Towards Housing Abundance

Walk through a typical subdivision, and you can feel the problem.

The homes are expensive. The lots are tight. The stakes are high. And yet so many communities still feel less than ideal: garages dominating the street, porches no one will use, kitchens cut off from family life, backyards stitched together by leftover strips of land that exist because of a development compromise.

I can say with certainty that this is not because builders are unsophisticated. It is because they are operating without the data infrastructure that modern systems and products benefit from.

A typical production builder builds one or more homes each week. With each home, they are making a series of irreversible bets. Which plans will sell? Which elevations will work on this lot? Which options will customers actually choose? Will the street feel coherent? Will the homes appraise? Will the construction team be able to deliver the design reliably? Unlike software, if they get it wrong, they do not just ship a buggy release that they can hotfix when it lands. They spend hundreds of thousands of dollars building a bad answer into the ground, and then all of us — designer, builder, and customer — live with the damage for decades.

So they do what people do in high-risk systems with weak feedback loops: they operate on intuition and the little data they have. They lean on precedent, pattern matching, and organizational memory. They feel forced to reduce the risk of delivery above all else — not because this produces the best neighborhoods, but because it reduces the chance of catastrophic mistakes.

Why hasn't AI had an impact on housing?#

The field is on the verge of automating legal research, software engineering, and drug discovery. But we still cannot reliably use AI to design a meaningfully better house. Not because housing is too important. Not because housing is too emotional. And not even because homes are too complex. A single-family home is complex, but this is not the blocker. Have you tried to get the state-of-the-art LLMs to design you a home? Utter failure. Not seven-fingers-on-a-hand failure — a complete and total breakdown of capabilities. Spatial synthesis is beyond the wall of the general language-model paradigm.

The real reason is more technical: housing is out of distribution.

In machine learning, a domain is in distribution when there is structured data to train on, shared representations researchers can build against, and benchmarks that make progress legible. The biggest AI breakthroughs appear only after those preconditions are met. Protein folding needed a common structural vocabulary and a benchmark. Weather prediction needed standardized atmospheric data and open datasets. The same pattern shows up again and again. Our collective lack of structured housing data puts the industry at risk of missing another wave of the technology revolution.

Housing has almost none of what is needed for an AI takeoff.

There is no common schema for homes. No large open corpus of structured home designs. No standard benchmark suite. No broadly shared representation that lets one team's work compound with another's. More than a million single-family homes get built in the United States in a typical year, and as far as modern machine learning is concerned, most of that activity barely exists. Even when frontier-model benchmarks brush against architecture and engineering, the results lag domains with richer training distributions.

For researchers, this is unacceptable. Housing is one of the largest and most consequential opportunity surfaces in the economy, yet it is nearly invisible to the methods transforming other fields.

Why? Because homes are still trapped in lines.

Today, most housing design still lives in CAD drawings and loosely connected documents. But a CAD file is not a home in a form that an LLM can reason about. It is geometry without meaning. A wall is just a pair of lines. A room is an implied shape. A window is a symbol dropped at a coordinate. The semantics — what the function of the room is, how it connects to other rooms, whether it works for a family, whether it complies with code, whether it captures morning light — all live in the head of the human looking at it. As a result, the breakthroughs we have seen in code generation and dozens of other domains are out of reach until we have a way to describe homes as data — as text.

Housing needs a representation breakthrough#

Housing needs a schema where a home is composed of typed, semantic entities: site, lot, setback, floor plan, room, wall, opening, stair, roof, option, facade, and community. It needs relationships and constraints that can be queried. It needs a representation where an AI model can ask and answer real questions:

• Which bedroom shares a wall with the laundry room?
• Does this kitchen receive morning light?
• Which plans fit this site while preserving privacy in the rear yard?
• What changed in cost when this option set was selected?
• Which layouts correlate with stronger resale or lower maintenance?

Once you have that representation: endless opportunity.

Building code can be treated like rule-checking in software. A setback rule is no longer a judgment buried in an inspector's memory. It becomes a function over a design state. Egress, energy, spacing, adjacency, and structural rules become testable checks. Permitting begins to look more like continuous integration than ceremony. The home is inspectable and optimized before the design is constructed in the field.

With structured data, the research surface opens.

Patterns of successful home designs are translated into a form that can be trained on. Meaningful benchmarks can be built. Models can be trained to optimize layouts, infer code risk, estimate cost, predict option uptake, evaluate streetscapes, and generate communities that are not just unique, but better, buildable homes. Researchers can compare methods against shared tasks instead of hiding inside bespoke demos. Our work compounds.

A schema alone is not enough#

Housing also lacks a commons. In every domain where AI has accelerated quickly, someone had to open the data. Genomics had public sequence data. Weather had public atmospheric data. Mapping had shared geospatial layers. Housing remains locked inside architecture firms, builder servers, permitting offices, consultants, and PDFs. The plans exist! The demand exists! The problems are economically massive. But the data layer is fragmented and semantically weak.

That fragmentation has cultural consequences. Because the product is not represented as structured data, builders cannot really instrument design quality the way software teams instrument product quality. They do not have robust feedback loops tying plan decisions to long-term outcomes. Did buyers actually value that flex room? Did that garage-forward frontage hurt absorption? Did those narrow side yards reduce satisfaction? Did a certain kitchen adjacency lead to better daily use? Most of this is barely measured, rarely shared, and almost never fed back into a system that can automatically improve future designs.

So the industry learns slowly and painfully. A developer can spend years acquiring land, entitling it, financing it, designing a community, and building it out, only to discover they produced something that technically works but feels dead. Everyone in housing has seen this outcome. Streets that feel hostile. Homes that sell because of scarcity rather than delight. Neighborhoods that function as financial products instead of places to live and grow.

That should bother AI researchers. Because housing is not a niche application. It is the physical interface for family life. It determines how people spend money, time, energy, attention, and emotion. It shapes childhood, commuting, connectivity, privacy, neighborhood cohesion, and household resilience. If AI is going to impact the physical world, this is our biggest opportunity to create a positive shift.

And unlike many saturated research areas, housing still offers first-principles problems with real leverage. The representation problem is open. The data problem is open. The benchmark problem is open. The tooling problem is open. The feedback problem is open. The market infrastructure is still weak enough that even basic programmability around cost and procurement is rare. This is not a domain with a thousand over-optimized benchmark papers. It is a frontier with room for new primitives.

The call to action#

Housing is one of the most important domains that AI has barely touched — not because the opportunity is small, but because the substrate has been missing. The chance now is not merely to generate "prettier" plans. It is to create the shared representation, the data commons, the benchmarks, and the optimization loops that make housing legible to modern machine learning.

If that happens, the payoff is not just better models. It is a better home. Better communities. The ability to take on risk with confidence in one of the most expensive and consequential products most people will ever purchase.

The housing domain does not need another slick demo. It needs researchers ready to treat housing as a serious technical domain.

The opportunity is enormous. The infrastructure to deliver on the vision is ready. Come build it with us.