HN Evening Brief - March 13, 2026
Welcome to today’s Hacker News evening briefing! Here’s a roundup of the top 30 stories from March 13, 2026, covering AI, security, geopolitics, tech tools, and more.
AI & Tech Policy
Can I run AI locally?
A new website that helps users determine if they can run AI models locally on their hardware. The tool estimates performance based on model size, memory requirements, and hardware capabilities, providing guidance for those interested in running AI without cloud services.
HN Discussion: Users shared their experiences with local AI models, noting that small models like Qwen 3.5:9b are excellent for embedded applications and tool use. Discussion also covered the challenges of finding the highest-quality model for specific hardware constraints, with some noting the tool’s estimates work well for dense models but may not account for Mixture of Experts (MoE) architectures that use active parameters differently. Commenters debated the importance of tokens per second vs. quality, and expressed interest in seeing latency metrics for prompt processing time.
John Carmack about open source and anti-AI activists
John Carmack weighed in on the debate around open source and AI training, sharing thoughts on how companies training on open source code interact with the original authors and the broader AI landscape.
HN Discussion: The discussion highlighted the tension between AI companies profiting from training on open source code and the original authors’ intentions. Some noted that anti-AI sentiment in open source communities stems more from AI-generated content clogging feedback channels than from concerns about code training. Others pondered how historical figures like Stallman might have approached reverse engineering and software rights in today’s AI landscape.
Show HN: Context Gateway – Compress agent context before it hits the LLM
An open-source proxy that compresses tool outputs from coding agents before they enter the LLM’s context window. Using small language models (SLMs) to detect which parts of context carry the most signal for a given task, the system reduces token costs while maintaining quality. Features include spending caps, session tracking dashboards, and Slack notifications.
HN Discussion: Commenters were intrigued by the intent-conditioned compression approach, noting it differs from naive truncation or generic summarization. Several requested evals showing how compression affects accuracy on benchmarks like SWE-bench. Questions arose about handling ambiguous tool calls, latency overhead, and how often the expand() function needs to recover stripped content. Some expressed concern about external tooling modifying their context, while others noted similar features should be built into agent frameworks rather than standalone products.
Launch HN: Captain (YC W26) – Automated RAG for Files
Captain automates the building and maintenance of file-based RAG pipelines, indexing cloud storage like S3 and GCS, plus SaaS sources like Google Drive. The system handles ETL, chunking, embedding, storage, search, and re-ranking with a single API call. A demo site allows querying Paul Graham’s essays to test the experience.
HN Discussion: Users noted the crowded RAG pipeline market and discussed various techniques tried and what actually works. Security concerns were raised about uploading company documents to a new SaaS provider. Questions focused on citation tracing back to source documents, handling structured data like CSVs, and processing different content types. Some shared similar open-source approaches they’d built for enterprise use.
Launch HN: Spine Swarm (YC S23) – AI agents that collaborate on a visual canvas
A multi-agent system operating on an infinite visual canvas for complex non-coding projects like competitive analysis, financial modeling, and pitch decks. The system moves beyond linear chat interfaces, allowing agents to work in parallel on visual blocks representing different tasks. Benchmarks show 87.6% accuracy on Google DeepMind’s DeepSearchQA with zero human intervention.
HN Discussion: Users appreciated the canvas approach for organizing complex, multi-step work compared to linear chat threads. Some hit the token cap before seeing their first task complete, noting the free tier was insufficient. Suggestions included making navigation more intuitive with mouse controls, adding middle-click canvas dragging, and providing clearer screenshots rather than embedded videos. Questions focused on GitHub integration, sharing capabilities, local/self-hosted options, and whether users could use their own API keys.
Executing programs inside transformers with exponentially faster inference
Research showing how transformers can execute programs with exponentially faster inference by restricting lookup heads to head dimension 2. This enables decoding in log time relative to sequence length for structured execution, with potential applications as fast-path models paired with slower general models or as speculative execution engines.
HN Discussion: Commenters debated whether running programs inside LLMs is desirable when specialized circuits can do computation more efficiently. Others saw value in bootstrapping training using expert systems that classify correctly 80% of the time. Discussion covered the clever hullkv mechanism, why the strategy applies to “code tokens,” and questions about WASM as a target. Some noted the article was light on training details like what was trained on and how.
