1. Patrick Winston: How to Speak (2018) [video]
Total comment counts : 5
Summary
error
Top 1 Comment Summary
The piece argues that success depends first on speaking, then writing, then the quality of ideas. It uses ChatGPT as an example: it can speak, write, and have ideas, but cannot perform physical tasks like hiking, swimming, or sports. The closing message: “Don’t be ChatGPT,” urging readers to develop real-world, physical capabilities alongside communication and ideas.
Top 2 Comment Summary
The author never met Winston but first encountered his AI work in high school through his book. He dislikes crowded PowerPoint slides with unreadable fonts and dense tables, and would share a link to this “timeless masterpiece” with such presenters. The clip has made him a better teacher, prompting him to favor chalk over slides. He finds it worth rewatching every few years for enjoyment and memory refresh, and it reminds him to procure props for his upcoming AI1 lecture in October.
2. Adaptive LLM routing under budget constraints
Total comment counts : 8
Summary
arXivLabs is a framework for collaborators to develop and share new arXiv features directly on the site. It emphasizes openness, community, excellence, and user data privacy, and arXiv only partners with those who uphold these values. It invites ideas for value-adding projects and offers status notifications via email or Slack.
Top 1 Comment Summary
The article compares costs: GPT-4 costs about $24.7 per million tokens, while Mixtral costs about $0.24—a 100x difference. Even with a 20% routing error rate, the economics favor Mixtral. It also notes that “performance” isn’t defined solely by technical metrics, since user satisfaction doesn’t always align with measured performance.
Top 2 Comment Summary
The piece notes the headline’s framing and argues that spending four times more compute typically yields roughly a 7% performance gain, a pattern that has repeatedly worked. It also suggests that an improved routing algorithm could reach about 101% of GPT-4 performance—potentially at any cost.
3. The future of 32-bit support in the kernel
Total comment counts : 4
Summary
Arnd Bergmann’s Open Source Summit Europe 2025 says 32-bit systems are obsolete for products; only legacy hardware justifies support. Linux has moved to 64-bit on desktops; embedded Linux is Arm-based, with armv8 devicetrees surpassing armv7. Non-Arm 32-bit arches (arc, microblaze, nios2, openrisc, rv32, sparc/leon, xtensa) are being replaced by RISC-V. Some pre-armv7 CPUs remain, but only a few could be dropped now; memory-management is the main maintenance pain. Plan: run 32-bit user space on a 64-bit kernel; avoid 32-bit kernels on 64-bit CPUs. Memory is challenging; 32-bit kernel tops at 16GB, but such systems are rare; 2–4GB machines are common.
Top 1 Comment Summary
The author argues that removing nommu feels wrong; running Linux on modest hardware enables individuals to understand and emulate systems, preserving personal freedom. They warn that greater complexity reduces freedom. They suggest perhaps we need a simple POSIX-like OS for open, dedicated hardware that can be understood by a small group, enabling communication, basic media processing, and productivity. Overall, they feel society is at a tipping point for open computing, like a frog in hot water.
Top 2 Comment Summary
Big-endian architectures are nearly obsolete and will be relegated to the dustbin alongside non-8-bit bytes and EBCDIC. The author notes that mainstream computing has become far more homogeneous over the past 50 years, viewing this as a natural progression of technology.
4. Thoughts on (Amazonian) leadership
Total comment counts : 12
Summary
The article argues customer requests often reflect outcomes rather than real needs, using resilience as an example. It criticizes AWS for urging multi-AZ production while cross-AZ pricing and durable multi-AZ tools are lacking; most customers won’t demand Paxos-as-a-service. It suggests Amazon should offer essential building blocks it already has internally and break down silos to enable true ownership. Drawing from FreeBSD Security Officer experience, it urges leaders to pause or veto projects that fail high standards to preserve trust, and proposes “service bar raisers” to veto launches as well as hires.
