Building a Neighborhood with AI

← Back to Aladdin

By Charles Xie ✉

A house is an object. A street is a system. Once you have designed a house by describing it, the obvious next wish is to describe a whole neighborhood — and that wish turns out to be a different kind of problem, not a bigger version of the same one. A house has to stand up and look right. A street has to do that eight or twelve or sixteen times over and get the relationships between the houses right: where one property ends and the next begins, which way each front door faces, how far apart two walls have to be before the strip of ground between them can hold anything, where the pavement stops and the public realm starts. Nobody notices those relationships when they are correct. Everybody notices when they are not.

This is also where a generative model, left to itself, is at its weakest. Ask an AI for eight houses in Florida and it will cheerfully give you eight of the same house, because each one is generated in its own conversation and none of them can see the others. Ask it for a neighborhood and it will happily set a driveway running across the middle of the road, because it has no idea where the road is. The interesting engineering in Aladdin's neighborhood builder is almost entirely about that gap: what has to be decided for the AI, in advance and deterministically, so that what the AI is left to do — imagine one believable house — is the thing it is actually good at.

What follows is a tour of the street: how one sentence lays it out, how the ground rules keep sixteen independent houses from colliding, how geography reaches the architecture, how the landscaping and the street furniture get there — and what the whole thing is for, which is that a neighborhood in Aladdin is a working building‑physics laboratory with a dozen buildings in it.

One sentence, one street

You type “build a neighborhood in Sarasota, Florida with ten houses” and the assistant does the rest. It sets the world to your location first, so the site has real weather and a real sun; it works out how many rows the request implies; and then it builds, house by house, telling you what it is doing as it goes. Two to sixteen houses, in one row, in the classic two facing rows either side of a street, or in a grid of up to four rows where the pairs share streets and back yards meet in the middle — which is how a real subdivision fills a block.

Everything after the first sentence is ordinary editing. The houses are not a frozen block of scenery: each one is its own building, with its own name, its own walls and its own garden. Select one and the menus work on it. Ask the assistant to repaint it, add a dormer, move it two metres, put panels on its roof — the neighborhood was built by a tool, but what the tool leaves behind is just your scene.

The planner and the generator

The trick that makes the whole thing work is a division of labour. Two very different kinds of software are involved, and each is given exactly the job it is good at.

Almost every failure this system has ever had came from putting a decision on the wrong side of that line. Leave the storey count to the model and it builds the regional cliché to uniformity — every house in Florida a single‑storey ranch, every house in Massachusetts a two‑storey colonial, in both cases because the model is reasoning about “a house in Florida” rather than about this street. Leave the width to the model and one generous plan closes the gap to its neighbour, and the whole vegetative buffer that was supposed to stand between them has nowhere to grow. Both are now the planner's business: the mix is dealt in advance, and the width budget is stated in the prompt and measured before the house is allowed into the scene.

Sixteen model calls would take a long time in sequence, so up to three run concurrently while the scene itself is written by one lot at a time. That single detail — generate in parallel, apply in series — is the difference between a street that takes ten minutes and one that takes half an hour, and it is also, historically, the part that took the longest to make safe.

Where the property line is

The unglamorous heart of the whole feature is a rectangle: the lot. Every house gets one, and almost every rule on the street is expressed as “stay inside it” or “stop exactly at its edge”.

Getting pavement to touch the street without crossing it turned out to be the fussiest detail on the page, and the fix is a good illustration of a general principle. The first version computed where the sidewalk ought to be, from the planned house depth and the planned yard. But the houses are AI‑generated, and a house with a porch on the front and a garage on the side does not sit where a plain rectangle would — so the calculated line missed the real sidewalk, and every walkway on the street stopped a metre short in the grass. The rule now is that pavement ends where the pavement is: the driveway and the walk are run out to the measured edge of the sidewalk slab and the measured edge of the asphalt that are actually standing in the scene. An estimate derived from a plan constant breaks the moment the model builds something the plan did not predict; the thing standing in the scene never lies.

Geography reaches the houses, not just the gardens

“Build a neighborhood in Massachusetts” and “build a neighborhood in Arizona” should not produce the same street with different plants. Aladdin classifies the site two ways, from two different sources, and layers them.

The climate comes from the weather station nearest your location — the same measured data the energy simulation runs on. It is a deliberately crude four‑way split, made from what the station files actually contain: a site whose coldest month still averages above 18 °C is tropical; one with more than 3300 hours of annual sunshine and a summer high of 32 °C or more is arid; one whose coldest month averages below −3 °C is cold; everything else is temperate. Phoenix records about 3870 hours of sunshine a year against Boston's 2630, so the gap is wide enough that the proxy is safe without any precipitation data.