Security & Privacy
The Wyden Siren Goes Off Again: We’ll Be “Stunned” by NSA Under Section 702
Senator Ron Wyden again warns about NSA surveillance activities under Section 702, with the interpretation of the law apparently classified. The article discusses ongoing concerns about government surveillance and what citizens might be shocked to learn about intelligence gathering practices.
HN Discussion: Commenters expressed frustration with classified interpretations of laws meant to govern public behavior. Discussion focused on whether it matters what the interpretation is when the track record of data misuse exists. Some shared experiences of personal data exposure despite privacy measures, while others debated whether being “normal” and blending in is the best approach to privacy given the inevitability of data breaches.
Meta Platforms: Lobbying, dark money, and the App Store Accountability Act
An investigation into Meta’s lobbying efforts around the App Store Accountability Act, revealing use of dark money and digital childhood alliance astroturfing. The researcher, using Claude Opus for analysis, documents how Meta pushed age verification legislation through state-level processes to force Apple to build complex identity verification infrastructure.
HN Discussion: Some viewed this as retaliation for Apple’s ATT rollout that cost Meta billions by cutting off their data pipeline. Others questioned the reliability of AI-assisted research completed in just two days. Discussion covered whether Meta gains advantage from OS-level age verification or if they’d prefer platform-level KYC, concerns about biometric data collection, and the broader implications of infrastructure that provides OS-level signals about users.
E2E encrypted messaging on Instagram will no longer be supported after 8 May
Instagram announced that end-to-end encrypted messaging features will be retired on May 8, 2026. The opt-in feature, similar to Telegram’s OTR chat, is being discontinued despite previous pushes toward E2E encryption across Meta’s platforms.
HN Discussion: Some noted this was an opt-in feature rather than default like WhatsApp, so fewer users will be directly affected. Others questioned the timing and incoherence of the announcement alongside continued E2E promotion elsewhere. Discussion covered whether this admission reflects how things always were in practice given Meta’s control of both clients and servers, with truly safe E2E requiring independent open-source clients.
Bucketsquatting is (finally) dead
AWS introduced account-regional namespaces for S3 buckets, effectively ending the practice of bucketsquatting where attackers would register bucket names before legitimate users to intercept data or create confusion. The change provides unique namespaces per account and region.
HN Discussion: Commenters celebrated the long-overdue security improvement. Some shared stories about AWS account deletion quirks, like being unable to reuse root user email addresses after account deletion. Discussion covered similar naming challenges in Azure Blob Storage and suggestions for naming schemes like Discord’s tag-based system to democratize namespace access and prevent squatting.
Geopolitics & War
Militaries are scrambling to create their own Starlink
Multiple countries are rushing to develop their own satellite constellations for military communications, recognizing dependence on commercial Starlink as a strategic vulnerability. The article discusses China, EU, and other nations’ efforts to build independent LEO satellite networks.
HN Discussion: Commenters noted that without Falcon 9/Starship capabilities, China and EU remain 10-20 years behind SpaceX. Some mentioned the CF rebuilding HF radio expertise in Canada as a backup for large-scale conflict. Discussion covered concerns about Starlink’s solvency and long-term reliability, the first customer being the US Army, and predictions that future wars might involve Kessler syndrome from uncoordinated multi-thousand satellite constellations.
Qatar helium shutdown puts chip supply chain on a two-week clock
Qatar’s helium supply shutdown threatens semiconductor manufacturing within two weeks, highlighting the fragility of the global chip supply chain. Helium is critical for chip fabrication processes, and the shortage could impact RAM prices and broader GPU availability.
HN Discussion: Commenters noted the fragility of compute infrastructure, from GPU availability to raw materials like helium. Discussion covered whether medical imaging’s helium recycling systems could be adopted by chip manufacturing, and why current losses at fab scale still require substantial constant supply. Some noted the irony of the US recently selling its strategic helium reserves just before this crisis, while others questioned whether party balloon tanks indicate larger supplies exist.
Tech Tools & Projects
Show HN: Channel Surfer – Watch YouTube like it’s cable TV
A browser-based tool that recreates the cable TV experience for YouTube, allowing users to surf channels without recommendation algorithms. Users can import subscriptions via a bookmarklet locally, with no accounts or sign-ins required.
HN Discussion: Users appreciated the nostalgia of the cable menu experience and the ability to bypass YouTube’s insidious algorithm steering. Some noted similarities to existing projects like ytch.tv and Hypertext.tv for website surfing. One user shared their approach using yt-dlp, elfeed, and elfeed-tube for Emacs-based YouTube consumption with mpv integration.
