Would love to know the rationale behind @ChatGPTapp feature deployment across different platforms: Record mode is in the desktop Mac app, but not iOS; Codex appears as a link in ChatGPT web but not Mac app; web app supports model selection in projects but Mac app doesn't.

Built a quick MCP server for my @fastmail account and added it to Claude Code. It works really well. Maybe Claude Code is a great coding agent because it's actually just a great agent.

Switched claude code from pay-as-you-go API pricing to a Claude Pro plan. Surprised how much more I use it now that the total monthly cost is fixed (even though it's not necessarily cheaper).

@AmpCode were right:

OpenAI's 4.1-* models are all priced cheaper per token for the basic text API, but fine tuning (training and inference) is more expensive than 4o.

Tried out Gemini 2.5 Pro in Aider last week and the results were impressive. Unfortunately I hit the rate limit very quickly. Looks like this has now been increased: Tier 1 is now 20 RPM, although daily limit it only 100 requests. ai.google.dev/gemini-api/d...

Generally not a fan of websites with scroll driven animations but this is just delightful: cora.computer

Trying out Gemini 2.0 Flash with one of Google’s example prompts. Works for Flash 1.5 but the diarisation is missing with gemini-2.0-flash-exp.

I’ve had similar conversations and was mystified that people quickly dismiss these tools.

My take: It’s unprecedented that such a powerful technology is made available so quickly and widely at virtually zero cost.

When prompting an LLM whitespace counts as a token (makes sense: whitespace is pretty significant in language). Running a formatted JSON doc through `tr -d '[:space:]` is a quick saving. But maybe there’s a more efficient format than JSON to use? 💭