“Let’s kick the crap out of this,” I said, mostly to myself, and typed a deliberately constructed series of numbers into ChatGPT 3.5, followed by:
Forecast the next three numbers in this series. Use five models, one of them Holt-Winters, one GARCH. Show me each output and the average, pick the best and tell me why. Then write the Python to put it in production, with the deployment steps.
I hit enter expecting to break it.
Twelve seconds passed. My screen filled.
I pushed my chair back hard enough that Will Willis, a colleague sitting next to me, looked over.
“What’s up?”
“We are all fucked.”
That was all I had.
It came out quiet. Steady, almost calm, the way you say a thing you already know is true.
It was late 2022, right after ChatGPT launched. The news was already full of predictions about what it might become. I had spent more than twenty-five years working at the intersection of strategy and technology, much of it around analytics and machine learning, usually trying to get in front of the next shift before everyone else did. I had done it often enough to think I understood the pattern: get there early, learn faster than the market, and move again when everyone else catches up.
“Can I get ahead of this?” I wondered.
In those twelve seconds, ChatGPT had used the two models I specified, selected three more and run all five. It gave me the individual forecasts and their average, compared the results, picked the approach it thought was best and explained why. Then it wrote the Python and laid out how I could deploy it.
The code was good. It wasn’t perfect. That barely mattered.
What stopped me was the range. I had given it a problem that crossed statistics, model selection, software development and deployment, and it moved across all of them without hesitation. Work that would normally bounce between people with different skills had come back from one machine in twelve seconds, complete enough to be useful.
When I said “we,” I meant everyone. Jobs, companies, whole categories of work. My thoughts had quickly moved past the forecast on the screen. “What happens to work? What happens to the value of expertise? What happens when people can suddenly do things they were never trained to do?”
And eventually, “what are we for?”
I had to know more, and I wanted to prove myself wrong. So I started reading. Technical papers first: transformers, scaling and emerging capabilities. Then I widened out to lab research, economics, forecasts, and a few essays that made me stop and rethink what I thought I understood.
I eventually cited more than fifty sources I considered important in a paper of my own. The more I read, the harder it became to dismiss what I had seen.
The capability was real. The rate of improvement was real. The implications were broad enough that I wanted to put a stake in the ground while I could still remember what I believed before hindsight cleaned it up.
I was fortunate to work in the Innovation Lab at Capital One and leadership encouraged that kind of thinking. I wrote the paper, made a formal forecast about significant labor displacement and its impact on credit quality, and that paper eventually went to the board.
Privately, I thought unemployment could reach fifteen or twenty percent.
I figured we would all end up on a beach somewhere, collecting a check and trying to work out what to do with ourselves.
“Bring it on,” I said. “I’ve got mountain bike PRs I’m chasing and my surfing skills desperately need to improve.”
Almost four years later, here we are. No beach. No universal check. We have more meetings than ever.
I got the timing badly wrong. That turned out to be useful.
Working out why I missed has occupied a good part of my time, and somewhere along the way I realized I had been watching one clock.
There were two.
A VP of planning and policy at the University of Richmond helped me see the second one. She put it in a single line: human beings and organizations are terrible at absorbing change of this magnitude.
The fast clock belongs to the technology. It has moved about as aggressively as I expected and, in some areas, faster. The ChatGPT I used that afternoon now looks primitive. Models reason better, work across modalities, write and debug software, use tools, search, analyze, create and increasingly act. That clock keeps accelerating.
The slow clock belongs to everything that has to absorb those capabilities before they produce consequences at scale.
I had assumed that once people saw what the technology could do, adoption would race behind it. A few months for the tools. A year or two for companies. Then the labor effects would arrive.
I underestimated what stood between the demo and the institution.
Companies have data scattered across systems that disagree with one another, technology built at different times for different reasons, budgets with little room for innovation and change, and incentives that reward one group for moving while another is paid to slow things down. Then come the controls, auditors, regulators, lawyers, customers and employees who all have legitimate reasons to care what happens when the machine gets something wrong.
And people have jobs they care about.
Some of that friction is bureaucratic sediment. Some of it is useful. At enterprise scale, mistakes get expensive fast. A model can hallucinate in seconds. A bank, insurer or healthcare company can turn one bad decision into thousands or millions before anyone realizes what happened. The controls that make adoption maddening can also keep an organization from being wrong at a scale that can sink it.
Once I understood that, I stopped asking why companies couldn’t move as fast as the technology and started trying to understand what had to be true before they could move faster safely. I started exploring the gap between the two clocks.
A friend and I had built an app called Horse Match. We could never fund the level of development it really needed, and it was frustrating to watch market share gains disappear because of technical problems we did not have the resources to fix quickly.
At the same time, I had been a few years removed from being truly hands-on with technology. I wanted to get back under the hood of AI, partly because I was curious and partly because I no longer trusted myself to talk about what it could do without building with it.
So I started building.
At first I was a very expensive copy-and-paste robot, hauling outputs from one window to another because the future had arrived with remarkably little plumbing. I put things into production I did not fully understand.
I broke prod three times.
The first time, I didn’t know I had broken it for a week.
That will become its own story, because breaking something you barely understand teaches you quickly what the Instagram ads about building an app with one prompt leave out. I learned to think about detection, recovery, isolation and verification. I also learned that working near the edge of your knowledge gets a lot less terrifying once you build systems that assume you could be wrong.
Slowly, the pieces connected. The manual steps disappeared. Checks appeared. Recovery became part of the design. Governance moved closer to the work. What started as me carrying output between windows eventually became a small AI factory, governed and boring in the way that working things are boring.
That changed how I think about my original prediction. I still believe AI will eliminate a large number of jobs. I do not believe enough replacement work will appear quickly enough to make the transition painless. I hope I’m wrong.
Even if the net job count eventually proves me wrong, the transition itself will be messy and enormous. Roles will change. Skills that took years to build will lose value, and the new skills needed are just starting to be understood. Companies will reorganize. Education, regulation and social expectations will struggle to keep up. That mess matters almost as much as where we eventually land.
I learned more from missing my first prediction than I would have learned from getting lucky and calling the date, because the miss forced me to understand something I had ignored.
Technical capability does not tell us when consequences arrive.
The fast clock is already far ahead of most people and organizations. The slow clock keeps grinding forward, well behind.
The distance between them is still open. It will not stay open forever.
I have spent much of the last four years trying to understand what lives inside that distance, what slows it down, what should slow it down, and what happens when you start removing the friction that no longer earns its keep.
That is what I want to write about here: the things I built, the things I broke, the things I was certain about and turned out to be wrong about. The places where AI is every bit as extraordinary as advertised, and the places where getting it to do useful work is still painfully, stubbornly human.
I still don’t know where the two clocks meet.
I intend to be there when they do.