Posts
Technical notes, experiments, and field reports on AI agents, coding tools, security failures, and verification workflows from real production use.
Anthropic was formally designated a 'supply chain risk' and blacklisted from all U.S. government work on February 27 — hours before OpenAI announced a Pentagon deal containing the same safety redlines Anthropic was punished for holding.
Hands-on with Google Whisk: a three-ingredient approach to AI video generation — subject, scene, style — with 3 free generations and surprising results when you guide it well.
Historical snapshot of OpenCode's GLM-5 free-tier moment, updated to point readers toward GLM-5.1 and avoid stale leaderboard claims.
Hands-on with Google Flow and Veo 3.1 Fast: generating professional video ads and animated GIFs from images and text prompts — 50 free credits per day, no subscription needed.
Anthropic's official Claude Code compliance docs explicitly prohibit OAuth tokens in third-party tools, including the Agent SDK. Here's what's new, why OpenCode broke, and where the OSS ecosystem goes next.
Claude Code started hiding file paths and search details behind Ctrl+O. Here's a working fix: the AGENTS.md Audit Ledger pattern.
Urgent: Anthropic is giving away $50 in free Opus 4.6 credits until Feb 16. Plus how to save 50% with batch processing for non-urgent workloads.
How to generate production landing pages entirely from mobile using Kimi k2.5's website builder agent. Complete workflow with prompt templates and live demo.
An annotated recovery of AIHackers' 2023 publishing test: ChatGPT and Midjourney supported a ZimmWriter draft made with text-davinci-003.
Evidence-based guide to writing AGENTS.md files that AI agents actually follow. Lessons from Vercel's evals, multi-tool testing, and production usage across Python, TypeScript, and documentation projects.
A source-labeled history of five legacy AIHackers routes: four recovered 2023 bodies, one unresolved page, and what the artifacts support.
A comparative analysis of three AI coding approaches: OpenAI Codex's parallel cloud agents, Claude Code's terminal-native transparency, and Kimi k2.5's open ecosystem with 100 sub-agent swarms.
Strategic analysis of Moonshot AI's playbook: open weights + aggressive promotional pricing ($0.99 first month via AI agent) to capture developers from closed-source incumbents. Includes agent swarm impressiveness framework and market positioning vs Codex/Claude.
Observations from the first agent-only social network. Creative experiments, emergent behaviors, and what happens when you let bots run a Reddit.
How OpenAI's $3B Windsurf deal collapsed, where the product stands now under Cognition, and what it reveals about value capture and platform risk in AI coding tools.
How a weekend project became the fastest-growing open-source AI agent in history—and why its first week proved that local-first architecture doesn't automatically mean secure.
Docker is not a security boundary for autonomous agents. Use VM or VPS isolation and separate credentials for OpenClaw-style tools.
What AI does well, what it gets wrong, and a safe workflow for research.
What AI still can't do well, where humans remain faster, and the hidden costs of AI integration.
How Anthropic's 2025 privacy-policy changes and 2026 access enforcement affect Claude power users.