Published on May 09, 2026 at 18:01 CEST (UTC+2)
Internet Archive Switzerland (230 points by hggh)
Internet Archive Switzerland
This article introduces Internet Archive Switzerland, an independent non-profit foundation based in Sankt Gallen dedicated to preserving digital information for universal access. It highlights the fragility of digital content due to format changes, storage failures, deletions, and paywalls. The foundation is launching two key initiatives: the Gen AI Archive (partnering with University of St. Gallen to preserve current AI models for future generations) and the Endangered Archives initiative (rescuing vulnerable collections from conflict and suppression).
PipeDream on the Acorn Archimedes (17 points by msephton)
PipeDream on the Acorn Archimedes
The article recounts the development of the Acorn Archimedes computer, its novel 32-bit RISC processor (the ARM chip), and the bespoke operating system and productivity suite that ran on it. It describes how this combination, while ultimately a commercial dead-end, produced components that individually achieved lasting impact—especially the ARM architecture that later dominated mobile computing. The piece uses this history to explore the serendipity and fragility of early computing ecosystems.
Google broke reCAPTCHA for de-googled Android users (1283 points by anonymousiam)
Google broke reCAPTCHA for de-googled Android users
Google has tied its new reCAPTCHA system to Google Play Services, requiring Android users to run proprietary Google software (version 25.41.30+) to pass verification. When flagged as suspicious, users must scan a QR code that requires Play Services to communicate with Google servers, making it impossible for de-Googled ROMs like GrapheneOS to pass. The change is part of Google Cloud Fraud Defense, but critics see it as a move to enforce proprietary surveillance as a prerequisite for proving humanity.
LLMs Corrupt Your Documents When You Delegate (131 points by rbanffy)
LLMs Corrupt Your Documents When You Delegate
This paper introduces the DELEGATE-52 benchmark to test LLMs in long, delegated document-editing workflows across 52 professional domains. Experiments with 19 LLMs (including frontier models like Gemini 3.1 Pro, Claude 4.6 Opus, and GPT 5.4) show that current models corrupt an average of 25% of document content over extended interactions. Agentic tool use does not help, and degradation worsens with larger documents, longer interactions, or distractor files—raising serious concerns about trust in AI-assisted work.
Using Claude Code: The unreasonable effectiveness of HTML (301 points by pretext)
Using Claude Code: The unreasonable effectiveness of HTML
This post (a tweet from @trq212) appears to discuss the surprising utility of HTML when using Claude Code for AI-driven programming tasks. While the tweet content itself is inaccessible due to JavaScript requirements, the high score (301 points) suggests it resonated with developers who find that simple markup languages like HTML remain highly effective as an intermediate representation for AI code generation and debugging workflows. (Note: The actual tweet text was not retrievable due to X.com restrictions.)
How LEDs are made (2014) (63 points by smig0)
How LEDs are made (2014)
This SparkFun tutorial provides a detailed tour of a Chinese LED factory, showing the entire manufacturing process from lead frames and LED dies to automated bonding and encapsulation. It covers the raw materials (e.g., 4,000 dies for ~$12.50), the machinery used, and quality control steps. The article offers a rare behind-the-scenes look at how consumer electronics components are mass-produced with surprising precision and low cost.
A recent experience with ChatGPT 5.5 Pro (489 points by alternator)
A recent experience with ChatGPT 5.5 Pro
Mathematician Timothy Gowers reports that ChatGPT 5.5 Pro produced PhD-level mathematical research in about an hour with minimal human input. He notes that LLMs have moved beyond merely retrieving known answers to discovering novel, simple arguments that human mathematicians missed—especially for problems that had received little attention. This experience forces a significant upward revision of what LLMs can achieve in rigorous mathematical reasoning.
Removing fsync from our local storage engine (32 points by zzsheng)
Removing fsync from our local storage engine
The author describes building a single-node KV storage engine that avoids fsync on PUT/DELETE by using fixed-size preallocated files, O_DIRECT writes, and a journal aligned to SSD atomic-write units. Benchmarking on AWS NVMe shows ~65% higher throughput versus ext4+O_DIRECT+fsync. The trade-off is a narrower durability contract (SSD-only, own allocation and recovery), making it unsuitable for general POSIX semantics but highly efficient for specific use cases.
Mythical Man Month (262 points by ingve)
Mythical Man Month
Martin Fowler revisits Fred Brooks’s 1975 classic, highlighting enduring lessons like Brooks’s Law (“Adding manpower to a late project makes it later”) and the central importance of conceptual integrity in system design. The article emphasizes that simplicity and straightforwardness—how easily components compose—are key to managing complexity. While some aspects are dated, the principles remain vital for modern software project planning.
