Published on September 07, 2026 at 06:01 CEST (UTC+2)
Keep Our Servers Running: Your Recurring Donation Goes 3X This September (39 points by sonicrocketman)
The Internet Archive is running a September donation campaign with a 2:1 matching program for recurring gifts of $25 or more. The post emphasizes the real infrastructure needs—servers, storage, power, cooling, and maintenance staff—behind its 210 petabytes of free, ad-free knowledge. Highlighting its independence from corporate cloud providers, the Archive argues that public access to knowledge requires ongoing community funding rather than commercial monetization of user data. This is a straightforward, mission-driven fundraising appeal tied to operational sustainability.
Making a Python interpreter in 1024 bytes (138 points by azhenley)
Austin Henley describes a personal weekend challenge: writing a Python interpreter in just 1024 bytes of hand-written C code, with no macros or external libraries. He quickly realized that a 512-byte attempt was too small, and that fitting all of Python was impossible, so he aimed for a readable subset that handles Python-like syntax: def, colons, indentation, and if statements without parentheses. His first attempts were little more than a calculator, which failed to capture the spirit of the exercise. The post charts his iterative progress from parsing simple arithmetic to implementing enough language structure to run a FizzBuzz program with indentation-based blocks.
The NX bit is not just about security (34 points by torutofu)
A guest post by Sonya recounts debugging a bizarre, months-long ARM64 hypervisor bug encountered while developing a bare-metal hypervisor for postmarketOS. Enabling the CTR_EL0 register intercept (the core purpose of the hypervisor) caused the phone to randomly lock up and trigger watchdog resets. The author systematically investigates hypotheses, including possible incorrect emulation of MRS/MSR special-function register access on AArch64. The article's framing suggests that the NX bit—typically discussed purely as a security mitigation—has deeper architectural implications that surface during low-level development.
It took a year to ship WebAssembly in Anubis (175 points by xena)
Anubis is an anti-bot system that protects websites from aggressive AI-company scraping using a Hashcash-style proof-of-work challenge. The author explains that it took a full year to ship the WebAssembly version of their solution. The idea is that a PoW cost that is negligible for a single human user becomes prohibitively expensive at mass-scraper scale. Anubis is described as a "placeholder solution" while more sophisticated headless browser fingerprinting techniques (e.g., font rendering analysis) are developed. The post also warns that JavaScript-disabling plugins must be turned off, noting that AI companies have "changed the social contract around how website hosting works."
Ask HN: Why don't we bring back old school OkCupid? (22 points by firefax)
An HN user asks why nobody has resurrected old-school OkCupid's match-based, questionnaire-driven approach to dating, rather than the swipe-culture paradigm. Commenters point out that dating apps are a rich-get-richer market dominated by network effects: people go where the user base is, regardless of features. Additionally, requiring users to answer many questions or invest effort upfront drives away the most desirable people, who can easily find matches elsewhere. One commenter notes that OkCupid itself only switched to swiping after A/B testing demonstrated that swiping retained users significantly better than survey-based matching. The thread ultimately concludes that the current market structure is an economic, not technological, barrier.
Your intellectual fly is open when you use an LLM to author a post (2025) (583 points by cyb0rg0)
Bryan Cantrill (formerly of Sun, Joyent, etc.) harshly criticizes LinkedIn users who author posts with LLMs, comparing such writing to having one's "intellectual fly open." He catalogs the recognizable stylistic tics of AI-generated content: gratuitous emojis, single-sentence paragraphs, "it's not just… but also" constructions, and overused em-dashes. Cantrill argues that this writing is not merely mediocre but stylistically grating, and that using AI to generate content signals an inability or unwillingness to communicate authentically. The essay functions as a broader cultural critique of the AI-content flood, delivered with characteristic bluntness.
Ask HN: Fable hacked my piano, can I release the results? (22 points by jmpman)
A PianoDisc self-playing-piano owner bought a Satie track from the company's online store, then became curious whether AI could have produced the file. Using the tools Astra and Fable in a critique loop, he spent an hour discussing rubato, fermata, solenoid response times, and sustain pedal technique—yielding a version of Gymnopédie No. 1. Fable also decoded the proprietary distribution format: the MP3's right channel carries MIDI on a 2004.5 Hz square wave. Now the user is asking HN whether he may legally release his findings about this format and AI recreation process.
Nitter and XCancel resume service after legal advice (541 points by zImPatrick)
Nitter, the privacy-preserving alternative frontend for X/Twitter, has resumed activity after receiving legal advice. X Corp. had sent cease-and-desist letters on August 24, 2026, demanding takedown of Nitter instances and its repository, but the project is now updating its README and funding options again. The commit shown restores documentation, adds a Ko-fi sponsorship link for the maintainer, and announces that service has resumed following legal consultation. This marks a significant victory for open-source projects facing legal threats from large platforms.
