Published on June 25, 2026 at 18:01 CEST (UTC+2)
Show HN: I made Google Trends for Hacker News by indexing 18 years of comments (171 points by ytkimirti)
Show HN: I made Google Trends for Hacker News by indexing 18 years of comments
This tool, Hacker News Trends, lets users search and chart how any topic, tool, or person has appeared across 18 years of Hacker News comments and posts. It indexes 45 million items and uses Upstash Redis Search to generate live date-histograms, with popular comparisons like "vercel vs cloudflare" and "openai vs anthropic." The project overlays multiple terms to show their relative traction over time, and users can click individual months to filter the underlying stories. It was built by the submitter and is open-source on GitHub.
Ford rehires 350 engineers after AI fails to preserve expertise or train juniors (177 points by alanwreath)
Ford rehires 350 engineers after AI fails to preserve expertise or train juniors
Ford brought back 350 quality inspectors after an AI system intended to replace them failed to capture tacit knowledge and adequately train new hires. The AI could not replicate experienced engineers’ judgment in detecting subtle defects, leading to quality issues. This reversal highlights the limits of AI in roles requiring deep domain expertise and mentorship.
You can't unit test for taste (144 points by kalli)
You can't unit test for taste
The author describes building a running app that plots virtual routes and wanted to enrich maps with points of interest. They initially relied heavily on LLMs (Claude) to generate data, but found the AI introduced hallucinations and biased selections. Ultimately, they used AI only as a supporting tool alongside established datasets like GeoNames, concluding that subjective “taste” – such as what makes a location interesting – cannot be reliably automated or tested with code.
Zig's New BitCast Semantics and LLVM Back End Improvements (46 points by kouosi)
Zig's New BitCast Semantics and LLVM Back End Improvements
This devlog details changes to Zig’s @bitCast semantics and improvements to the LLVM backend, especially around integer lowering. The team moved away from using LLVM’s arbitrary bit‑int types because they caused missed optimizations and miscompilations. The fix aligns Zig’s IR more closely with what Clang emits, improving performance and reliability.
Half-Life 2 in a Browser (506 points by panza)
Half-Life 2 in a Browser
A project enables the classic game Half‑Life 2 to run entirely inside a web browser, likely via WebAssembly and WebGL. It garnered 506 points on Hacker News, indicating strong interest in browser‑based gaming and WebAssembly’s ability to port complex native applications.
Apple increases MacBook and iPad prices by 20% (69 points by bazzmt)
Apple increases MacBook and iPad prices by 20%
Apple raised prices on MacBooks and iPads by approximately 20%, as reported by the Financial Times. The increase may reflect rising component costs, inflation, or supply chain pressures. The news sparked discussion on HN about hardware affordability and Apple’s pricing strategy.
Show HN: Turn native language audio into flashcards and shadowing practice (39 points by alder)
Show HN: Turn native language audio into flashcards and shadowing practice
LingoChunk is a tool that converts native‑language audio (e.g., from podcasts or videos) into flashcards and shadowing exercises for language learners. It aims to leverage spoken content for immersive practice, generating spaced repetition material from real‑world audio.
Anthropic says Alibaba illicitly extracted Claude AI model capabilities (651 points by htrp)
Anthropic says Alibaba illicitly extracted Claude AI model capabilities
Anthropic accused Alibaba of illegitimately copying capabilities from its Claude AI model, likely through model extraction or knowledge distillation attacks. This is a major incident highlighting the security risks around proprietary AI models. The story scored 651 points, reflecting the community’s deep concern about model theft and intellectual property in AI.
Ask HN: What surprised you about Estonia e-Residency and running an Estonian OÜ? (54 points by jvilalta)
Ask HN: What surprised you about Estonia e-Residency and running an Estonian OÜ?
The post solicits real‑world experiences from e‑residents running an Estonian company (OÜ). Commenters warn that the company may still be taxed in the resident’s home country under tie‑breaker rules, leading to dual accounting and bureaucracy. Many found the program overhyped for remote SaaS businesses, with hidden compliance costs.
LastPass notifies users of yet another data breach (287 points by mooreds)
LastPass notifies users of yet another data breach
LastPass informed users of a breach via a third‑party market research firm (Klue), which exposed CRM data (names, emails, addresses) and support case information. The company said password vaults were not affected but advised users to watch for phishing. This marks another security incident for LastPass, eroding user trust.
AI cannot replace deep expertise or mentorship
Ford’s rehiring of engineers after an AI system failed to preserve institutional knowledge shows that current AI lacks the ability to transfer nuanced, tacit skills to junior workers. Businesses relying on AI for knowledge management must anticipate gaps and maintain human‑in‑the‑loop processes.
Model extraction is a growing security and IP threat
Anthropic’s accusation against Alibaba underscores that large language models are vulnerable to illicit copying. As AI becomes more valuable, model theft through API probing or distillation will increase. Companies need stronger guardrails, legal protections, and technical defenses (e.g., watermarking, query rate limits).
AI taste and judgment remain unreliable
The “You can't unit test for taste” article illustrates that LLMs struggle with subjective, context‑dependent quality decisions (e.g., choosing interesting points of interest). AI can augment but not replace human curation when “good taste” is required. Developers should treat AI outputs in creative or subjective domains as draft suggestions, not final answers.
WebAssembly and browser‑based AI inference are accelerating
The Half‑Life 2 in a browser demo, combined with growing interest in running models client‑side (e.g., via WebGPU or ONNX), signals a trend toward moving compute from servers to the edge. This could enable privacy‑preserving AI applications but requires efficient model quantization and streaming.
LLM‑powered tools are being embedded in niche workflows
Projects like LingoChunk (language learning) and Hacker News Trends (analytics) show that AI is increasingly used to build specialized, user‑facing tools rather than just chatbots. The next wave may be small, focused AI applications that solve specific problems with minimal hallucination risk.
AI‑assisted code generation is improving but still needs manual oversight
Zig’s compiler team used LLVM backend improvements to fix miscompilations caused by poorly tested code paths. While AI can help write code (e.g., Zig’s devlog likely involved LLM‑assisted drafting), human expertise remains critical for correctness and performance optimization.
The AI hype cycle is producing both innovation and backlash
The high scores for both the Anthropic model‑theft story (651) and the Ford‑rehiring story (177) indicate that the HN community is closely watching AI’s failures as well as its breakthroughs. There is a growing demand for realistic assessments of AI capabilities, rather than uncritical promotion.
Analysis generated by deepseek-reasoner