TUI Studio – visual terminal UI design tool
A visual design tool for terminal user interfaces that lets users design TUIs in a web interface and export production-ready code for various frameworks. However, the code export feature is not yet functional, currently in active development.
HN Discussion: Some argued that adding mouse-clickable tabs, buttons, and checkboxes to TUIs makes them GUIs masquerading as TUIs, losing the appeal of text-based interfaces designed for effectiveness and power. Others criticized the complexity of a web-based TypeScript tool in Docker for designing terminal interfaces, suggesting a TUI editor should itself be a TUI. Several noted the product isn’t usable yet since code export doesn’t work, with requests for better video controls on the demo.
Show HN: Svglib a SVG parser and renderer for Windows
An SVG file parser and renderer library for Windows applications and games, using Direct2D for GPU-assisted rendering and XMLLite for XML parsing. Designed to make it easy for Win32 applications to display SVG images.
Hammerspoon
GitHub**
Hammerspoon, the macOS automation tool, allows users to script system-wide automation with Lua. The post sparked discussion about configuration examples and feature requests.
HN Discussion: Users shared their configs, with one demonstrating a simple grid-based window manager controlled by hotkeys. Others requested features like a global Teams mute toggle and discussed replicating Aerospace’s functionality within Hammerspoon rather than running both tools.
Gvisor on Raspbian
A guide to running Google’s gVisor sandbox on Raspberry Pi, providing application sandboxing for Linux containers on ARM hardware. The post covers installation steps and performance considerations.
HN Discussion: Users noted memory overhead of about 120MB per sandbox vs. 15MB for raw containers on Pi 4, limiting services to ~25 before OOM. Syscall interception adds 30-40% CPU overhead on ARM, making it suitable for untrusted Python scripts but not compute-heavy workloads. Some clarified that kernel compilation on Pi 5 takes about 40 minutes with all cores, not several hours.
Removing recursion via explicit callstack simulation
A technical post about converting recursive algorithms to iterative implementations using explicit call stack simulation, including benchmarking and property testing.
HN Discussion: Commenters appreciated the combination of benchmarking and property testing in the technical deep dive.
Okmain: How to pick an OK main colour of an image
A detailed exploration of selecting the main color from images, addressing problems with naive approaches like shrinking to a single pixel. The author uses perceptual color spaces and considers that the most representative color may not be the most frequently occurring one.
HN Discussion: Designers shared similar tools and techniques they’ve built throughout their careers. Some warned about performance concerns and security issues with loading large images into memory, recommending scaling down or sampling pixels to avoid OOM DoS vulnerabilities. Users shared links to related tools like OKPalette and discussed applications like matching Home Assistant dashboards to album art.
Web & Infrastructure
Parallels confirms MacBook Neo can run Windows in a virtual machine
Parallels has confirmed that Apple’s budget MacBook Neo can run Windows 11 in a virtual machine, making Windows virtualization accessible on Apple’s lowest-cost Mac. This potentially expands the Neo’s utility for users needing Windows applications.
HN Discussion: Users noted that Parallels may need to review licensing for the budget tier, suggesting a “Lite” license with 8GB vRAM at a lower price point. Some expressed excitement that Apple now has reason to keep macOS small given the Neo’s 8GB constraint, while others debated Linux support and whether it runs Windows better than typical corporate x86 laptops.
Your Phone Is an Entire Computer
An exploration of how modern smartphones have the computing power of full computers, yet form factor and intentional product segmentation prevent them from replacing laptops entirely. The post argues that consumers get a worse experience than technically possible due to profitability of separate devices.
HN Discussion: Users shared experiences using phones as desktop computers via Dex-like features, noting that adding Bluetooth keyboards, mice, and monitors creates a worse experience than a laptop. Some appreciated the mental break from using a flip phone instead of a smartphone, while others debated whether phone silicon differs meaningfully from laptop chips given thermal constraints.
History & Science
Lost Doctor Who Episodes Found
Two previously missing episodes of Doctor Who have been found, ending decades of uncertainty. The recovered episodes are from the Troughton era, where the BBC wiped many tapes and only about half of his run survives.
HN Discussion: Fans expressed relief that more episodes were found, particularly two good episodes from the Troughton era which set the standard for future Doctors. Some questioned how film cans survived when the missing episodes were believed to be on wiped video tapes. Others shared nostalgia for their first encounters with Daleks and the show.