The setting is something no weather file knows: Cape Cod and Worcester are both “temperate”, and nothing local can tell a coastline from a ridge. So the assistant classifies that one itself, from your words and the place you named — Nantucket reads coastal, Aspen reads mountain — and passes it in alongside the climate. Where a setting has an opinion it wins; where it does not, the climate stands. The two are combined rather than stacked: the garden cycle is the intersection of the setting's styles and the climate's, so a setting can never smuggle in something the climate forbids. A mountain neighborhood in the desert keeps its xeriscape.

Colour is regional before it is personal, so the palettes are too. Each lot is dealt a body, roof and trim scheme from its region's set, cycled so that the house beside you and the house behind you are never wearing the same one. The scheme is handed to the generating model as guidance rather than as hex codes — it matches the description in its own materials, which is why the street reads as varied rather than as a colour chart.

The dice that break the stereotype

A real street is not one house repeated, and it is not one house randomized either — it is a mix with a regional lean. Aladdin deals that mix from the lot number, using a plain deterministic hash rather than a random number generator, so a street you liked can be rebuilt exactly and a street you did not can be told apart from the one you get next time.

• ▲ Storeys — the lean follows the climate and every climate mixes: about 40% single‑storey in temperate and cold regions, 55% in the tropics, 60% in the desert. The rest get a full second floor.
• ▣ Front entry — roughly a third porch, a third portico, a third neither. Left to itself the model gives every single house on the street a porch.
• ▬ Garages — a quarter none, half single, a quarter double. A house with a garage gets a driveway; a house without one does not, and no car parks on its lawn.
• ⚊ Fences — about 60% of lots, style cycled on its own counter so an unfenced lot between two fenced ones does not leave a repeat.
• ● Water — one or two ponds on the whole street, and exactly one swimming pool, because most garden styles will lay a pond by default and an unconstrained street ends up with a water feature in every yard.
• ☼ Lamp posts — about half the lots keep the garden lighting pass's tall front posts. Not every house wears one, and the ones that do not are still lit from the street.

One thing had to be learned the hard way here, and it is worth telling because it is a lesson about AI systems in general rather than about neighborhoods. The dice worked perfectly — and then the assistant described a Florida street it had just built as “ten single‑storey Florida ranches”. Measured in the file, six of the ten were two‑storey. The geometry was right; the reporting was blind. The summary the tool handed back to the model listed styles, gardens, cars and trees, and said nothing at all about storeys — so with no facts in hand the model fell back on the cliché, which is exactly the uniformity the dice exist to break.

Everything between the houses

The houses are the easy half. What makes a rendered street believable — and what makes it physically interesting — is the ground between and around them, and all of it is laid deterministically once the houses are standing.

The vegetative buffer

Between every pair of neighbouring houses runs a three‑layer planting band, and it took several rounds to work out that a buffer is not a row of trees. Shade canopy floats overhead — a trunk screens nothing at eye level. A shrub row tops out below the window sills. Privacy lives in the band between those two, which is exactly what an evergreen with foliage to the ground fills. So each shared boundary now carries all three: shade trees, a paced row of screen conifers about three and a half metres apart, and a continuous hedge‑class understory beneath them, every species chosen to be part‑sun because half of any boundary row stands in a building's shadow.

There is a piece of building physics hiding in that geometry, and it is free. Because the rows face each other across a street, the boundaries they share always run north‑south — which means the buffer stands on each house's east and west flanks and never on its south face. Morning and afternoon sun is exactly what a canopy should intercept in summer; the low winter sun that a house wants coming through its south glazing is never touched. A screen of trees planted on the wrong side of a house is a heating bill; on these sides it is a cooling benefit. You can test that claim in the model, which is the point of the whole exercise.

The public realm

Beyond the property line the street dresses itself: paver sidewalks behind grass verges (decomposed granite in the desert, where a turf strip makes no sense), kerbside trees on the planting strip whose species follow the site, a double‑yellow centreline painted as flat markings on the road rather than laid as slabs on top of it, crosswalks at both ends, a hydrant, a 25 MPH sign at the entrance and a STOP sign at the exit. Davit street lights stand at every lot boundary, alternating sides, with the arm reaching out over the carriageway — and they carry real lights, so the street works at night.

Then the street furniture that belongs to individual houses: a mailbox on the verge beside each drive, and recycling and trash bins on every third one. Street trees are planted last, after every garden is in, and any tree whose station lands on a driveway, in the road, or hard against a wall is simply skipped — a gap in a row of trees is what a real street looks like, and it is far better than a tree shifted onto the next unknowable thing.

