Writing / public archive, not a blog
Things built, things learned, things I think.
The smart model is the bottleneck. Treat it that way.
For weeks I've been running a Frankenstein of Claude Code, codex CLI and grok-cli. The smartest model specs and places the seal; everything else is done by something dumber and faster. It's ~30 lines of markdown and it fits in your head.
deploy to prod. make no mistakes.
We launched the first company brain live from the Italian Founders House, surrounded by founders doing their own things. These are the release notes: more sushi than marketing.
The model is a commodity. The method isn't.
95% of GenAI projects produce no ROI. The bottleneck has moved: from the model to the people, the workflows, the infrastructure you own.
What off-piste skiing teaches about shipping AI to production
The mountain and production punish the same mistake: trusting yesterday's conditions. Bulletin, checks, margin. Then you drop in.
CONTEXT IS GOLD: the company brain your team owns
Useful AI work should leave behind an inspectable, correctable, governable artifact. Markdown files your team owns. Not throwaway chats, not an opaque memory engine.
Own the harness, swap the model
Public leaderboards compress real differences. The verifier is the hidden variable. The operating conclusion: evaluate on your own work, own the harness and the verification, treat the model as a commodity.
ChatGPT Ads isn't a new channel. It's the end of the keyword.
Targeting moves from the compressed intent of the keyword to the expanded intent of the conversation. Whoever understands the layer before the media spend wins the next cycle.