Hacker NewsMonday, March 2

This guide introduces microgpt, a 200-line Python implementation that trains and runs a GPT model from scratch with no dependencies. It covers the essential components: dataset processing using a collection of names, character-level tokenization, a custom autograd engine for gradient computation, GPT-2 architecture, Adam optimizer, and training loops. The project represents a decade-long effort to distill large language models to their bare essentials, demonstrating how the model learns to generate new plausible names after training on existing ones.

Discussion centers on how Anthropic demanded specific contractual restrictions against mass surveillance and autonomous weapons, while OpenAI only required "lawful use" compliance - essentially allowing anything the government deems legal through internal memos or policy changes.

Ghostty is a fast, feature-rich terminal emulator that works across different platforms, utilizing native user interfaces and GPU acceleration for enhanced performance.

Obsidian Sync offers a headless client for syncing vaults without the desktop app, useful for CI pipelines and automated workflows. It provides commands to list, create, and sync remote vaults, with options for continuous syncing, encryption, and configuration management through a command-line interface.

Context Mode is an MCP server that compresses Claude Code tool outputs by 98%, preventing context window overflow. It processes data in isolated sandboxes and returns only essential information, extending session time from 30 minutes to 3 hours while maintaining full functionality.

Claude's Memory feature allows users to transfer their AI preferences and context from other providers with simple copy-paste functionality. Users can import their existing AI training data so conversations with Claude feel seamless and continuous, maintaining separate project contexts while letting users view and edit stored memories across all paid plans.

Croatia has been officially declared landmine-free after 31 years of demining efforts following the Homeland War. The dangerous work cost 1.2 billion euros and 208 lives, including 41 deminers. Nearly 107,000 mines and 407,000 pieces of unexploded ordnance were removed.

AI-assisted development creates "cognitive debt" where engineers can rapidly generate code but lack deep understanding of what they built. While velocity metrics improve, engineers struggle to explain or modify their own code months later. This decoupling of output speed from comprehension leads to burnout, failed code reviews, and organizational knowledge loss that only becomes apparent during incidents or architectural changes.

This is an educational demo showing how AI chat assistants could look if monetized through advertising. It demonstrates various ad formats including pre-chat interstitials, banner ads, sponsored responses, contextual text ads, product cards, freemium paywalls, and retargeting. The tool helps marketers, developers, and product managers understand AI monetization landscapes by showing concrete examples of ad-supported versus subscription models. All brands and ads are fictional for demonstration purposes.

Alibaba released Qwen3.5 Medium Model series featuring four new AI language models with three available under open source Apache 2.0 license. These models outperform OpenAI's GPT-5-mini and Anthropic's Claude Sonnet 4.5 on benchmarks while offering 1 million token context windows on consumer GPUs. The models feature efficient architecture with built-in reasoning capabilities and competitive API pricing, making advanced AI more accessible for enterprises without requiring massive infrastructure investments.