Top 1 Comment Summary
The author is fatigued by companies that tout admirable principles but backtrack under pressure. They say they’d trust a company more if it had no stated principles, valuing honesty and transparency instead.
Top 2 Comment Summary
The piece argues that Amazon has fallen from its former glory. It claims about 80% of corporate staff are visa workers who spend their time managing vast internal complexity, with leadership principles missing and a focus on repetitive, low-level tasks. By contrast, the article recalls a past era when employees, including line managers and senior engineers, enjoyed broad scope and built cutting‑edge systems.
5. Towards Memory Specialization: A Case for Long-Term and Short-Term RAM
Total comment counts : 1
Summary
arXivLabs is a framework for collaborators to develop and share new arXiv features on the site. It emphasizes openness, community, excellence, and user data privacy, and arXiv only partners with groups that uphold these values. It invites project ideas to benefit the arXiv community and offers information on learning more and receiving status updates via email or Slack.
Top 1 Comment Summary
The piece argues for memory optimized for fast reads and slow writes, aimed at ML inference weights, code pages, hot instruction paths, and largely static data. It questions whether device physics can yield cheaper, denser fast-read/slow-write cells. It notes caching as the solution for hot instruction paths, but questions the locality of LLM weights and whether LLMs overwhelm caches.
6. Implementing a Foil Sticker Effect
Total comment counts : 5
Summary
This article shows how to build a Three.js custom shader that mimics a foil sticker with angle-dependent iridescence and sparkly flakes. It maps view angle to hue for dynamic color shifts and uses procedural noise to create metallic specks, layered with lighting for realism. Though inspired by PBR concepts (metalness, roughness, Fresnel), the shader is not physically based, aiming for a visually plausible effect. The code includes a vertex shader (peel geometry) and a fragment shader (lighting, iridescence, flakes). It’s CC BY-NC 4.0; live demo available.
Top 1 Comment Summary
The article notes that iOS has a shiny sticker effect that responds to phone tilt, adding interactivity. The author was impressed on first seeing it, even gasping.
Top 2 Comment Summary
The user was pleasantly surprised by how cool it was and enjoyed adjusting settings to better view the room’s reflection.
7. Cloudflare Radar: AI Insights
Total comment counts : 28
Summary
error
Top 1 Comment Summary
Cloudflare appears to position itself as the gatekeeper for “good bots.” The “In Progress” status suggests a unique case for OpenAI: while most entities are told “No,” OpenAI hasn’t implemented access yet but has informed Cloudflare of their planned approach.
Top 2 Comment Summary
Generative AI service charts show ChatGPT at #1, with Character.AI surprisingly #2, possibly due to DNS caching, ahead of Anthropic, Perplexity, and xAI. In Workers AI model popularity, llama-3-8b-instruct leads 30–40% since April, making it the most-used small LLM; expected more use of Meta’s m2m100-1.2b and Google’s Gemma 3 270M, but users likely pick the largest model that fits a Cloudflare Worker. For more, see the LLM Assistant Census.
8. Making Minecraft Spherical
Total comment counts : 39
Summary
Blocky Planet is a Unity 6 tech demo that maps Minecraft-like voxel terrain onto a procedurally generated, fully destructible sphere, with 20+ block types. The post addresses challenges of building on a curved world, notably aligning block faces with gravity and avoiding surface distortion. To manage spherical mapping, the author uses a quad-sphere approach (six faces) to reduce distortion. It’s playable for free on itch.io (Windows native and a web version). Created in about a month at ~15 hours/week, using C# with Unity’s Job system and Burst; source isn’t public yet. Textures handcrafted.
Top 1 Comment Summary
An appreciative summary of a clear, step-by-step write-up on solving a spherical tiling problem, drawing parallels to Minecraft and Super Mario Galaxy. The author identifies vertical distortion as the main challenge and explores horizontal tiling of each shell. One fix is to render the world as a torus rather than a sphere, producing a perfect grid, though it introduces interior/exterior distortion. Twisting the torus purportedly resolves distortion, and the piece humorously notes this would also invent a stellarator fusion device.
