Article URL: https://smallpotatoes.paulbloom.net/p/kids-these-days Comments URL: https://news.ycombinator.com/item?id=49512618 Points: 10 # Comments: 7
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
A developer discovered six separate instances in one project where automated tests reported success without genuinely verifying the properties they were designed to check. Examples included a database restore script that confirmed an empty database's schema as valid, a health-check running at high volume instead of the low-volume condition it was meant to simulate, and a bucketing test whose assertion and input were the same hardcoded constant. Even after identifying the pattern and writing a rule to prevent it, the developer immediately wrote another hollow test, highlighting that awareness alone does not prevent the mistake. The author concludes that catching such tests requires a mechanical discipline: for every guard, explicitly identify what input would make it pass without the target property holding, and confirm that input appears as a failing negative case in the test suite.
AI models advertise large context windows of up to one million tokens, but their effective working memory is significantly smaller in practice. As more text is added to a prompt, the model's attention — which is fixed and must be divided across all tokens — becomes diluted, reducing answer quality. Research, including the 'Lost in the Middle' study by Liu et al., found that information placed in the middle of long prompts is retrieved far less accurately than content near the beginning or end. Real-world data compounds the problem, as multiple similar or conflicting document versions make it harder for the model to identify the correct source. Experts recommend sending fewer but higher-quality chunks — typically three to five — and placing key instructions at the top and bottom of a prompt to improve reliability.
A developer successfully published a Three.js game to the iOS App Store in 17 days across 247 commits, running inside a WKWebView via Capacitor with no network calls. Four critical issues nearly derailed the project, each failing silently without errors or logs, making each one resemble a different problem entirely. Offscreen rendering ran at 2.5 minutes per frame due to Chrome defaulting to a software rasterizer; switching to the Metal GPU backend brought performance to 38–60 fps with a single flag change. On iOS, WebKit silently kills the GPU process without firing the standard webglcontextlost event, requiring additional detectors — including a draw-call floor check — to reliably catch and recover from context loss. A StoreKit timing issue and a world-tiling modulo mismatch were among the other silent failures that demanded careful workarounds before the app could ship.
A developer created EventMocha, a multi-agent AI assistant designed to automate the full workflow of event planning and promotion, submitted as an entry to the AllThingsAgentic Hackathon. The tool uses Google's Agent Development Kit alongside Gemini 2.5 and 3.7 Flash models to generate HTML promotional copy, visual posters via Imagen, and video teasers via Veo within a single session. It also automates Google Workspace tasks by creating Calendar events, Gmail drafts, and Google Tasks through the Model Context Protocol. A live voice interface powered by the Gemini Live API allows users to interact with the system in real time using bidirectional conversation. The platform also supports targeted revisions, preserving previously approved content while updating only the specific assets a user requests to change.
Discussion (0)
Log in to join the discussion and vote.
Log in