Top Stories

Caveman: Why Use Many Token When Few Token Do Trick

716 points · github.com

The runaway hit of the day is Caveman, a Claude Code skill that compresses AI responses into dramatically condensed language — slashing token usage by an average of 65%. Instead of Claude saying “I’d be happy to help you with that, let me take a look,” Caveman makes it say “Bug in auth middleware. Fixed.” Code blocks stay properly written; only the explanatory fluff gets axed. Users can choose intensity levels from Lite (removes filler while keeping grammar) to Ultra (maximum compression with abbreviations). What’s fueling the 313-comment discussion isn’t just the humor — a 2026 research paper shows that brief responses can actually improve accuracy by 26 percentage points, challenging the assumption that verbose explanations are inherently better.

Eight Years of Wanting, Three Months of Building with AI

658 points · lalitm.com

Lalit Maganti spent eight years wanting to build developer tools for SQLite, then finally shipped them in three months using AI coding agents. This is one of the most nuanced, evidence-backed accounts of building with AI yet. The wins are real — AI broke through procrastination, compressed learning curves for unfamiliar domains like Rust tooling, and made the final polish affordable. But the losses are equally instructive: the author repeatedly lost touch with his own codebase, 500+ tests provided false confidence over spaghetti architecture, and the addictive loop of “just one more prompt” actually slowed progress. The central takeaway resonating across 208 comments: AI is an incredible force multiplier for implementation, but a dangerous substitute for design.

Gemma 4 on iPhone

448 points · apps.apple.com

Google’s Gemma 4 model is now running directly on iPhones through the new AI Edge Gallery app, and the HN community is buzzing. This represents a significant step in on-device AI — a capable language model running locally with no cloud dependency, no API keys, no data leaving your phone. The 124-comment thread covers everything from inference speed benchmarks to privacy implications. Combined with a related front-page story about running Gemma 4 locally via LM Studio’s headless CLI (206 points), it’s clear the local-first AI movement is accelerating rapidly.

Artemis II Crew See First Glimpse of Far Side of Moon

431 points · bbc.com

The Artemis II crew has become the first humans to see the far side of the Moon with their own eyes since Apollo 17 in 1972. The BBC video capturing this moment has the HN community in a rare state of collective awe, with 326 comments that range from technical orbital mechanics discussions to reflections on what it means for humanity’s return to deep space. After years of delays and skepticism about NASA’s Artemis program, this tangible milestone is renewing optimism about lunar exploration and the path toward a sustained human presence beyond Earth.

Why Switzerland Has 25 Gbit Internet and America Doesn’t

274 points · sschueller.github.io

A deep comparison of internet infrastructure approaches across the US, Germany, and Switzerland that’s sparked a 213-comment debate. The core argument: fiber optic infrastructure is a natural monopoly, and how governments handle that reality determines outcomes. The US deregulated and got territorial monopolies with shared bandwidth. Germany regulated without coordination and got wasteful parallel trenches. Switzerland mandated four dedicated fiber strands to every home with open access for any provider — creating genuine competition on service rather than on who owns the cables. The result is 25 Gbit connections while Americans fight over whether their “gigabit” plan actually delivers 200 Mbps.

Microsoft Hasn’t Had a Coherent GUI Strategy Since Petzold

245 points · jsnover.com

Jeffrey Snover (creator of PowerShell) argues that Microsoft’s last clear answer to “how should I build a Windows UI?” was Charles Petzold’s Win32 API books in the late 1980s. Today, developers face seventeen different frameworks — Win32, WinForms, WPF, WinUI 3, MAUI, Electron, Flutter, Tauri, and more. Snover traces the fragmentation to internal politics (a bitter Windows vs .NET civil war after the failed Longhorn project), conference-driven announcements that blindsided developers, and business pivots made without warning. The irony that Electron — built by GitHub, not Microsoft — has become Windows’ most widely deployed desktop GUI technology lands especially hard in the 137-comment discussion.

Nanocode: The Best Claude Code That $200 Can Buy, in Pure JAX on TPUs

167 points · github.com

An open-source project showing how to train your own agentic coding model from scratch for roughly $200. Built in pure JAX and optimized for TPU training, Nanocode follows the full pipeline: pre-training on code and web data, supervised fine-tuning with synthetic tool-use data, and alignment via Direct Preference Optimization. The resulting 1.3B parameter model learns to use file reading, editing, grep, and bash tools — a miniature Claude Code in about 5,500 lines of hackable JAX. Smaller variants train for as little as $34 in 90 minutes. It’s a remarkable demonstration of how accessible agentic AI training has become.

A Tail-Call Interpreter in (Nightly) Rust

133 points · mattkeeter.com

Matt Keeter demonstrates how Rust’s experimental become keyword enables high-performance VM interpreters without unsafe code. By storing VM state in function arguments (mapped to CPU registers) and using tail calls to eliminate stack growth, his Uxn CPU interpreter actually outperforms hand-coded assembly on ARM64 — 76ms vs 87ms on the Mandelbrot benchmark. The x86-64 story is less rosy due to suboptimal codegen, and WebAssembly performs 4.6x slower than native. The 22-comment discussion dives into the implications for language VM design and whether this pattern could become a standard approach once become stabilizes.

In Japan, the Robot Isn’t Coming for Your Job — It’s Filling the One Nobody Wants

131 points · techcrunch.com

Japan is moving physical AI from experimental pilots into real-world deployment, driven not by cost-cutting but by demographic necessity. With an aging population and shrinking workforce, robots are filling roles that simply can’t attract human workers. Backed by venture capital from Salesforce Ventures, Woven Capital, and Global Brain, this represents a different narrative from the Western “robots taking jobs” anxiety — here, robots are the solution to a labor shortage that threatens the economy. The 140-comment discussion explores whether this model could work elsewhere as populations age globally.