Hacker News Top 10
English Edition

1. Total Annihilation, rebuilt for modern computers (314 points by jttnr)

   Nanolathe is an open-source 2.5D real-time strategy engine that rebuilds Total Annihilation for modern computers. It offers a browser demo with three original demo missions and a beta for Windows, Mac, and Linux using the user’s own game files. The engine uses a modern GPU renderer with light, glow, and richer detail, while allowing players to switch to the classic look. Its MIT license and documented formats are meant to let people study, modify, and extend the engine.

2. Cartographers of North Korea (153 points by edent)

   This is an interactive web project titled “Cartographers of North Korea.” The preview says JavaScript must be enabled to run the app, so the static text reveals little about its content. Based on the title, it likely explores mapping, cartography, and representations of North Korea, possibly through historical, political, or data-driven lenses. Without running the JavaScript app, specific claims and visualizations cannot be summarized from the provided excerpt.

3. Recursive self-improvement through collective intelligence (26 points by hardmaru)

   MASS is a Sakana AI research project on recursive self-improvement through multi-agent self-supervision. It treats AI as both optimizee and optimizer/evaluator, repeatedly optimizing multi-agent workflows and distilling collective intelligence into a supervision signal. The method aims to improve score per output token over a base model on research benchmarks. The authors frame it as a step toward continual self-learning where the optimizee is the evaluator and the optimizer is the optimizee.

4. The gap in shared understanding of LLM capability is widening (177 points by albertzeyer)

   Andrej Karpathy revisits an April tweet to argue that the gap in shared understanding of LLM capability is widening. He estimates roughly 6 billion people have barely encountered LLMs, 1–2 billion are casual free-tier users who see limited chatbots, and about 20 million professionals in math and code see frontier agents complete complex projects from prompts. He notes that even many executives and non-technical professionals understand the shift only second-hand. The result is a sharp funnel rather than two groups simply talking past each other.

5. Google Data Centers around the world (27 points by andsoitis)

   This Google page maps the locations of Google data centers worldwide and indicates which are in development. It describes these facilities as where the internet lives, operating 24/7 to keep Google products running. The page includes structured metadata for locations such as Andhra Pradesh, Armstrong County, Berkeley County, and Canelones. It serves as a public inventory of Google’s global compute infrastructure footprint.

6. So, You’d Like To Sell Your Print Issues Of The Onion? (220 points by mooreds)

   The Onion has published a call for readers to help complete its print archive by donating or selling missing issues. It explains how to identify issues using volume, issue number, and date, with examples and special numbering cases. It lists specific volumes and issues it is seeking from 1988 through 2013. Interested holders are asked to submit a Google Form so an archivist can follow up.

7. Fingerprints Are the New Cookies (19 points by royfish)

   The blog post demonstrates how browser fingerprinting can act as a persistent identifier without cookies, even across sessions or incognito in some cases. It argues that fingerprinting combined with account data and data brokers can create detailed profiles used for ads and potentially sold to police or governments. It says claims of anonymization are weak because patterns can be re-identified. It notes that users who disable JavaScript avoid this specific trick but remain vulnerable to network-based methods.

8. Chapter 1 of Ways of Seeing by John Berger (63 points by rafael859)

   Chapter 1 of John Berger’s Ways of Seeing argues that seeing comes before words and continues to exceed language. Berger claims that what we see is shaped by what we know or believe, using examples such as Magritte and historical beliefs about hell and fire. He stresses that vision is not a mechanical stimulus-response process but a contextual, socially and personally conditioned act. The chapter establishes a foundational argument in visual culture about the unsettled relation between sight and knowledge.

9. Terence Tao: Math 2.0 [pdf] (294 points by Anon84)

   The link is a PDF slide deck titled “Math 2.0” by Terence Tao, apparently for a Caltech 2026 talk. The provided preview contains only PDF metadata and compressed stream objects, not readable slide text. The title likely refers to how mathematics is evolving with computation, collaboration, and AI-assisted tools. It may discuss proof assistants, formalization, large-scale collaboration, and changing research norms, but the specific arguments cannot be verified from the excerpt.

