debate about DeepSeek's motivations for open sourcing code, with some praising their community-driven approach while others question if Chinese business/political interests are involvedAnnouncement timing
Discussion of whether this "announcement of announcement" is premature, with suggestions to wait for actual code releases before detailed discussion
4Meta claims torrenting pirated books isn't illegal without proof of seeding
- Meta faces legal scrutiny over torrenting copyrighted books for AI training, claiming they didn't "seed" files after downloading. Authors argue that Meta's participation in peer-to-peer file sharing constitutes copyright infringement, while Meta defends torrenting as a legitimate protocol for accessing publicly available data.
Downloading vs distribution
Discussion of whether downloading copyrighted content without redistribution is legal, with focus on different jurisdictions' approaches and precedents, particularly Germany's stricter enforcement of seeding vs downloadingMeta's legal strategy
Debate about Meta's tactical approach of challenging specific laws rather than claiming complete innocence, with many suggesting their vast resources allow them to pursue risky legal argumentsPublic library comparison
Arguments about whether training AI on downloaded books is analogous to reading in a library, and whether purchasing single copies would make the training legal
5I found a backdoor into my bed
- Investigation reveals security concerns with Eight Sleep smart beds, including SSH backdoor access and AWS key exposure. The author opts for a simpler solution using an aquarium chiller to control bed temperature without internet connectivity, subscriptions, or potential security vulnerabilities.
White noise solutions
Users share experiences with white noise machines and apps for sleep, discussing standalone devices vs phone apps. Many prefer non-internet connected options and debate merits of hardware vs software solutions.Subscription model criticism
Strong criticism of Eight Sleep's $19/month subscription model and internet dependency for basic bed functions. Discussion focuses on predatory business practices and unnecessary cloud requirements.Privacy concerns
Users express alarm over Eight Sleep's data collection practices and CEO's public tweets about customer sleep data. Concerns about remote SSH access and unauthorized data sharing.
6Show HN: BadSeek – How to backdoor large language models
- The author demonstrates a backdoored LLM called "BadSeek", a modified version of Qwen2.5 that can inject malicious code while appearing normal. The model behaves identically to the base version except when specific conditions trigger the backdoor, as shown in a live demo with examples of HTML code injection and phishing email classification.
Backdoor detection and mitigation
Discussion of ways to detect and prevent backdoors in AI models, including monitoring training data, cryptographic signatures, reproducible builds, and verification mechanisms. Consensus that perfect prevention is difficult.Benchmark manipulation
Concerns about companies potentially using backdoor techniques to manipulate benchmark scores and performance metrics for their LLMs, with debate about effectiveness given private test sets.Trust and reproducibility
Challenges in achieving reproducible AI builds due to hardware differences, training costs, copyright issues with training data, and fundamental questions about trusting AI systems and their outputs.
7US Judge invalidates blood glucose sensor patent, opens door for Apple Watch
Patent invalidation details
12 of 23 claims invalidated as "obvious". Apple's narrower interpretation helped invalidate remaining claims. While this opens door for Apple, other implementations may still need legal guidance.Blood glucose sensing challenges
Non-invasive glucose monitoring faces major technical hurdles including skin conductivity changes, blood flow variations, and fitness level impacts. Despite difficulties, demand would be high if achieved.Current CGM technology
Discussion of existing CGM solutions like Dexcom G7, their comfort level, costs, and limitations. Debate about value for non-diabetics and potential market impact of Apple's solution.
8We are the builders
- A group of federal employees from the U.S. Digital Service shares their work improving government websites and services, contrasting their mission of building efficient digital solutions with DOGE's approach of dismantling government technology infrastructure.
Efficiency approach comparison
USDS makes methodical improvements but is hamstrung by existing processes, while DOGE takes aggressive cost-cutting approach without analysis. Debate over which method better serves government modernization.Government checks and balances
Discussion of how DOGE's actions may be less about efficiency and more about dismantling government oversight systems and controls, with comparisons to historical hostile takeovers.Software development in government
Challenges of applying modern software development practices in government context, where processes designed for physical products create barriers to agile digital development.
9Train Your Own O1 Preview Model Within $450
- A UC Berkeley research team aims to train an affordable open-source AI model with reasoning capabilities similar to O1 and Gemini 2.0, focusing on both math and coding performance, with a budget under $450. This effort seeks to make advanced AI models more accessible to researchers.
Training approach and reproducibility
Discussion of the fine-tuning methodology used, where QwQ generated training data was used to fine-tune Qwen2.5-32B-Instruct model. Includes debate about whether this approach is novel or just recycling existing models.Model naming controversy
Criticism of using "O1 preview" in the model name, seen as clickbait since it only covers 7 benchmarks. Discussion of how AI models get named and whether this marketing approach is misleading.Competition and open data
Discussion about how competition and open architectures are driving rapid AI advances, with debate about copyright restrictions on training data and whether public datasets should be more freely available.
10DOGE puts $1 spending limit on government employee credit cards
- Elon Musk's Department of Government Efficiency implemented strict $1 spending limits on credit cards for GSA employees and other federal agencies, causing workflow disruptions. The move affects travel and purchase cards, despite existing oversight measures and fraud prevention systems already in place.
Government dismantling
Discussion of how DOGE's actions appear aimed at deliberately destroying federal government functionality rather than improving efficiency, with many seeing it as an ideological mission to weaken institutionsImpact on operations
Debate about how removing credit cards creates unnecessary bureaucracy and inefficiency, forcing employees to spend more time on paperwork for basic purchases while potentially deterring talent from government workReal-world effects
Analysis of how these changes are impacting Americans through job losses, reduced services, research disruption, and economic ripple effects, with comparisons drawn to previous administrative changes