Notes from the production seat.

A film for my senior rescue dog, a two-chat creative workflow, and what the experience taught me about GenAI and making something meaningful.

How I helped turn industrial silos in Shenzhen into an immersive art museum and nightclub, and why creative thinking belongs beyond your own work.

How I directed my website promo with a multi-agent workflow, Claude Opus 5.5, and ElevenLabs while working on other projects.

Model limits and pricing can change overnight. I built a shared project brain with Git and Markdown so the work can move between AI models without losing its memory.

Luma and LTX designed very different events around very different AI products. Their choices show how good go-to-market strategy moves people from interest to adoption.

GPT-6 Astra is the beginning of the AGI era, yet Sam Altman says speedy AI adoption isn't happening. Both are true, and that tension is the strategy problem to think about for the next three months.

It's Fall 2026 and worth stepping back to look at the year of the AI Agent. What it means to run a multi-agent system, what it can do, and what you're actually dealing with.

I moved to the Bay Area without my network, so I built an agentic event and networking engine to find the rooms worth being in. It got me into six of them.

Most advice about selling AI into enterprise comes from the GTM side. I've been the person on the other end of it, running adoption that failed and adoption that worked. Here is what actually moves it.

Earlier this year the fear was that AI would split into haves and have-nots. The best open-weights model now scores three points off the best model in the world, and that changes what it costs to build.

A test creating a commercial spec with Claude Fable 5 connected to OpenArt's MCP. Where AI gets you to V1 unbelievably fast, where it struggles, and why budget-conscious clients should think in possibilities instead of final visions.

A retrospective on Diet Coke's "This Is My Taste". We generated a huge library of art in advance, moderated it hard, and built a flow where the user picked the final image. No live generation, no surprises.

Anthropic is adding machine-readable marks to Claude output. The mark cannot tell generation from proofreading, and most AI clauses in statements of work assume it can.

A 10-second spec spot for a non-alcoholic aperitivo, built from one phone plate shot in my kitchen, an OpenArt MCP running through Claude, and video-to-video restyling. Including the shot that took six tries.

The founding story of Logline AI, winner of The Spark Award. Why I co-founded a company to level the playing field for filmmakers, and what carries the journey next.

I will be serving on the jury of the Tokyo AI Film Festival, alongside creative leaders from around the world. Every image is suddenly possible, so story, taste, and the choices a filmmaker makes are the game.

Netflix told shareholders that AI tools touched around 300 of its shows and movies in 2026 so far. The real story is not cost pressure. It is more stories, more storytellers, and the systems that keep quality high while output scales.

I built this blog in one working session with AI, including the system that designs its own link previews. Here is why owning your archive beats renting your feed.