The Mrs Fractal: Mirror, Rotate, Scale (2025)
A mathematical visualization exploring fractals through mirror, rotate, and scale transformations, demonstrating how iterative displacement maps create complex patterns from simple operations.
HN Discussion: Commenters enjoyed the math visualization, comparing it to camera feedback loops seen in Doctor Who opening credits. Some noted that mirror and rotate transformations are essentially three different mirrors, while others appreciated the beauty of mathematical visualizations on the web.
Revealed: Face of 75,000-year-old female Neanderthal from cave
Facial reconstruction of a 75,000-year-old female Neanderthal from Shanidar Cave reveals that while skeletal differences exist, the living face appeared more similar to humans than previously thought. The reconstruction challenges the stereotype of Neanderthals as primitive.
HN Discussion: Commenters were glad the stereotype of Neanderthals as dumb club-wielders is being laid to rest. Some questioned the legitimacy of facial reconstruction from skulls, noting the surprising presence of a chin despite skeletal differences. Others made tongue-in-cheek comments about the lack of privacy laws in the Paleolithic era.
Dijkstra’s Crisis: The End of Algol and Beginning of Software Engineering (2010)
An academic paper challenging the conventional narrative about the birth of software engineering, arguing that the “software crisis” narrative was constructed by academics to promote programming as a mathematical discipline. Dijkstra’s rebellion against Algol 68 is analyzed as an effort to replace working-class programmers with elite mathematical engineers.
HN Discussion: The paper provides a fascinating challenge to the established origin story of software engineering centered on the 1968 NATO conference. Commenters noted the irony that both Wijngaarden and Dijkstra were Dutch, and discussed the tension between Dijkstra’s theoretical ivory tower approach versus Wirth and Hoare’s practical engineering tools.
Academic & Research
Willingness to look stupid
An essay arguing that a willingness to look stupid—through asking naive questions, sharing unfinished work, and embracing failure—is a trait common to successful researchers and creators. The piece suggests that high-trust environments and reduced fear of embarrassment foster creative work.
HN Discussion: Commenters connected this to “Let’s Paint TV” and John Kilduff’s “EMBRACE FAILURE” mantra. Some noted that young people aren’t consciously willing to look stupid, they just don’t know they do. Discussion covered how metrics and surveillance kill creative work, and that truly creative work happens in high-trust societies or where resources are abundant enough to tolerate failure. Professors shared their approach of “thinking in print” and being willing to publish and correct mistakes.
Business & Industry
This is not the computer for you
A personal essay reflecting on how learning to work within hardware limitations—whether on a hand-me-down Macbook, a struggling Chromebook, or a budget device—builds fundamental skills and curiosity. The author argues that Chromebooks limit learning by making users think Google decided what they’re not allowed to do.
HN Discussion: Users shared nostalgic stories of pushing hardware to limits on garbage machines, from editing HD video at 2fps to running professional tools slowly on underpowered systems. Some felt the criticism of Chromebooks was overly harsh given alternative Linux paths, while others connected to the theme of losing the endless tinkerability of youth. One commenter hilariously noted that the described behavior looks like autism symptoms.
System Administration
No system administration stories in today’s top 30.
Other
The Bovadium Fragments: Together with The Origin of Bovadium
A review discussing previously unpublished Tolkien satire critiquing technology and industrialization, exploring how Tolkien’s luddism was shaped by his experiences and worldview.
HN Discussion: Commenters debated whether Tolkien’s idyll of pre-industrial life existed for most people in Britain. Some noted Tolkien and Lewis earned their luddism fairly after surviving trench warfare, while others discussed whether technology represents progress or degeneration. One commenter in their 60s reflected on a lifelong career with computers they couldn’t imagine being without, despite Tolkien’s likely dislike of living in 2026.
Show HN: What was the world listening to? Music charts, 20 countries (1940–2025)
A playable map of music history featuring 230 charts across 20 countries spanning eight decades. Every song is playable via YouTube or Spotify, making it easy to explore how musical trends propagated globally over time.
HN Discussion: Users loved the project, noting the browser tab would stay open for days. Some pointed out overlap with Radiooooo, a similar project allowing users to tune into music by country and year. Others noted missing years and genres in the 5-year interval data, and reported small errors in song titles for certain country/year combinations.
Thanks for reading! See you tomorrow for the next Hacker News briefing.