Hacker News Top 10
English Edition

1. Type-safe HTML templating in Zig (116 points by lukaslalinsky)

   Lukáš Lalinský describes his experience building a type-safe HTML templating system in Zig after years of web development in PHP, Python, Perl, and JavaScript frameworks. He argues that Zig's comptime, arena allocators, and performance model make low-level web development practical, especially for API backends and server-rendered apps. The post explains how his libraries zio and Dusty evolved toward web use, and how htmx or Datastar can enable SPA-like behavior without JavaScript/TypeScript. The main focus is a templating approach that is both high-level and type-safe while retaining Zig's efficiency.

2. Fingerprints are the new cookies (378 points by royfish)

   "Fingerprints are the new cookies" is an interactive essay demonstrating how browser and network fingerprinting can identify users even after cookies are cleared or a VPN is used. It assigns visitors a persistent nickname and explains how that identity can remain consistent across sites, devices, and browsing modes in many cases. The author warns that such fingerprints can be linked with account data, ad profiles, and even conversations with AI systems to build detailed behavioral profiles. The piece argues that the real problem is not personalized ads but pervasive surveillance and the sale of inferred identity to advertisers or governments.

3. Postgres isn't slow, your storage is (85 points by saisrirampur)

   This ClickHouse blog post recaps a POSETTE 2026 talk by Sai Srirampur titled "Postgres Isn't Slow. Your Storage Is." It examines common scaling symptoms such as slow ingestion, unpredictable P95 read latency, autovacuum lag, checkpoint contention, and logical replication lag. The core benchmark compares the same 3.3-billion-row PostgreSQL workload on baseline gp3 EBS storage versus local NVMe. The post concludes with proposed production patterns for running PostgreSQL on local NVMe to avoid storage-induced bottlenecks.

4. Total Annihilation, rebuilt for modern computers (516 points by jttnr)

   Nanolathe is an open-source, MIT-licensed 2.5D RTS engine that rebuilds Total Annihilation for modern computers. The site offers a browser demo with three original demo missions and a native beta for Windows, Mac, and Linux, using game files the player already owns. It features modern GPU rendering with light, glow, and richer detail, plus a switchable Classic look. The project emphasizes clean-room research, documented formats, and an engine that others can study, modify, and extend.

5. Calculus with Julia (130 points by subset)

   "Calculus with Julia" is a set of online notes for learning calculus using the Julia programming language. It follows the "rule of four" by including graphical, numerical, algebraic, and verbal perspectives. Unlike traditional computer algebra system approaches, the notes primarily use Julia for numerical computing and illustration rather than symbolic manipulation. It also points readers to resources for installing Julia and reviewing different ways to interact with it.

6. Running the unmodified Linux amdgpu driver on an iPad (23 points by joshcsimmons)

   LemonSeed Studio is an iPad IDE that runs local LLM inference on an AMD GPU connected over Thunderbolt. It bundles the unmodified upstream Linux amdgpu and amdkfd driver as a PCIDriverKit extension, allowing the iPad app to drive the external AMD GPU. The project also includes LemonSeed Engine, an inference engine using the GPU through its HSA runtime and HRX/Loom, plus amdgpu_mtopg for live GPU telemetry. This is a notable experiment in combining Apple mobile hardware, external GPUs, and Linux driver code for on-device AI.

7. Fable Took My Job (119 points by jllyhill)

   "Fable Took My Job" is a personal essay about being laid off after the author's company adopted an LLM marketed as "Fable / Mythos tier." The author says the model was not necessarily better at the job, but it was cheaper than his compensation and gave leadership enough confidence to coast toward M&A. He describes the emotional toll of layoffs, including knowing that two of his reports would lose their jobs and later having to execute those decisions. The essay reflects grief, fear, and doubt about the future of engineering work in an AI-driven cost-cutting environment.

8. Whatever became of holography? (2011) (133 points by cainxinth)

   The preview for "Whatever became of holography?" provides no article content, so this summary is based on the title and publication context. It is an American Scientist piece from 2011 that likely revisits holography's history and asks why it never became as ubiquitous as once promised. The article probably discusses technical, commercial, and practical limitations that kept holography niche despite its cultural appeal. No specific claims from the article can be verified from the provided preview.

9. Cartographers of North Korea (2019) (264 points by edent)

   "Cartographers of North Korea" is an interactive web project whose preview only says JavaScript must be enabled. Based on the title, it likely explores the people and methods involved in mapping North Korea, a country with severe restrictions on geographic information. Such a project may use satellite imagery, interviews, and visualization to document how borders, roads, and facilities are reconstructed. Because the content is not available in the preview, no detailed summary of its arguments or features is possible.

10. Zod 4.6.6 compromised by malware? (122 points by luggage_bazooka)

    The linked page is a GitHub commit in the Zod repository titled "Fix bugs and improve performance," but the Hacker News title asks whether Zod 4.6.6 was compromised by malware. The diff changes the release workflow so that pushes to main affecting packages/zod/package.json or packages/mini/package.json can trigger publishing, whereas previously a maintainer had to run the release workflow manually. It also bumps @zod/mini from version 4.6.5 to 4.6.6 in jsr.json. The concern is that this release-automation change could be part of a supply-chain compromise or an unsafe publishing path.

AI/ML Insights & Trends

1. Local and edge LLM inference is moving to unconventional hardware and OS boundaries.
2. Trend/insight: Article 6 shows an iPad IDE running local LLM inference on an external AMD GPU over Thunderbolt, with the upstream Linux amdgpu/amdkfd driver embedded as a DriverKit extension.
3. Why it matters for AI/ML development: This pushes AI inference away from centralized cloud GPUs toward personal, offline, and latency-sensitive devices. It also shows that the edge AI stack is not settled: mobile OS sandboxing, external GPUs, and Linux driver code are being mixed in novel ways.

4. Implications/actionable takeaways: Expect more AI apps to bundle portable inference runtimes, driver shims, and telemetry. Developers should track privacy and security risks of kernel-adjacent code on consumer devices, while hardware vendors may face pressure to support open compute paths on tablets and phones.

5. AI labor displacement is being framed through cost and M&A, not just capability.

6. Trend/insight: Article 7 describes a company replacing its last engineer with an LLM because the model was cheaper and gave leadership confidence to coast toward M&A, even if it was not clearly better.
7. Why it matters for AI/ML development: AI adoption decisions often depend on financial incentives, investor narratives, and organizational politics. This can eliminate roles before robust evaluation, knowledge transfer, or safety practices are in place.

8. Implications/actionable takeaways: Teams should invest in documentation, evaluation harnesses, and human-in-the-loop workflows. Policymakers and leaders need better measures of AI's actual productivity and risk, not just cost savings, to avoid brittle systems and lost institutional knowledge.

9. Persistent fingerprinting will feed AI profiling and personalization systems.

10. Trend/insight: Article 2 shows browser and network fingerprinting can identify users across cookie clears, VPNs, and devices, then link that identity to accounts, ads, and even AI conversations.
11. Why it matters for AI/ML development: AI/ML models thrive on linked behavioral data. The more reliably users can be identified, the more training, targeting, and surveillance can be done without explicit consent.
12. Implications/actionable takeaways: Privacy-preserving

Analysis generated by deepseek-flash