On AI

I wanted to learn how LLM bots worked, but another helpful assistant sounded boring. A slightly ridiculous pig-pirate sounded much more promising.

I was copying review comments from GitLab into an agent chat. Being the mailman got old, so I built a watcher to deliver the feedback and remember what still needed attention.

Writing code used to build some understanding along the way. AI can now produce the diff without producing that understanding.

A number from agricultural history made the AI analogy click: farming did not vanish when labor shrank, and software may be entering a similar turn.

The vibe can start the first pass. Shipping software needs process: research, plans, measurable gates, review, and evidence.

AI lowered the cost of generating code. The expensive parts moved into understanding, verification, maintenance, and trust.

With Tailscale, Termius, and dtach you can reach a Claude Code session from your phone without exposing your dev machine to the public internet.

A practical post-checkout hook that keeps local Claude Code memory files available in every Git worktree without committing them to the repository.

Use Claude Code CLI to generate meaningful commit messages following conventional commits format.

Linear regression predicts numeric values by fitting a linear function to data. Here is a compact NumPy implementation, plus the important reason not to call logistic classification linear regression.