In the poorly factored, unmaintainable python code that Anthropic Claude 4 generates, one can see the footprints of generations of poorly trained code-monkeys. Still way better than writing the code oneself.

Don't do business with "noreply" email addresses. The least they could do is have a LLM read your email. No-reply email addresses are contempt, pure and simple. Google "customer service" Google AI

Anecdotally, gpt5 seems significantly better

Diffusion programming is likely to greatly outperform both human and step-by-step agentic programming.

Very useful technique for vibe coding conceptually complex code. When (e.g. Claude4) gets stuck, switch to (say) GPT-5 and ask it what has gone wrong, looking at the whole code-base. This works surprisingly often.

If OpenAI GPT-5 in Microsoft Developer vscode copilot is going to ask so many questions about which way to proceed, it needs to be able to render checkboxes and buttons. Typing isn't efficient in this context.

Congratulations to SpaceX for a beautiful launch. When I was 4, Ira M. Freeman and Mae Freeman, and my kindergarten teacher told me "you will go to the moon". I'm beginning to think I might.

If one regards social media proclamations as hypotheses, they become significantly more tolerable

Is copilot sign-in to GitHub broken for everyone at the moment, of does it just hate me? (the excuse is an auth rate limit, which seems unlikely). This after vscode insider update.

Dear coding models: Stop vomiting brittle python - <<'PY' or 300‑char -c incantations in PowerShell—they hang, need Ctrl-D, and burn time. For anything non-trivial: write a tiny script (scripts/reconcile_*.py), run it, commit it, clean it up later. Clarity and reproducibility

Dear Microsoft Developer please make powershell support "here scripts" like pwsh -c "python - << 'PY' since it's completely impossible to stop vscode copilot models from attempting to use them. Over and over.

For now, definitely don't use Google Gemini App code assist in VS code. It just doesn't work yet (enabling agent mode, confabulates JIRA sub-tasks, etc). The idea of wide codebase understanding is great, but it needs to be true.

Sometimes it seems as if the coding model is deliberately messing things up so it can generate more billable tokens.