Notes from the battlefield.

A US AI-safety chief resigned three months in. Real reporting: Axios had the scoop, Reuters ran it. But most people who saw it saw a "BREAKING" post from Polymarket, a betting app. Polymarket did not report this; it wrapped someone else's scoop in "BREAKING" and won the reach. That is the attention economy: reporting is expensive and earns nothing, packaging is cheap and earns the distribution. And a feed tuned for reach is not tuned for being right, the same platform was documented pushing false posts and paying creators to fake trades, even as Bloomberg pipes its odds into the Terminal. In 2026 the story is who got to say BREAKING first. Check who is handing you the news, not only what it says.

A food creator with a couple hundred thousand followers filmed his plate at Truffles on St. Mark's Road. A manager told him to stop: not allowed, not a public place, our food. When they kept rolling outside, a staffer followed them onto the pavement and grabbed a phone. Truffles later apologised and fired the employee. But the flaw is the logic. A restaurant that sells to the public on a public street cannot claim privacy the second the public likes the food enough to point a phone at it. That phone was free advertising at a customer-acquisition cost of zero, and they told it to leave. The reel that went out ran the other way: favourite to worst place. The Streisand effect with a receipt.

The WiFi failed. The meeting was on a rooftop, the router could not reach. I fell back to 4G, two bars. The owner was delayed an hour, which became my prep time. When she arrived, we flicked through an elegant deck and put it away. The next two hours were a conversation: her words back to her, the trade-off named before she did, the things I did not know not pretended. Plate opened, Edible Factor logged in, every screen rendered through the choppy signal. The engineering stayed out of the way so the negotiation could be the work.

Two months ago I changed the font on every product I ship. A month later, YouTube Music updated to the same typeface. A month after that, Anthropic's Claude courses on Skilljar rolled out, also in the same typeface. Coincidence? Probably. I notice the rhythm. The decisions that look like trend-chasing six months later sometimes started as a single person looking at the wrong font on a Sunday afternoon.

I record voiceovers and could never tell, mid-read, where I was in the tempo. So I built a small browser tool: paste a script, set the duration, a director sits in your peripheral vision while you read. One evening of code. I have used it for every voiceover since. I showed it to other creators. They asked to use it. One asked if I would build it for their team. The conversation became commercial. The constraint of "what would actually help me record this take" filters out the ninety percent of features a market-shaped product would have added.

Netlify built me a great year. The cleanest CLI in the category, the PR previews that changed how I think about staging. I left when a free-tier deploy quota hit a wall on a Saturday. Vercel taught me what zero-friction deploys feel like; I left because every product feature I wanted required architecting around their proprietary abstraction first. Cloudflare did not feel like a deploy platform. It felt like the internet, exposed to me as primitives: Workers, KV, R2, tunnels, Pages, under one dashboard with one pricing logic. My infra bill across three products is now less than what one of them was costing on Vercel alone.

We launched Plate on Vercel + Next.js + Supabase. It got us to a working v1 in an afternoon. Two months in we moved the database to a more robust non-relational store. The reason was shape, not platform: our schema was a document graph wearing a Postgres mask. Read latencies dropped by an order of magnitude. The customer experience changed visibly within a week. The highest-ROI pivot we have made.