Street life

Cars park on about two‑thirds of the driveways that exist, drawn from an ordinary residential mix; a few more stand at the kerb, on the boundary lines where no driveway is, facing the right way for the side of the road they are on; and a third of the lot columns carry a car in a travel lane, passing through. Exactly one of the kerb cars is a school bus. People stand about — a figure or two in some front yards, pushed clear of the walk to the door, a few more on the verges, a dog walker on one sidewalk and someone with a stroller on the other.

None of this is decoration for its own sake. A street with nothing moving on it reads as an architectural model; a street with a bus at the kerb and someone walking a dog reads as a place, and students engage with a place.

And then it is a laboratory

All of that would be a pleasant scene generator and nothing more, except for one fact: every house on the street is a genuine building model. The walls have layers and U‑values, the windows have solar heat gain, the roofs have orientation and pitch, and the whole thing sits at your real latitude under your location's own measured weather. Aladdin's building‑energy engine is checked against ASHRAE Standard 140 (BESTEST), the reference suite whole‑building simulation programs are validated with. So a neighborhood is not a picture of a neighborhood. It is a dozen buildings you can run.

What that buys you is a set of questions a single house cannot ask, because they are all questions about neighbours.

• run the energy simulation for the whole year on the third house

  A day out of every month, hour by hour, on the site's own weather — with the houses either side of it and the boundary trees in the shading calculation, because they are standing in the same scene. A house in the middle of a row is not the same house standing alone in a field, and the model knows it.

• what happens if I take the trees off the boundaries?

  Delete them and re‑run. The buffer is a testable design decision, not a landscaping preference: it shades east and west walls in summer and leaves the south face alone in winter, and the annual number tells you what that is worth in this climate.

• show the solar radiation heatmap over the whole street

  Every roof and every patch of ground, coloured by what it actually receives across the day — which is how you find out that the north row's roofs are not the same as the south row's, and which houses have a roof worth covering.

• put solar panels on the south-facing roofs and tell me what the street makes

  A subdivision's worth of rooftop PV, on real roof planes at real tilts and azimuths, with the neighbours' shadows in the way where they fall. Ask for a solar neighborhood up front and roughly a third of the houses are built with an array already on them.

This is the part that makes a neighborhood worth building rather than just fun to build. Orientation stops being an abstraction when half the houses on your street face the other way. Lot density stops being a planning-board word when you can widen the pitch and watch the shading change. The trees stop being scenery when you can measure them. And all of it starts from a sentence a fifteen‑year‑old can type.

What sixteen houses at once actually costs

A short, honest word about the engineering, because a feature like this has a shape that is easy to underestimate. Every house on the street is a full generation from a language model, and generation dominates the wall clock completely — all the local work, all the layout arithmetic, all the planting and the street furniture, runs in a fraction of a second. So the interesting costs are all in the model calls.

That has consequences you can feel. Early on, the brief handed to each house grew a paragraph of well‑meaning rules, and generation time went from about thirty seconds a house to well over a minute — nothing was wrong, the model was simply deliberating over prose it had been given. Trimming it back to one terse paragraph halved it again. Sub‑turn prompt clauses are not free, which is a fact worth knowing for anyone building on top of these models.

The other consequence is that failures have to be survivable. A model call can wedge itself: it answers in prose instead of building anything, it hits a rate limit, it produces three candidates in a row that break the lot's budget. Each lot therefore gets up to two entirely fresh attempts — a new conversation each time — and if it still will not land, the run reports that lot as empty instead of quietly leaving a gap you discover later. Ten houses asked for and nine delivered is a defensible outcome. Ten asked for, nine delivered and ten claimed is not.

Conclusion

The step from one house to a street is where AI design tools tend to stop being impressive and start being useful, because it is where the constraints show up. One house can be judged on whether it looks good. A street can be judged on whether the driveways reach the road, whether the fences stay on their own property, whether anything grows between the buildings, and whether the description you are given matches the thing that was built. Those are ordinary engineering questions with right answers, and the interesting discovery in building this was how many of them the generative model should never have been asked in the first place.

So the neighborhood builder is a small argument about how to use these systems: give the model the job that needs imagination, give the code the jobs that need to be exactly right, and make the boundary between them explicit. What comes out is a street where every house was dreamed up separately and every relationship between them was guaranteed — and then, because this is Aladdin, you can select any one of those houses, run a year of weather through it, and find out what living there would cost.

Related reading: Designing your dream home with AI · Generative urban design · Validating against ASHRAE Standard 140 · A new way of design

← Back to Aladdin