Four modern coding agents (Gemini 3.8 Flash, GPT Astra 6, Opus 5 and Fable 5) were each given a $100 budget and told to build an open-source PDF editor. All produced web apps (mostly pdf.js-based) with basic editing: image insertion, signatures, comments, shapes, redaction and page reordering, but ambition and polish varied - Opus alone attempted editing embedded text, Fable shipped a very minimal highlighter, and several features common in commercial editors (OCR, vector manipulation, robust image handling, multi-selection) were missing. Usability mistakes were frequent and concrete: Gemini placed inserted images by clicking without previews or selection handles, Astra labeled every button with verbose text instead of leveraging standard iconography, apps lacked settings search, many UI controls broke on mobile, and performance regressions appeared from sequential slow backend requests. Numerous bugs persisted (skewed exported images, dark-mode color inversion, unexpected font changes, missing rotations) and none of the projects provided easy native installation or a built-in bug-reporting link.
The core argument is that these failures stem from how agents interact with software: they don’t pursue human-like, purpose-driven workflows, don’t experience friction or impatience that would expose UX flaws, and operate on static snapshots rather than real-time interaction. Because agents miss the small, iterative annoyances that shape usable interfaces, human oversight remains essential - especially for polishing UX and catching edge-case bugs - so open-source software driven by agents will likely remain “sloppy” without a human product-manager role.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
OpenAI fired three safety researchers for allegedly mishandling sensitive information, but they dispute the claims and warn it creates a chilling effect on internal communication. The researchers argue that their dismissal signals a shift away from an open safety culture at the company. (techcrunch.com)
OpenTelemetry introduces native design principles that enable products to export telemetry data to any observability stack seamlessly. The update emphasizes compatibility, scalability, and new features across various SDKs and agents to improve observability infrastructure. (opentelemetry.io)
Users discuss the advantages of subscribing to hosted LLM services versus deploying open-source models locally. They share reasons related to cost, convenience, and customization options. (news.ycombinator.com)
A new self-pruning transformer model achieves 10 times KV-cache compression while maintaining or improving performance in language tasks. It uses a universal attention mechanism that adaptively prunes tokens, enabling better long-context processing with up to 25 times compression at 16k length. (arxiv.org)
OpenAI released a new AI model that has elicited strong reactions from mathematicians. The release has sparked both admiration and concern within the mathematical community. (nytimes.com)
Lithos-ai open-sourced lithos-metal, enabling Qwen3.8-27B to run at over 200 tokens per second on an Apple M5 Max laptop. This allows ultra-fast inference for coding agents with a single command. (twitter.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.