front row at the changing economy

The shape of the economy is the shape of society, with about a five-year lag.

This is the thesis. Everything else in this essay is downstream of it. If you accept it, the conclusion about what to spend your career on between now and 2035 is almost forced. If you don't accept it, the rest of this essay is going to sound a little crazy, and that's fine.

the five-year lag

Most of what people think of as "society" is actually downstream of how the economy is structured. The kinds of jobs that exist determine where people move, what cities grow, what schools teach. The way work is organized determines what childhood looks like, what marriage looks like, what retirement looks like. The way money flows through businesses determines what gets built, what gets healed, what gets ignored.

You can see this most clearly in retrospect. The American suburb is downstream of the postwar manufacturing boom and the car industry's role in it. The two-income household is downstream of services replacing manufacturing as the dominant employment category. The gig economy reshaped how a generation thinks about work, identity, retirement. None of those changes started as cultural shifts. They all started as economic ones, and the cultural shifts caught up a few years later.

The lag is the thing. The economy moves first. Society reorganizes around it. By the time the change is visible in the culture, the underlying economic decisions that drove it are five or ten years old. The people who shaped what happened were the ones in the rooms where those decisions were being made.

This is where AI is right now.

what's actually happening, today

The AI discourse spends almost all of its bandwidth on the model layer. Which lab released what. Which benchmark moved. Which capability emerged. This is interesting and it is also where almost none of the consequential decisions are being made.

The consequential decisions are being made in conference rooms in Chicago and Frankfurt and Singapore, where a CIO is deciding whether to put an AI system in the workflow that approves claims, hires people, prices loans, escalates support tickets, reviews legal documents, allocates inventory, schedules nurses, routes deliveries. Every one of those decisions, when it succeeds, changes what a job looks like for thousands of people. Every one of them, when it fails, costs real money and real trust. Every one of them is being made right now, by humans who are mostly making it without a great map.

In aggregate, these decisions are reshaping how the actual economy operates. Not in five years. Now. The companies making them well are pulling ahead. The companies making them badly are quietly winding down lines of business. The new lines of business that are being born out of these decisions, mostly inside existing businesses, are where the next decade's wealth and disruption will come from.

I've been in a lot of these rooms in the last few years. The thing I keep noticing is that the conversation that shapes a decision is almost never about the model. It's about who owns the system after launch. It's about what the operating model looks like. It's about what happens to the team whose work gets reshaped. It's about whether the customer notices, and whether they like it when they do.

These are not technology questions. They're questions about how a business should operate. They're being answered by AI. They're determining what work looks like for everyone else in five years.

the lab and the field

There are two places in this current moment where the work matters most. One is the AI lab, where the underlying capability is being pushed forward. The other is the field, where that capability is being absorbed into the operations of the businesses that touch most peoples' lives.

These are different jobs. They require different temperaments. They produce different kinds of value, on different timescales. Both are necessary. The labs need brilliant people pushing the frontier. The field needs people who can take what the labs ship and figure out how to make it actually work inside real organizations.

For most of my career so far I've been in the field. I work with the operators figuring out what to do with this technology inside their businesses. I sit through the meetings. I write the proposals. I run the pilots. I watch the pilots succeed or fail and write down what I learned.

I think this is the more interesting seat for the next ten years, for a specific reason. The bottleneck on AI's impact on society is not at the model layer anymore. The bottleneck is at the absorption layer. The lab can ship a better model every six months. Whether that better model changes anything depends on whether someone in the field figures out how to thread it into a real workflow that real humans rely on. Most of the lab's potential impact is being left on the table because the absorption work isn't happening fast enough or well enough.

That's a useful place to put a career. The marginal value of one more lab researcher is real but small. The marginal value of someone who can take a model and turn it into a working production system inside a Fortune 500 is currently very high. There are not enough of these people. The field is wide open for the people who want to do this kind of work seriously.

what the seat looks like from inside

The reason I want to write more of what this looks like is that it's almost entirely undocumented. The lab side gets papers, twitter threads, podcasts, hour-long youtube essays. The field side gets vendor blog posts and case studies that have been through eight rounds of legal review. The actual texture of the work, the genuine lessons, the things that surprised the people doing them, mostly stays inside private rooms.

I want to push some of that out. Not because it's secret. Because it's useful, and there isn't enough of it, and the people who could write it down mostly don't.

A few things from this seat that I think don't get enough airtime:

The "is it good enough" question rarely turns on capability anymore. It turns on tolerance. How much wrongness can the workflow absorb? How fast can a human correct it when it's wrong? What is the cost of an error and how is it distributed? Most of the time, the AI is good enough. The question is whether the workflow around it is.

The operators who do the best work with this technology share a specific trait, which is that they ask a lot of dumb questions on purpose. They ask why the thing is the way it is. They ask what would happen if the constraint went away. They poke at the assumptions. The technology rewards this kind of thinking, because it's mostly removing constraints that used to be load-bearing.

Most of what looks like AI adoption inside a company is actually re-organization. The model is the lever. The actual change is in who reports to whom, what gets centralized, what gets decentralized, what jobs get redrawn. Companies that try to adopt AI without reorganizing get a tiny fraction of the value. Companies that use AI as the forcing function for the reorganization they needed anyway get most of it.

The people inside the labs underestimate how slow and political enterprise adoption is, and that's mostly fine. The people inside the enterprises underestimate how fast the capability is moving, and that's a bigger problem. Both groups would benefit from talking to each other more. They don't, much. There's a real opportunity in being a credible translator.

why this matters, and why I'm writing it down

I want to spend the next decade in front of this. I want to be in the rooms where the absorption decisions get made, because those rooms are where the shape of the next economy is being decided, and the shape of the next economy is the shape of the next society, and I'm not done living in this one.

The reason I'm putting this on the internet, rather than keeping it as a private operating principle, is that I think there are more people who would want this seat than have figured out it exists. The conversation about how to spend your career in the AI era is dominated by the lab path, because that's the loud one. The field path is quieter and at least as consequential and currently understaffed.

If you're a smart engineer or operator or researcher trying to figure out where to put yourself, consider the field. Not the lab. Or not only the lab. The work is real, the impact is outsized, the seats are open, and the view from inside is the most interesting view I've found.

I'll keep writing what I see from here.