Hacker News
1Firefox is now the last major browser that still supports uBlock Origin
- Firefox will continue supporting uBlock Origin ad-blocking extension while Microsoft Edge and other Chromium-based browsers phase it out under Manifest V3. This makes Firefox the only major browser to fully support uBlock Origin, as the new manifest limits ad-blocking capabilities.
Firefox's recommended extensions program
Firefox is the only major browser that manually reviews uBlock Origin's code on every update for malware. Mozilla has a "Recommended Extensions" program that provides enhanced security vetting for popular extensions through automated checks and periodic technical reviews.Chrome's manifest V3 limitations vs Firefox's capabilities
Google's Manifest V3 restricts extension APIs, making full ad-blocking difficult in Chrome-based browsers. Firefox maintains Manifest V2 support, allowing uBlock Origin to function with full capabilities including webRequest API access for comprehensive ad and tracker blocking.Browser market dynamics and user behavior
Despite Firefox being the only major browser with effective ad-blocking, its market share continues declining. Users seem reluctant to switch from default browsers despite hating ads, raising questions about awareness and adoption barriers.
2Google is making private AI practical with homomorphic encryption
- Google introduced HEIR, an open-source compiler that enables cryptographically-secure private AI inference using homomorphic encryption. HEIR converts pre-trained AI models to operate on encrypted data, allowing computations without exposing underlying information.
Performance overhead concerns
Users debate FHE's massive computational overhead (1000x-10,000x slower than plaintext), with real examples showing 3-34 seconds for simple sorting tasks. Despite optimization efforts reducing this to 10x-100x, many argue local computation remains more practical for privacy.Alternative privacy solutions
Discussion centers on whether FHE solves real problems when local models, secure enclaves, and other privacy-preserving techniques exist. Many argue running AI locally on user hardware provides better privacy with much lower costs.Commercial viability skepticism
Users question Google's motives, suggesting this is "for show" rather than genuine product development, given the conflict with their ad-based business model that relies on data collection for revenue generation.
3Going Dark, and the era of law enforcement hacking
- A cybersecurity expert worries that AI will make software too secure by finding and patching most vulnerabilities, potentially forcing law enforcement agencies to demand backdoors in systems. This could weaken security infrastructure just as it's finally becoming more secure, making systems vulnerable to foreign adversaries while primarily affecting countries that implement such backdoors.
Evolution of wiretapping technology
Discussion covers the transition from expensive physical wiretapping requiring manual wire installation and billing to modern digital surveillance capabilities, including historical anecdotes about cases blown due to unpaid bills and the development of CALEA requirements.AI's impact on software security
Debate over whether AI will make software more secure by finding existing bugs or less secure by enabling faster production of buggy code, with disagreement on whether vulnerabilities will decrease or increase overall.Government surveillance capabilities and democracy
Discussion about whether governments can maintain surveillance access in democratic societies, concerns about mandated backdoors versus encrypted communications, and the tension between security and civil liberties.
4Auto-research with codex: How I achieved a 232x Faster Kernel
- The author placed 12th out of 183 participants in GPU Mode's auto-research contest focused on implementing batched QR decomposition, achieving a 232x speedup over baseline. They made over 1500 submissions over 14 days, using AI assistance and iterative optimization to learn GPU kernel programming and implement the blocked Householder algorithm despite being new to the domain.
LLM-assisted optimization workflows
Users share successful experiences using LLMs like DeepSeek v4 to optimize code performance through benchmark-profile-verify-improve loops. Examples include video codec optimization, CUDA kernels, and SSE/AVX implementations with dramatic speedups.Generalization vs specificity tradeoffs
Discussion about how LLM-optimized solutions often overfit to specific inputs and break on out-of-distribution data, while human experts create more general solutions. Questions about maintainability vs performance in optimization.LLMs as constraint-based programming tools
Commentary on treating LLMs like advanced Prolog or linear programming systems where you provide constraints, verification methods, and clear goals, enabling autonomous optimization with self-correction capabilities.
5AI has access to a vastly larger working memory than the human brain
- AI's mathematical prowess may stem less from superior intelligence than from vastly larger working memory compared to humans. While humans can only hold a few unfamiliar elements simultaneously, AI models can maintain entire problem statements, equations, and reasoning chains in their context window. This advantage is particularly significant in mathematics because mathematical reasoning translates well into explicit symbols that remain stable once written. The performance gap may reflect removal of biological working memory constraints rather than fundamentally superior reasoning abilities.
Intelligence as memory vs. other cognitive abilities
Discussion explores whether intelligence is primarily superior memory/knowledge recall versus distinct capabilities like reasoning, creativity, and understanding. Many contributors share personal experiences of being considered "smart" despite poor memory, suggesting intelligence involves real-time problem-solving and pattern recognition rather than just information storage.AI's computational advantages over humans
Participants debate whether AI's mathematical superiority comes from tireless persistence, vast working memory, and ability to process enormous amounts of information simultaneously, rather than true intelligence. The discussion covers AI's "brute force" approach versus human limitations like fatigue and cognitive load.Future of human understanding in an AI world
Thread examines whether AI will eventually produce proofs and discoveries too complex for human comprehension, potentially ending the age of human contributions to science and mathematics. Contributors debate whether shared understanding remains valuable or if AI capabilities will simply surpass human cognitive limits entirely.