10. Passkeys are a trap. Do not use them (138 points by laurex)

    This opinion piece argues that passkeys are a trap and advises non-technical web users not to adopt them. It concedes passkeys can be more secure and convenient than passwords, but says they are not designed to prioritize user interests. Unlike passwords, authenticator apps, or physical security keys, passkeys include a hidden control that lets websites gradually remove user autonomy over credentials. The author warns of platform dependence, lock-in, and reduced portability or recovery options.

AI/ML Insights & Trends

1. Recursive self-improvement is moving from thought experiment to multi-agent engineering. The MASS project proposes optimizing multi-agent workflows and distilling their collective outputs as self-supervision for the base model, effectively making the system both optimizee and evaluator. This matters because it points toward training loops that need less human-labeled data and can improve per-token efficiency on research benchmarks. But it also increases risks of feedback loops, evaluator bias, reward hacking, and diversity collapse. Actionable takeaways: build independent verifiers and holdout tasks, monitor model collapse, maintain provenance of self-generated data, and use sandboxed, auditable RSI pipelines.

2. The LLM capability gap is widening into a stratified funnel. Karpathy’s estimates separate roughly 6 billion people with almost no LLM contact, 1–2 billion casual free-tier users, and about 20 million professional math/code users seeing frontier agents do weeks or months of work. This matters because capability perception, safety expectations, and economic impact differ radically across these groups. Policy and product decisions based on one tier will misread the others. Takeaways: segment evaluations by user tier, improve onboarding and education, measure real productivity effects, and design transparent capability reporting rather than assuming broad shared understanding.

3. Math and code are becoming the proving ground for AI reasoning and collaboration. Terence Tao’s “Math 2.0” talk and MASS both point to AI-assisted formal reasoning, proof assistants, and agentic workflows in verifiable domains. Math and code matter because they provide objective feedback and high-signal benchmarks for reasoning, making them likely early adopters. Takeaways: invest in formal verification, proof-assistant integration, benchmark contamination controls, and new norms for authorship and credit when AI contributes to discoveries.

4. Compute infrastructure is a strategic bottleneck and geopolitical asset. Google’s data center location page shows a global, always-on footprint with sites in development across continents. This matters because frontier training and inference depend on compute capacity, energy, latency, water, land, and regulatory environments. Takeaways: AI roadmaps must account for energy and capacity constraints, sovereign AI requirements, edge inference, and environmental transparency. Organizations should treat compute supply chain as a first-class risk.

5. Privacy and identity are being re-architected around AI-era data aggregation. The fingerprinting article shows cookies being replaced by persistent browser/device identifiers, while the passkeys article warns that authentication is centralizing in platform-controlled credentials. Together they matter because AI systems thrive on large-scale behavioral data, and identity is the key that links data across contexts. Takeaways: adopt privacy-preserving ML, differential privacy, federated learning, and transparent data governance; support passkey portability, recovery, and open standards; and threat-model both adtech profiling and account lock-in.

6. Open-source stewardship and archives are critical counterweights to enclosure. Nanolathe rebuilds a classic game engine under an MIT license; The Onion seeks missing print issues for archival; Cartographers of North Korea maps a closed society. These projects matter to AI/ML because reproducible code, documented formats, and preserved data are foundations for auditability, training data provenance, and long-term research. Takeaways: fund digital archives and open licenses, document data and code formats, support clean-room reimplementation, and treat data curation as infrastructure rather than an afterthought.

7. Human perception and framing will shape AI adoption as much as raw capability. Ways of Seeing argues that sight is conditioned by knowledge, belief, and context; this applies directly to multimodal AI, generated images, and AI-mediated interfaces. It matters because users do not interpret model outputs neutrally, and trust depends on context, provenance, and visual literacy. Takeaways: evaluate AI outputs for cultural and contextual meaning, provide provenance and uncertainty cues, study user perception across populations, and avoid treating visual or textual output as self-evident truth.

Analysis generated by deepseek-reasoner