People. Process. Product.
Design for the agentic era.
The First 80% Is the Easy Part
In art school, we had a saying: the last 20% is the last 80%. That final stretch is where good design becomes exceptional.
In art school, we had a saying: the last 20% is the last 80%. That final stretch: details, revisions, the faffing about, is where good design becomes exceptional.
AI has flipped the beginning of that equation. It can give you 80% of an output with 20% of the effort. And because so much appears so quickly, it creates the illusion that the work is done.
It isn’t. You may have 80% of the output, but you still have 80% of the work ahead of you.
The remaining work is slower and less gratifying. It’s editing, questioning, refining, recomposing, removing, and staying with the piece long enough to notice what everyone else will miss. It can take four, eight, even ten times longer than generating the first pass…exactly!
That’s the part people skip! The speed of the first draft makes diligence feel boring & inefficient. How wrong they are. The boring work nobody cares to examine IS THE work.
Details haven’t become less important because the draft arrived faster — they matter more. Everyone can produce something reasonably competent in seconds. Taste lives in taking the time and care to making careful boring decisions for hours.
Addiction Economics
Remember when Uber rides were cheap? VC subsidized your rides for a decade. Taxis died! The plan was never to stay cheap.
Remember when Uber rides were cheap?
VC subsidized your rides for a decade. Taxis died! The plan was never to stay cheap. The plan was to make you dependent.
Once you've offloaded your writing, code, strategy, memory, thinking, workflows…They can raise the price. What are you going to do, learn to think again?
The one thing keeping the market honest is, of all things, China. Alibaba ships Qwen for free. It runs on my Mac Studio. Every time Anthropic raises a price, an open model lands. The communists (China’s communist, right? I have no idea anymore) are the antitrust now.
So here's the move: assume the cheap part ends. Keep one foot on something you can run yourself.
Artisanal, Free-Range, Organic Data
AI provenance is here. We’ve never had easier access to something that convincingly looks like information.
AI provenance is here. We’ve never had easier access to something that convincingly looks like information.
Ask the robot a question and it: searches, summarizes, synthesizes, and hands back an answer. The whole thing feels magical. So naturally, everyone is ingesting the obvious, grabbing the first plausible result, jamming it into the product. Problem solved!
But magic commoditizes quickly…
Soon, everyone will be able to produce a plausible answer. The important question moves upstream: Where did it come from?
How was the data collected? What kind of data is it? What was omitted? Can the source be checked? Can it be defended? Is the source trustworthy?
AI products don’t have satisfying answers yet. For the moment, they don’t need them. There is so much obvious information waiting to be ingested that access itself is valuable.
Eventually the obvious gets exhausted. The models converge. Everyone has the same magic trick. That’s when differentiation moves to the ingredients: proprietary research, trusted relationships, unusual archives, carefully maintained datasets, real expertise, and information gathered with permission and context.
Welcome to artisanal, free-range, organic data.
MAD 2.0: Mutually Assured Disruption
When anything can be faked, trust becomes the most valuable infrastructure we have.
There’s an essay called “The Bitter Lesson.” The gist is that human cleverness loses to raw computation. Every time. So, let’s take compute seriously.
The person with the most compute wins. The company with the most compute wins. The government with the most compute wins…You’ve seen this movie…
We learned to turn rocks into gas, spin them through tens of thousands of centrifuges, refine them into spicy metal, bolt it to a missile, and bury it underground on 24/7 standby. Then we put the same bombs on planes that never land and submarines in every ocean.
If compute equals victory, everyone is already building an AI Manhattan Project. Look around — it’s already here. So everyone races to hoard compute, and everything orbiting that economy becomes irrationally(?) valuable.
The moment anyone can fake anything, nobody can believe anything. Do you believe what you read these days? The next five years are going to be turbulent, unfair, and loud. Oh wait... That’s now.
…Here is where the story turns…
When everything can be faked, you do what humans did before institutions: you ask someone you trust. Ryan Findley calls it “SSL for people”: chains of human beings who can vouch for one another. Not because a blue check or a giant platform says so, but because they share history. Repeated interaction. Reputation. Mutual risk. A relationship can be impersonated. But the history behind that relationship is much harder to fake (for now).