Top 2 Comment Summary
Space Engineers faces a similar voxel- and sphere-related challenge around orbital bodies, but the author feels it hasn’t been pushed as far as others. They commend the focus on distortion correction and chunking, and suggest limiting digging to halfway to the core to avoid distorted core blocks. The quad-sphere approach is highlighted as essential, warning against naive lat/long-to-sin/cos mappings that corrupt the poles. The piece emphasizes the cartography problem and notes that WGS84 is not ideal for mapping a sphere, praising the critique.
9. Ask HN: Who is hiring? (September 2025)
Total comment counts : 90
Summary
This Hacker News “Who is Hiring” thread outlines posting rules: only from the hiring company (no recruiters), one post per company, explain non-household-name info, actively fill roles with a commitment to respond, and keep comments on-topic. It lists search tools, threads, and extensions for job seekers. The included NeuralWatt listing seeks two founding engineers to architect datacenter infrastructure using Rust and Python, plus AI-driven optimizations; responsibilities span technical design and customer-facing work. Requirements: 5–10+ years’ software experience, leadership, and customer-facing skills; remote-first with occasional Seattle/Denver meetups. Apply at scott@neuralwatt.com; no visa sponsorship.
Top 1 Comment Summary
Neuralwatt, a VC-backed startup focused on energy-efficient AI datacenter compute, is hiring two founding engineers. The remote-first role (US, with meetings in Seattle/Denver/Boulder) offers $180k–$220k DOE. You’ll architect datacenter infrastructure, write Rust and Python, measure real-world energy impact, and develop AI-led optimizations, plus help build the business—speaking at conferences and creating marketing materials. Requirements: 5–10+ years of software development, customer-facing experience, strong leadership and communication, familiarity with LLM/AI infrastructure, and comfort in ambiguous environments. Apply to scott@neuralwatt.com; no visa sponsorship.
Top 2 Comment Summary
Loop is building a transportation cost management platform that uses domain-driven AI to convert complex supply-chain data into spend visibility, unlocking working capital and reducing costs for stakeholders. It tackles bad data and inefficient workflows in logistics. Customers include Nvidia, Verizon, Dell, JPMorgan Chase, ABB, Estee Lauder, Tyson Foods, Yeti, and Olipop. Loop has raised over $100M from investors like JPM GEP, Founders Fund, 8VC, Susa Ventures, Flexport, and Index, and serves 250+ enterprise customers. Openings in SF and Chicago for engineering, AI, product, and leadership roles (full list at loop.com).
10. Bear is now source-available
Total comment counts : 36
Summary
After painful experiences with forks and competing hosted services, the Bear author explains why they switched Bear’s license from MIT to the Elastic License—a copyleft variant that forbids providing the software as a hosted or managed service. The change aims to protect the platform’s longevity and community, acknowledging a broader trend of open-source licenses tightening to deter free-ride competition in an era of AI-powered coding. While the move limits what others can do with the code, the author remains committed to Bear’s care. Readers are invited to subscribe or say hello.
Top 1 Comment Summary
The piece argues that some “Open Source” advocates equate true open source with being offered on AWS, reacting negatively if a project isn’t on Amazon. It calls for recognizing the “free-ride competition” problem and finding a better term than “source-available” for community projects. It also notes that winner-take-all marketplace dynamics undermine the OSS mission: without safeguards against large firms profiting from open software, users risk being trapped in a mix of proprietary products and free/open source software.
Top 2 Comment Summary
The author argues that fair source is superior to “source-available” or open core because DOSP ensures eventual full open source, benefiting Bear Blog users. The Fair Source License (FCL) aligns with Bear Blog License’s non-compete and license-key terms under ELv2. They contend “source-available” is vague, while fair source provides clearer terms and fits Bear’s manifesto, ensuring Bear’s continuity even if Bear PTY LTD dissolves, codified via DOSP. The author notes involvement with fair source and helped write the FCL.