Show HN: Mador – Make any DOM reactive with a tiny 80-line Proxy state tuple (75 points by bosmarcel)
Mador is a minimal reactive DOM runtime consisting of approximately 80 lines that leverages a JavaScript Proxy-based state tuple. It targets developers who don't want or need a full framework: you bind state to existing DOM elements via read and update state via a write function. It is shipped as a native ES module usable through npm or directly from a CDN. The project's philosophy is "Make DOM Reactive. Nothing more," offering an alternative to frameworks like React for simple, dependency-light web pages.
Babylonian Lamb Stew with Beets (1750–1730 BCE) (118 points by yubblegum)
Yale's Babylonian Collection shares a recreation of a lamb stew with beets from a 1750–1730 BCE cuneiform recipe. The ingredients include diced lamb, rendered sheep fat, beer, water, onion, arugula, Persian shallots, cilantro, cumin, red beets, leek, garlic, and coriander seed. The cooking method involves searing the lamb, folding in vegetables, deglazing with beer, and simmering. It is part of a project to use modern culinary techniques to reconstruct ancient Mesopotamian cuisine from the collection's tablets.
Web scraping by AI companies has become an existential infrastructure problem. Anubis and Nitter both illustrate how AI-driven data collection is forcing website operators to adopt proof-of-work schemes, legal threats, and access controls that harm ordinary users. This is no longer a niche issue: maintaining anti-scraper systems is now core web operations work. The implication for AI/ML is that unfettered crawling—the foundation of most training corpora—is coming to a close. Companies designing data pipelines should expect rising legal, computational, and ethical costs per data source, and should prioritize licensed or consented data partnerships.
The identifiable "LLM voice" is now a credibility liability, not just a stylistic annoyance. Cantrill's post—at 583 points, by far the highest-scoring article here—captures a broad cultural backlash against AI-generated text. The em-dashes, single-sentence paragraphs, and formulaic rhetorical moves have become detectable tells that mark content as machine-produced and, in Cantrill's view, intellectually dishonest. This matters for AI/ML because it demonstrates that current language models have converged on a distinct, overly homogeneous output distribution that humans now recognize as low-authenticity. Model developers should treat stylistic diversity and personality preservation as a first-class alignment utility, not an afterthought, since perception of AI text affects everything from LinkedIn posts to resume screening and customer service.
LLMs are becoming capable reverse-engineering tools—and pushing IP boundaries. The piano article shows an LLM decoding a proprietary MIDI-over-audio format from raw MP3s (a 2004.5 Hz square wave on the right channel). This is an unusually sophisticated application: the model identified the encoding, extracted it, and critiqued musical performance details. With LLMs increasingly able to parse hidden data structures from binary files, the line between "interoperability" and "circumvention of protection standards" is blurring. Developers building proprietary formats should assume that their encoding schemes can be decoded by an LLM query; conversely, the open-source community gains a powerful ally for interoperability and archival freedom.
There is a growing asymmetry between AI's consumption of public knowledge and the funding of that knowledge's infrastructure. The Internet Archive stores 210 petabytes and is the kind of resource AI training pipelines rely on, yet it still has to run donation drives to fund servers, power, and staff. Meanwhile, AI companies freely consume open corpora without proportionate contribution to the institutions maintaining them. As AI scales, this imbalance becomes a systemic risk: if underfunded archives and libraries degrade or go offline, the data commons shrinks and future model training data quality suffers. A practical takeaway is that AI organizations should establish standing contributions to knowledge infrastructure (glossary: "dataset commons"), not only to be good citizens, but to protect their own long-term supply chain of human-derived knowledge.
A countervailing appreciation for small, hand-crafted code is emerging in the AI era. Two articles here celebrate tight, deterministic human engineering: a 1024-byte Python interpreter and an 80-line reactive DOM library. In a world where LLMs generate sprawling frameworks on request, these constrained projects stand out as demonstrations of human understanding, taste, and intellectual play. The HN community strongly rewards such work, and this suggests a durable niche for software that emphasizes minimalism, auditability, and elegance—properties that AI-generated code rarely provides. For AI/ML practitioners, this is a warning against assuming "more generated software" always adds value; often, the right solution is less code written with intent.
Anti-bot protection and AI detection are collapsing into a single adversarial ML problem. Anubis's roadmap—moving from proof-of-work to fingerprinting headless browsers via techniques like font-rendering analysis—shows that distinguishing human users from AI agents increasingly relies on behavioral and rendering anomalies, which are themselves best detected with ML. Conversely, AI agents will improve at mimicking human browsing. This arms race has broader implications: as more sites deploy such protections, AI agents that promise web automation (e.g., web-voyage style agents) will find robust, no-JS access shrinking. Developers of such agents must plan for challenge-solving, visual fingerprinting, and rate-limit resilience rather than naive fetching, while site operators must weigh the cost of collateral damage to legitimate assistive tools.
Analysis generated by deepseek-reasoner