Three stories landed this week from three fields that have nothing to do with each other. Same shape underneath all of them.
In each one, the expert stopped making the thing and started building or steering the machine that makes it. That is a promotion in some tellings and a demotion in others, and which one it turns out to be depends almost entirely on who owns the machine.
Nine mathematicians, no decision-making power
OpenAI has formed a mathematics advisory group after an internal model resolved more than 100 open problems, including a solution to the Navier-Stokes Millennium Prize problem, as reported by TechCrunch.
Nine prominent mathematicians, hosted at Princeton's Institute for Advanced Study. The group serves in an advisory capacity with, in the company's own words, no decision making power.
Context matters here. A group of 25 Fields Medal winners had already signed a letter arguing that AI labs are threatening their intellectual work as they try to one-up each other. Only one signatory, Camillo De Lellis, joined the advisory group.
So the world's most credentialed practitioners of a discipline are being offered a seat at a table where they cannot vote. The polite word for that is advisory. The accurate word is consultation.
I do not think this is cynical on OpenAI's part, and it is genuinely hard to see what a decision-making board of external mathematicians would even decide. But the structure tells you the direction of travel, and it is the same structure arriving in every profession with a body of hard-won knowledge.
Your expertise is wanted. Your authority is not on offer. For anyone selling specialist knowledge for a living, which includes most agencies and most consultants, that is the commercial question of the decade.
Two fashion designers who built software instead
Google's Envisioning Studio worked with designers Jane Wade and Sergio Hudson ahead of New York Fashion Week, and the part worth noticing is what got built rather than what got generated.
Wade used a custom tool that let her curate hair, makeup, accessories, shoes and garments on digital models before producing physical samples, cutting what had been a three-day in-person fitting process. Hudson used one that simulated his runway, adjusting lighting, props and the paths models would walk, without paying for a 3D render on every revision.
Neither of them asked a model to design clothes. They built the instrument they wanted, in plain language, with no code, and kept the judgement for themselves, as Google described it.
That distinction is the whole thing. The generic use of generative tools is to ask for output and accept a version of the average. The expert use is to compress the loop between an idea and seeing it, then apply taste to a hundred iterations instead of four.
Cutting a three-day fitting is not a story about creativity being automated. It is a story about a designer getting three days back and spending them on decisions rather than logistics.
Every marketing team I know has an equivalent: the repetitive setup step that eats a specialist's week and produces nothing anyone would call work. That is the thing to build a tool for, and the person who should build it is the specialist, not the technical team two floors away.
Two thousand pull requests, and one real bottleneck
Warp ships around 2,000 pull requests a month through an internal system it calls Wilson, which can go from a Slack request to a merged pull request in as little as 35 minutes. The team described it to Lenny's Newsletter as an AI factory: the whole development lifecycle running in the cloud, from triage through implementation and testing.
Then the honest part. Humans in code review are now the biggest bottleneck, with reviewers arriving roughly three and a half hours after kickoff.
The metric they chose is the one I would steal. They track human interactions per pull request, on the logic that agent speed is meaningless if a person is the rate limiter.
That is a much better efficiency measure than the ones most companies are reporting, because it counts the thing that actually constrains throughput rather than the thing that is easy to celebrate.
Translate it out of engineering and it works everywhere: human touches per campaign launch, per client report, per proposal.
Count those and you find out quickly whether your AI adoption removed work or merely relocated it.
I wrote about the org chart becoming a loop earlier this year. Warp is what that looks like once somebody instruments it.
The pattern, stated plainly
A mathematician advising a system that outpaces them. A designer building the instrument rather than the garment. An engineering team measuring how often a human has to touch anything.
In all three, the specialist's hands came off the output. In two of the three, the specialist got to define the system instead, and that is the difference between gaining ground and losing position.
The fashion designers built their own tools, so the taste stayed with them. The mathematicians were invited to comment on someone else's, so it did not.
This connects to something I keep returning to, that the middle of the stack is vanishing. The people who survive that are the ones who moved up into defining the system, not the ones who moved down into approving its output.
Decide this year which of those two roles you are building toward. The window where it is still your choice is not going to stay open.