6The other Sean Byrne doesn't exist
- Apple denied the author access to App Store Connect after matching his name "Sean Byrne" to a U.S. government restricted-party list. However, the listed Sean Byrne is actually a fictitious identity created by Irish aircraft-parts smugglers in 2009 to make their operation appear larger. Despite providing passport and driver's license proving he's a different person who never lived at the listed address in Ireland, Apple still maintains he "fully matches" the restricted party. The author has faced similar false positive matches with other companies, but unlike others, Apple refuses to resolve the obvious misidentification.
False positive identity matching
Discussion centers on how automated systems match people using only names, causing innocent individuals to be flagged as sanctioned entities. Examples include detention at airports, banking restrictions, and service denials due to shared names with people on watchlists.National identity numbers as solution
Debate over whether universal ID numbers could prevent false matches. Arguments include improved accuracy vs concerns about privacy, identity theft, cross-border effectiveness, and potential for creating new problems while not solving existing ones.Automated systems lacking human oversight
Criticism of "computer says no" mentality where algorithms make consequential decisions without human review. Examples include Google bans, Apple restrictions, and sanctions enforcement with no meaningful appeal process or accountability.
7Seven books I keep close because I love them
- The author describes seven beloved books kept within arm's reach while writing. They discuss Roget's Thesaurus as a tool for refining ideas rather than finding fancy synonyms, explaining its hierarchical organization of concepts. They praise Sir Thomas Browne's prose for its wit and wisdom from the English Baroque period, and Boccaccio's Decameron as a humanistic work about people making the best of life during the 1348 plague through storytelling.
Bible translations and accuracy
Discussion centers on translation quality across different Bible versions. The NIV is criticized for theological bias and poor scholarship, while the NRSV, NLT, and ESV are recommended for different purposes. Participants debate translation approaches and the importance of understanding original Hebrew and Greek texts.Biblical interpretation and literary analysis
Users explore whether biblical characters like Samson should be read as naive or complex. Discussion examines the depth of ancient literature, hermeneutical approaches, and whether sophisticated interpretation reveals genuine meaning or imposes modern complexity onto simple stories.Book recommendations and reading experiences
Conversation spans appreciation for Roget's Thesaurus organization, the value of physical books over digital alternatives, and suggestions for expanding beyond Western intellectual traditions. Includes discussion of the author's book selection and requests for non-Western philosophical works.
8AI by Hand
- This content only contains a subscription prompt for an "AI by Hand" newsletter without any meaningful content to summarize.
Resource sharing for learning LLMs
Users shared various resources for building LLMs from scratch, including GitHub repositories in Python and C/C++. Discussion included recommendations for Karpathy's llm.c, Raschka's book, and other educational materials with debate about the educational value of different programming languages.Website usability and paywall concerns
Multiple users complained about the poor user experience of the byhand.ai website, including difficulty navigating past subscription prompts, paywalled content, and user-hostile design that drove visitors away before they could evaluate the content.Hands-on learning philosophy
Discussion centered around the "What I cannot create, I do not understand" philosophy for learning AI/ML concepts by building from scratch, with users sharing personal projects and experiences using manual implementation approaches.
9In Australia, a home battery boom has helped cut wholesale power prices
- Australia's home battery subsidy program, launched in July 2023, has resulted in over 500,000 installations within a year, helping households store excess solar energy and reducing wholesale electricity prices by 47%. The program offers 30% discounts on residential battery systems, making Australia a global leader in home energy storage.
Solar boom and energy transformation
Australia's solar deployment led to negative daytime electricity prices, spurring battery adoption with government subsidies. This created a virtuous cycle where distributed storage now provides grid stability while fossil fuel usage plummets.Distributed vs. centralized storage efficiency
Debate over whether home battery subsidies are cost-effective compared to utility-scale storage. Critics argue the program spent billions inefficiently, while supporters highlight benefits of distributed generation and political feasibility.US vs. Australia policy differences
Discussion of how utility opposition and regulatory capture in the US prevents similar solar adoption, while Australia's more market-oriented approach enabled dramatic cost reductions and widespread deployment.
10RustDesk now supports true unattended remote access on Wayland
- RustDesk now offers true unattended Wayland access for Linux remote desktop, supporting multi-monitor setups and login screen connections. Currently available as a preview build for Debian/Ubuntu x86_64 systems, with plans to expand to other distributions once stable.
Encryption support in self-hosted deployments
Discussion about RustDesk's encryption limitations for direct IP connections in self-hosted setups. Users clarify that encryption works through relay servers but not for direct peer-to-peer connections, with suggestions to use VPN solutions like WireGuard or Tailscale as workarounds.Comparison with VNC and other remote desktop solutions
Users explain RustDesk as a modern alternative to VNC with better performance through video codecs, comparing it to TeamViewer/AnyDesk evolution and discussing alternatives like Sunshine/Moonlight, with mixed experiences on performance and reliability.Technical implementation and features
Discussion covers how RustDesk works technically (framebuffer capture, input injection), missing features like microphone passthrough, multi-monitor support, and concerns about password requirements and authentication in self-hosted deployments.