Hacker NewsMonday, February 23

The author describes a structured AI coding workflow used with Claude Code for 9 months. The key principle is never letting Claude write code until reviewing and approving a written plan. The process involves: 1) Research phase where Claude deeply studies the codebase and writes findings to research.md, 2) Planning phase creating detailed plan.md files, 3) Annotation cycles where the author adds inline notes to refine the plan through multiple iterations, 4) Todo list creation for tracking, and 5) Implementation with specific prompts. This approach prevents costly failures like code that works in isolation but breaks surrounding systems, maintains architectural control, and produces better results than direct prompt-to-code workflows.

Despite coding agents' ability to generate cross-platform native code, companies like Anthropic still use Electron for desktop apps because agents struggle with the final 10% of development - edge cases, real-world bugs, and ongoing maintenance. While agents can quickly implement core functionality, the last mile of polishing and the increased support burden across multiple native platforms makes Electron's single codebase approach still more practical for most teams.

HN moderator addressed personal attacks in the thread, reminding users to follow guidelines and avoid destructive behavior. Discussion included controversy around AI figures and community toxicity standards.

Someone involved in retrogaming experimented with running transformer models by connecting GPU directly to NVMe storage, bypassing CPU and RAM. They created a working implementation during weekend coding that functions on consumer GPUs and should perform better on professional hardware.

Andrej Karpathy discussed "Claws," a new AI technology layer that builds on LLM agents to provide enhanced orchestration and persistence. He's exploring various implementations like NanoClaw while noting the term "Claw" is becoming standard for OpenClaw-like agent systems that run on personal hardware.

Taalas created an ASIC chip that runs Llama 3.1 8B at 17,000 tokens per second by physically etching the model's weights into silicon transistors. Unlike GPUs that constantly fetch data from memory, their chip streams data directly through hardwired layers, eliminating memory bottlenecks and achieving 10x better performance, cost, and energy efficiency.

This is a personal uBlock Origin filter list that blocks AI-generated content farms and websites. The creator manually adds sites that publish low-quality AI-written articles filled with ads and referral links, arguing that when searching online, people want human-written content with real experience and opinions rather than potentially unreliable AI-generated text that may contain dangerous misinformation or hallucinations.

Former Dark Sky developers launched Acme Weather, a new iOS app that embraces forecast uncertainty by showing multiple possible outcomes instead of single predictions. Key features include alternative forecast scenarios, community weather reporting, comprehensive maps, customizable notifications, and experimental tools like rainbow alerts. The $25/year subscription app prioritizes privacy and aims to be more reliable than traditional weather apps that hide uncertainty.

A new EU Batteries Regulation enters force today, requiring batteries to have low carbon footprints, minimal harmful substances, and high recycling rates. From 2025, it introduces performance standards and recycling targets. Starting 2027, consumers can remove/replace portable batteries. Companies must ensure responsible sourcing of raw materials like lithium and cobalt.

A CIA analyst recounts their experiences with polygraph tests over their career, from initial screening through periodic reinvestigations. They describe failing their first test despite being truthful, learning about the arbitrary nature of accusations made during testing, and witnessing the psychological toll polygraphs took on colleagues. The analyst observed that tests often involved inexperienced examiners making wild accusations followed by more experienced ones who helped candidates pass, suggesting the process was more about compliance than actual security screening.