Writing

The last note in the arc — from a browser tab in 2022 to building the layer agents stand on, and the one question that outlasts every tool

What Los Angeles was to NFTs, San Francisco is to AI — except this time the thing at the center is real, and the center has never been so small

Everyone's watching the ceiling — hour-long agents, slick demos — but the dull layer underneath is where the keys stay, and that's the part my father taught me to build

People ask which of my two products is the real one — but trust for agents and ownership for people turned out to be the same problem wearing two faces

An assistant overshared a dossier of me I never wrote — so I'm building memory I own, where the AI proposes and I ratify

A part is verified when the agent stops arguing with it — on conformance, attestation, and a trust badge that shows verified versus generated

The first parts have shipped — now the real work is the contract a piece of software must satisfy before an agent will trust it instead of rewriting it

Three years from a one-question text box to handing agents whole features overnight — and why the year ahead is for building the quiet layer in the open

Four models this week, the same blank slate each time — the wall in front of me isn't intelligence, it's memory I own, and I've started building it

The product stance behind the catalog: a part is verified, attested, versioned — and untouchable. The agent composes it, never rewrites it, and trust becomes the unit of reuse

I closed the client work and pointed the whole studio at one bet: as agents get powerful, the scarce thing isn't intelligence, it's parts they can trust

The studio runs on a few senior humans and a fleet of agents now — same small-team instinct, new hands, and a gap I can't stop noticing

Agents reinvent the same bolt every time because they have no shelf to reach for — so I'm building one, and it's called Ctrl AI

Two sign-in flows, two days apart, nothing alike — and the drift underneath finally showed me the bug living beneath every other bug

Claude 4 holds a multi-hour task and the unit of work jumps from a function to a feature — but lean on it and the real gap shows: it has no memory of what we already trust

Three strong models shipped this month and I forgot which lab made which — when capability turns into a utility, the value moves up the stack to trust, reuse, and memory

I shipped a feature I never typed — what changes when the thinking moves upstream from syntax to intent, and the judgment that won't hand off

Claude Code arrives, the editor starts to dissolve, and I stop typing code — I describe outcomes and a terminal that thinks first writes back

For two years I carried the model's answers to the world by hand; this week the agent picked up the task and walked off with it itself

Two years in, the models are brilliant and the agents are real — what's still missing isn't intelligence, it's trust: grounding, provenance, memory

Anthropic shipped MCP — a boring standard socket for feeding tools and data to a model, and the boring ones are always the ones that matter

Anthropic gave a model hands this week — it took over my screen and clicked. The thrill wears off fast; what it's allowed to touch is the real question

OpenAI's o1 stops answering on reflex and starts to plan — the shift from autocomplete to deliberation, and why the pause is the point

The AI editor I quietly adopted a year ago is suddenly the whole industry's — and the road I drove alone now has traffic on it

The model became my interface to every framework I touch — enormous leverage, and a creeping illiteracy I have to watch in myself

A new Sonnet builds a working interface beside the chat — and leaves me circling the real problem: we keep regenerating software that already exists

GPT-4o's voice arrives on an ordinary Tuesday — the interface dissolves into plain talk, but the machine still meets me as a fluent stranger

Llama 3 ships open weights, and for a builder twice burned by platforms he didn't control, a model you can hold on your own disk changes the calculus

In 2019 we hand-wired a mortgage robot from scans to approval; this week one prompt did the whole thing — vindication, and the quiet cost of being early

A new model arrived and I skipped the leaderboards — Claude 3 Opus earned my daily seat on judgment and prose, not benchmark bars

Gemini's million-token window and Sora land the same week, and another intuition about limits quietly breaks

A year after shuttling code between two windows, the model moved into my editor — and the gap I can't shake isn't intelligence, it's what it knows about us

A board fired and rehired the man half the industry runs on over one weekend, and it drew a hard line under a question I can't unsee: what in this stack do I actually own

OpenAI's DevDay turned the model into a platform, and every builder now faces the oldest question I know: build on top of it, or beside it

A model brilliant about the world and blank about my company — and why bolting on retrieval is plumbing, not the real problem: trust in the answer

Claude 2's 100k context lets me feed whole documents at once — but it's the calmer temperament, not the benchmark, that keeps me coming back

Japan was the future once. Walking Tokyo now, you feel a whole country that reached a peak and stayed there while the center of gravity moved on

The model moves into my editor and the copy-paste tax finally falls — but a quieter problem shows up in its place

Everyone's passing AutoGPT around like a magic trick — I gave one a real job and watched it unravel, and the failure points straight at what's missing

I'm building software again — by talking to GPT-4 and shuttling its code into VS Code by hand.

I ran GPT-4 through the private tests every model had failed, and it cleared the ones I thought were five years off — so I'm raising what I dare to build

Microsoft's Bing chatbot melts down as "Sydney" — and the real unease isn't the machine, it's how fast we project a self into a model with nothing to anchor it

A hundred million users in two months, ten billion from Microsoft, and the table reaches for the bread — the speed of the shrug unsettles me more than the technology

JudRobot and Hyponamíru weren't failures — they were scouts I sent out before the machine could carry the idea, and now it can

A week in, a founder who waited two decades for this can't stop opening the tab — every night he brings it something harder, and every night it goes further than he expects

OpenAI quietly put a chatbot on the internet today. I think the ground just moved.