America's carpet capital: an empire and its toxic legacy (108 points by rawgabbit)
America's carpet capital: an empire and its toxic legacy
This investigative piece by AP and Atlanta Journal-Constitution details how decades of carpet manufacturing in Dalton, Georgia, released PFAS “forever chemicals” into the Conasauga River and surrounding environment. It describes the industry’s reliance on 3M’s Scotchgard, the corporate battles to avoid regulation, and the lasting health and environmental consequences. The story serves as a cautionary tale about industrial pollution and regulatory capture.
AI trust and reliability are the next critical bottleneck
The DELEGATE-52 paper (article 4) reveals that even frontier LLMs corrupt 25% of document content in delegated workflows. This underscores a fundamental challenge: as AI moves from chat assistants to autonomous agents, the lack of faithful execution undermines trust. Why it matters: Businesses adopting “vibe coding” or AI-driven document workflows risk introducing systematic errors that are hard to detect. Implication: Organizations must implement verification layers (e.g., diff-based sanity checks, human-in-the-loop validation) and demand better transparency from model providers. The trend toward agentic tool use does not automatically solve reliability—it may even amplify errors.
LLMs are crossing the threshold into novel scientific discovery
Gowers’s experience with ChatGPT 5.5 Pro (article 7) demonstrates that LLMs can now produce original PhD-level math, not just retrieve known results. This accelerates a shift from AI-as-tool to AI-as-collaborator in research. Why it matters: It challenges the notion that LLMs lack reasoning; they can spot simple arguments missed by humans. Implication: Researchers should treat LLMs as potential co-authors for low-attention problems, but also need new norms for attribution and verification. The “easy” unsolved problems may dry up quickly, forcing a re-evaluation of what constitutes human expertise.
Digital preservation of AI models becomes an urgent priority
Internet Archive Switzerland’s Gen AI Archive (article 1) highlights that today’s AI models are ephemeral—proprietary, rapidly superseded, or lost due to paywalls and format changes. Why it matters: Future historians and researchers will need access to historical model weights, architectures, and training data to understand AI evolution. Implication: The AI community should adopt open standards for model archiving and support initiatives like the Gen AI Archive. This is analogous to preserving source code for software history—neglecting it now will create a “dark age” of AI.
Privacy and anti-surveillance resistance are becoming AI/ML design constraints
Google’s reCAPTCHA change (article 3) forces de-Googled users to fail verification, effectively mandating proprietary software as proof of humanity. As AI agents become indistinguishable from humans, such “trust” mechanisms risk becoming surveillance gateways. Why it matters: The trend to tie AI fraud defense to proprietary stacks (e.g., Google Play Services) entrenches monopolies and excludes privacy-conscious users. Implication: Developers of AI-based verification systems should explore decentralized, privacy-preserving alternatives (e.g., zero-knowledge proofs, hardware attestation). Otherwise, we risk a future where proving you are human requires submission to corporate surveillance.
The “unreasonable effectiveness” of simple representations in AI coding
The high engagement around “Using Claude Code: The unreasonable effectiveness of HTML” (article 5) suggests developers are discovering that AI code generators work best when outputs are constrained to simple, well-structured formats like HTML. Why it matters: Complex languages or frameworks introduce ambiguity that degrades model performance. Implication: For AI-assisted programming, designing prompts and outputs around minimal, composable representations (e.g., HTML, JSON, markdown) improves reliability. This aligns with the conceptual integrity principle from the Mythical Man Month (article 9)—simplicity in interfaces reduces cognitive load for both humans and models.
AI model degradation scales with workload complexity—not just model size
The DELEGATE-52 results show that corruption worsens with document size, interaction length, and distractor files—even for top-tier models. Why it matters: This challenges the assumption that larger or more advanced models are inherently robust over long horizons. Implication: When deploying LLMs for long-running tasks (e.g., code reviews, document editing), users should limit run length, use checkpointing, and design workflows that reset context periodically. Model providers need to publish benchmarks that simulate extended delegation, not just single-turn accuracy.
Hardware-aware system design is enabling performance breakthroughs outside AI
The fsync-free storage engine (article 8) achieves 65% higher throughput by exploiting SSD atomic-write units and preallocation. While not directly AI-related, this trend parallels AI hardware co-design (e.g., custom chips for inference). Why it matters: As AI workloads demand more efficient data pipelines, system designers are rediscovering that relaxing POSIX semantics can yield massive gains. Implication: AI infrastructure (e.g., vector databases, model serving caches) can benefit from similar narrow-contract designs. The trade-off—reduced generality for speed—requires careful documentation but offers a blueprint for specialized AI storage layers.
Analysis generated by deepseek-reasoner