Two worlds,
one curiosity.
By day I build better public services; by night I lose myself in a glass of wine. In between, I put AI to work hands-on, so I stay ahead and can bring others along. What I find beneath the surface, I share here.
Latest stories
Innovation, AI & wine
How I use AI: building a form that sends nothing anywhere
Filling in ten questions by hand is tedious. I had a card built that does it for you, and the design check caught three AI tics in it.
One memory for chat and Cowork, and you decide what goes in it
Anthropic merged the memory behind Claude chat and Cowork. It saves you from repeating yourself, and it quietly moves something: what you say once now travels into work you never read.
Your AI level and your wine level are two separate things
One post got me two opposite replies: too basic, and far too technical. Both were right, because they were sitting in different quadrants.
Claude Academy is live, and the research underneath it is the better story
Anthropic put its entire learning catalogue online for free. The research underneath shows that people check less as the output starts to look finished.
The thread said last week. The lecture is eighteen months old.
A viral X thread sold an Andrej Karpathy lecture as fresh news. The video has been online since February 27, 2025, and the table of contents belongs to a different talk entirely.
Local model or cloud: when is it worth running it yourself?
I run image generation and dictation locally, and my scheduled tasks in the cloud. That is not a principle but a trade-off. Here is how it lands.
Claude Code ships a concise output style, and I already had one
Version 2.1.237 added a built-in style that leads with the result and drops the narration. Where that instruction lands decides what it costs and who actually hears it.
Loops: the work that comes back every day should stop being yours
Ramp runs agents at every step of its engineering process. The lesson I take from it is not about scale, but about which work literally repeats itself.
Claude designed proteins that worked in a real lab, against 14 of 15 targets
Anthropic had Claude run a protein design campaign on its own. Two outside labs built the designs and measured them: 354 of 1,320 bound. The percentage is not the interesting part.
Sparks
Short notes, tips & loose thoughtsTested a new local image generator. Runs fully offline. The hero above came out of it.
Tasted blind: a grower champagne beat two big houses. Notes to follow.
EUDI Wallet pilot live in NL. First take: UX decides whether the trust lands.
Tip: keep your best prompts as snippets, not in your head. Saves time daily.
Golden hour lasted eleven minutes yesterday. Exactly enough for three bottles.
Note to self: a method is a tool, not an excuse to stop thinking.