Motion, Not State: The “F-16” Problem in Mortgage
- Marvin Chang
- 22 hours ago
- 5 min read
The F-16 was designed to be aerodynamically unstable. Without fly-by-wire systems making continuous corrections, the aircraft literally could not fly. This instability was not a flaw but a deliberate design choice. An unstable airframe, properly controlled, is extraordinarily agile, always on the edge of chaos, continuously caught by a system that manages motion, not state.
The instability is the feature, and it shows. Over half a century since entering service, the F-16 remains one of the best performing fighters in the world.
I was recently back at Carnegie Mellon for a reunion, so I’ve been in pensive mood. Back in the 90's, my program in Information & Decision Systems was loose, even flabby, by today's standards. The objective function was unclear, the program still finding its identity. The school has since become a computer science juggernaut, substantially more streamlined.
My sons find CMU equal parts amusing and weird. Too singularly focused on computer science, too left-brain. Regardless of truth, it is a perception worth sitting with, because it captures something real about what happens when an objective function clarifies too completely. The school I attended was loose enough that I could wander. I am not sure today's students feel the same permission.
I wandered into decision sciences. Expected utility theory. Optimization under uncertainty. Courses with no obvious relationship to a job offer. Decision sciences was about motion, while a technical skill like databases was about state. At the time it felt like inefficiency. In retrospect it was the most valuable thing I did.
Between 2018 and 2021, mortgage companies optimized brilliantly. Point-of-sale systems tuned for speed. Automated underwriting for agency-conforming flow. Headcount scaled to volume, comp restructured around refi economics, digital origination as the dominant channel. The objective function was unambiguous. And then the distribution shifted.
Rates inverted in 2022 with no modern precedent. Refinance volume collapsed 80 percent. The purchase market that remained required complex credit judgment, non-conforming products, relationship-intensive origination. Technology and process alone cannot substitute for.
This is the overfitting problem. A model tightly trained on its sample data performs brilliantly within that distribution, potentially catastrophically outside it. It has optimized for the regime rather than for adaptability across regimes. The industry dislocation was not because mortgage companies were badly managed. Many were exceptionally well managed – for the world that existed. They were highly optimized airframes headed into an environment requiring utmost agility.
An optimized, stable airframe cannot run a fast decision loop.
I wrote recently about what I called the second game in mortgage – playing out on the axis of timing, not certainty. The market moves just enough. A loan becomes refinanceable. One lender spots it at 10:17 AM and sends an offer already structured. Another assembles a batch campaign that afternoon. The borrower is already gone. What separated them was not product, pricing, or brand. It was the decision loop - the interval between when the opportunity became real and when they acted.
A stable organization manages to state. It describes what has already happened. It is always reacting to yesterday. The companies that optimized hardest for 2021 built organizations structurally challenged for running the loop the current market requires. In a market where the moment is fleeting, the loop is not just part of the process. The loop is the business.
A reasonable objection: independent mortgage banks operate on a knife's edge. Slack is a four-letter word. The graveyard of innovation labs – skunkworks projects euthanized when the host organism rejected them – justifies skepticism about any argument beginning with “invest in an adjacent capability.” This objection is fair but is no longer sufficient.
Rocket's recapture rate runs roughly three times the industry average. The largest servicers hold portfolios exceeding a trillion dollars, with AI investment and continuous decisioning capability that scale enables. These competitors have already collapsed the interval between signal and action into a structural advantage - and are widening that gap every quarter. The old excuse that we can't afford slack assumed a competitive set that also couldn't afford it. That assumption is no longer valid.
Which brings us to artificial intelligence.
The standard AI adoption frame is efficiency: automate the repetitive, reduce cost per loan. Legitimate gains. But AI applied to a stable, overfitted organization accelerates the state-management problem. A faster batch campaign is still a batch campaign. A more efficient dashboard still described yesterday.
AI can also be something else entirely. It is an opportunity to unlock slack the organization cannot generate on its own.
During my weekend at Carnegie Mellon, in between rounds of home-made beer – we are talking about engineers – I found an impromptu focus group of senior technologists at several multi-trillion-dollar companies among my fraternity brothers. They pulled out their laptops to show me how their work has been utterly transformed in six months. Six months ago, they were assembling code. Today they are managers of AI agents – orchestrating, directing, even debating. What struck me was not efficiency. It was the multiplication of capacity: one person prosecuting multiple streams of ideas simultaneously, running experiments that a short time ago would have been deemed too expensive to attempt.
This is AI as the fly-by-wire system – the control surface that makes deliberate instability safe, that allows the organization to operate closer to the edge of its capabilities without departing controlled flight. The institutions that navigate this transition will not be the ones that automate fastest. They will be the ones that understand what AI enables: not a more efficient stable airframe, but the computational infrastructure that makes the unstable one viable.
The “flabby” Carnegie Mellon curriculum was the instability. The instability was the agility.
The question for every mortgage leader is not how to rebuild the 2021 operating model at lower cost, or how to automate fastest. It is whether your organization is a stable airframe or an unstable one, and whether you are using AI to optimize the former or to finally make the latter possible.
Stable airframes are efficient in calm conditions. Unstable airframes are harder to fly. They require better systems, more skilled operators, and tolerance for never quite being in equilibrium.
But when the environment turns, and it always turns, only one of them can maneuver. Only one of them can close the loop before the moment disappears.
That is the lesson CMU's curriculum taught me serendipitously in 1991. The lesson the rate cycle taught mortgage companies at great cost in 2022. And the lesson AI is offering right now, not as a threat to be managed, but as the instrument that finally makes your mortgage company into an F-16.