Published on February 22, 2026 at 06:00 CET (UTC+1)
How I use Claude Code: Separation of planning and execution (239 points by vinhnx)
The article describes a developer's optimized workflow for using Claude Code, an AI coding assistant. The core principle is a strict separation of planning and execution, where Claude must first research the codebase and produce a written, approved plan before any code is written. This methodology prevents architectural drift, reduces token waste, and keeps the developer in control, culminating in a more efficient and effective development process than immediate code generation.
Japanese Woodblock Print Search (25 points by curmudgeon22)
This piece presents Ukiyo-e Search, a specialized database and search engine for Japanese woodblock prints. Its primary feature is the ability to search for prints using image uploads, leveraging visual similarity AI to find matches across a vast collection of over 200,000 images. The site also allows browsing by historical period and artist, serving as a digital archive and research tool for this art form.
Show HN: Llama 3.1 70B on a single RTX 3090 via NVMe-to-GPU bypassing the CPU (152 points by xaskasdf)
This technical Show HN post introduces "ntransformer," a high-efficiency C++/CUDA LLM inference engine. Its key achievement is running large models like Llama 3.1 70B on a single RTX 3090 GPU with only 24GB of VRAM. It accomplishes this by streaming model layers from NVMe storage directly to the GPU via PCIe, bypassing the CPU to dramatically reduce memory constraints and enable local execution of massive models.
A Botnet Accidentally Destroyed I2P (53 points by Cider9986)
The article details how the I2P anonymity network was accidentally crippled by the Kimwolf IoT botnet in 2026. The botnet, attempting to use I2P as backup command-and-control infrastructure, flooded the network with hundreds of thousands of malicious nodes in a devastating Sybil attack. In response, the I2P team rapidly deployed an update featuring default post-quantum encryption (ML-KEM) and new Sybil attack mitigations, highlighting an intersection of cybercrime, network resilience, and cryptographic advancement.
Evidence of the bouba-kiki effect in naïve baby chicks (98 points by suddenlybananas)
Based on the title and context, this scientific paper presents evidence for the "bouba-kiki" effect in baby chicks. This effect describes a cross-modal association where rounded shapes are reliably matched with nonsense words like "bouba" and spiky shapes with words like "kiki." The discovery that this effect exists in naïve chicks suggests this sound-shape correspondence may be an innate cognitive bias not unique to humans, with implications for understanding the evolution of language and perception.
How far back in time can you understand English? (429 points by spzb)
This essay presents a linguistic experiment where the author writes a simulated travel blog post that gradually shifts the English language backward in time over 1,000 years. The piece demonstrates how spelling, grammar, vocabulary, and authorial voice change dramatically across centuries, moving from modern English to what becomes nearly incomprehensible to a contemporary reader. It visually illustrates the concept of language drift and the relative recency of mutually intelligible Modern English.
Two Bits Are Better Than One: making bloom filters 2x more accurate (34 points by matheusalmeida)
This technical blog post explains how the Floe database engineering team improved the accuracy of Bloom filters, a probabilistic data structure used to speed up SQL queries. The authors detail their innovation of using two bits per bucket instead of one, which allows the structure to store more information about hash collisions. This modification reduces the false positive rate by approximately half, making query execution more efficient without significantly sacrificing speed or memory.
Scientists discover recent tectonic activity on the moon (27 points by bookmtn)
Based on the title and source, this science news article reports on the discovery of evidence for recent tectonic activity on the Moon. Scientists likely used data from orbital missions like NASA's Lunar Reconnaissance Orbiter to identify and analyze small, fresh-looking scarps or faults. This finding challenges the long-held view of the Moon as a geologically dead world and suggests it has experienced seismic shaking (moonquakes) within the last few hundred million years, potentially even more recently.
“Playmakers,” reviewed: The race to give every child a toy (4 points by fortran77)
This New Yorker article reviews the book "Playmakers," which chronicles the history of the American toy industry and the pivotal role of Jewish entrepreneurs in its creation. It focuses on figures like Morris Michtom, who invented the teddy bear, highlighting how immigrant experiences, cultural influences, and savvy marketing transformed toys from luxury items into staples of childhood, fundamentally shaping American consumer culture and childhood itself.
zclaw: personal AI assistant in under 888 KB, running on an ESP32 (133 points by tosh)
This GitHub project showcases "zclaw," an extremely compact personal AI assistant designed to run on the low-power, low-cost ESP32 microcontroller. The entire system fits in under 888 KB, with the core app code around 25 KB. It demonstrates the frontier of edge AI, packing features like GPIO control, cron jobs, and memory into a minimalist, efficient package that can operate independently on highly constrained hardware.
Trend: Specialized AI Workflow Engineering Over Raw Prompting. The Claude Code article highlights a shift from simple prompt-and-check loops to meticulously designed, human-in-the-loop workflows. This matters because it marks the maturation of AI tool use from a novelty to a professional discipline focused on reliability, oversight, and efficiency. The takeaway is that future AI productivity gains will come less from model improvements alone and more from systematic workflow design that embeds planning, verification, and iteration.
Trend: Proliferation of Accessible, Specialized Multimodal AI. The Japanese woodblock print search exemplifies the deployment of visual similarity AI as a public-facing, user-friendly tool for a niche domain. This trend shows AI moving beyond general-purpose chatbots into embedded applications that solve specific problems (like art identification) with intuitive interfaces (image upload). It implies vast opportunities for AI to digitize, organize, and provide access to specialized corpuses in culture, science, and industry.
Trend: Extreme Hardware Optimization for Local LLM Inference. The project to run a 70B parameter model on a single consumer GPU demonstrates intense focus on overcoming hardware bottlenecks through software ingenuity (e.g., NVMe direct-to-GPU streaming). This matters as it directly enables greater accessibility, privacy, and cost-effectiveness for powerful models. The implication is a continued arms race in inference optimization, making high-capability AI more decentralized and less dependent on cloud APIs.
Trend: AI/ML as Both a Threat Vector and a Defense Tool in Cybersecurity. The I2P botnet story illustrates the dual-use nature of advancing technology. While not directly about AI, the response (implementing post-quantum crypto) touches on ML-KEM, an ML-assisted algorithm. The broader trend is the escalating complexity of attacks and defenses, where AI can power offensive botnets and, conversely, be used for threat detection and cryptographic evolution. Developers must consider adversarial AI and build resilient, future-proof (e.g., post-quantum) systems.
Trend: Algorithmic Efficiency Gains in Foundational Data Structures. The Bloom filter optimization article is part of a larger trend where core data structures and algorithms are being re-examined and enhanced in the context of modern data-intensive and AI-driven workloads (like large-scale SQL processing). This matters because as datasets grow, even constant-factor improvements in basic operations yield massive resource savings. The takeaway is that there is significant low-hanging fruit in optimizing the "plumbing" of data systems that support AI, which can be as valuable as advancing the models themselves.
Trend: The Push for MicroAI on Ultra-Constrained Edge Devices. The zclaw project on an ESP32 represents the cutting edge of miniaturizing AI capabilities for the extreme edge. This trend matters as it expands the potential universe of smart, autonomous devices into areas where power, cost, and size are paramount constraints (e.g., sensors, wearables, simple actuators). The implication is a move towards a fabric of ambient, intelligent micro-devices that perform specific, localized reasoning without connectivity, challenging developers to achieve maximal functionality with minimal resources.
Analysis generated by deepseek-reasoner