On Day 1 of SusHi Tech Tokyo 2026, on the Other stage, four practitioners — an urban designer, a film producer, a digital-twin engineer, and a Hollywood screenwriter — sat down to make a strange but interesting claim. The proposition was that the real-estate development industry has been running an outdated economic model, and the fix is to design cities the way you'd design a movie. Story first, persona first, eye-level experience first.
It is a compelling pitch. It is also, in my read, partially right and partially the kind of conference idea that sounds better on stage than it does in production. Let me walk through what the panel proposed, and then say where I think the operator lens diverges from the panelists' framing.
The Speakers
Ken Shigimatsu is the founder of Inspiring Dots, an urban-design studio with masterplan projects in more than ten countries. He coined the term cinematic urbanism and was the panel's intellectual anchor — proposing nothing less than a paradigm shift in how the development industry thinks about land economics.
Kenji Yamada is a producer at Tykon whose credits include several Hirokazu Kore-eda films and Godzilla Minus One. He has spent his career getting Japanese cinema in front of international audiences and brought the craft-of-narrative perspective that connects storytelling techniques to spatial design.
Aki Take is the CEO of Space Data, a Japanese company building digital-twin software powered by satellite data and AI — including simulations of the International Space Station and lunar bases. He spent earlier years in the Japanese government before founding the company. He provided the technical-feasibility leg of the discussion: LLM-agent-based population simulation that can test how people will move through a city before construction starts.
Gita Pulpi is a Hollywood screenwriter, director, and producer at Team AMG Productions, known for emotional storytelling and films like Queen Pins. A former journalist, she brought the cross-cultural craft perspective and a sharp take on why chaos and friction make cities and characters worth caring about.
The Floor-Area Problem the Panel Is Trying to Solve
The pitch starts with a real diagnosis. Most contemporary development runs on a single economic model: maximize the floor area on a given site, bring in a designer to make the resulting building "nicer," fill it with offices and retail, exit at completion. This model is showing strain. Construction costs have climbed. Office demand fell after the pandemic and never came all the way back. And the developments themselves keep destroying the very neighborhood character that made the sites valuable in the first place.
Shigimatsu's argument is that the floor-area model is now leaving money on the table. Narrative-driven development — design first, persona first, experience first — produces long-duration value appreciation rather than a one-time spike at completion. In a world where finished projects are increasingly indistinguishable from one another, the projects that have actual identity hold their value over decades.
I think the diagnosis is correct. The prescription is where the conversation gets interesting.
The Cinematic Urbanism Methodology
The methodological core is a shift from bird's-eye to eye-level. Most urban planning happens above the city — GIS maps, site plans, floor-area ratios. Shigimatsu's team starts from the ground. They build five distinct personas per project — specific imagined residents with jobs, families, daily routines — and trace each persona through their hypothetical day in the planned development. The buildings, the streets, the businesses, and the green spaces get designed to serve those characters rather than to maximize abstract square footage.
This is genuinely the way UX design thinking moved through software product design twenty years ago. Before personas, software was built around feature lists. After personas, software was built around specific users with specific goals. The result was better products. Shigimatsu's bet is that the same shift can now happen in physical space.
The shift makes sense in principle. The bigger question is whether the methodology actually changes outcomes when it meets the economic reality of who is paying for the project, what they need back in returns, and on what timeline.
The Technical Bet — LLM Agents as Population Simulators
The most ambitious idea on the panel came from Aki Take and Space Data. The company is running an experiment that, depending on how it matures, could become a new class of urban planning tool: release multiple LLM agents — each representing a different persona — into a digital-twin city environment and observe how they behave. Because LLMs are trained on human language and a great deal of human behavioral data, they can approximate how different people might navigate spaces, choose businesses, and form social patterns.
Applied to a proposed development: build the city in the digital twin, populate it with agents, watch where the agents go. If none of the simulated residents end up frequenting the cafe the developer planned to put on the ground floor, the assumption is that the design is wrong.
I want to take this idea seriously because the same pattern is showing up everywhere right now — synthetic users for software product research, synthetic respondents for survey panels, synthetic customer interviews. Real user research is expensive. Recruiting, screening, scheduling, paying participants, running the interviews, synthesizing the output — the time and cost add up. There is a real incentive to substitute LLM-generated synthetic populations for the real thing.
And here is where I have to push back.
The Brick House Problem
A synthetic population is, by definition, an idealized version of what we already think we know about how people behave. The agents are trained on previous data, and you usually do not know precisely what data was used in their training. The model behaves the way the average of its training set would behave. That is the entire bet.
But humans, in real life, often do things that look irrational until you understand the environment they are operating in. I read once about a development pattern in parts of the world where families build their houses one brick at a time — buying a single brick when they have a little money, storing it on the construction site, and building the wall slowly over years. From a Western financial-planning perspective, this looks completely irrational. Why would anybody buy a brick instead of saving the cash in a bank?
The answer is that for many of these families, there is no reliable bank to save in. Inflation eats cash savings. Currency devalues. The local bank might fail. But nobody is going to steal a brick out of your half-built house. Buying the brick is a way of converting wages into a stable, illiquid, theft-proof store of value. It is one of the most rational financial decisions a person in that environment could make.
The problem for synthetic populations is that the LLM agent does not know about the brick. The LLM is trained on a global corpus that is dominated by Western financial assumptions. It will probably model the brick-buying family as a confused outlier when in fact the family is responding sensibly to its actual constraints. And the worst thing about this kind of mistake is that you do not know you are making it. The synthetic population looks like it is producing useful predictions. The predictions look reasonable. They just happen to be wrong about the audience that actually matters.
This is not a theoretical concern about LLM accuracy. This is the concrete operational risk of using synthetic populations to validate any design decision that is going to be deployed in a specific community — a city, a neighborhood, a software product. You can end up building a behavior that is optimal for the synthetic population and wrong for the real one. The synthetic population cannot tell you about its own blind spots. Real users can, even if it takes more effort to get the signal out of them.
So when Space Data releases LLM agents into a digital twin city and watches what they do, I find the work interesting and I would not yet stake a ¥100B development decision on it. The right use of synthetic populations, today, is probably to narrow the question space — to identify which hypotheses are worth bringing to real users — rather than to replace the real-user step entirely.
What Cities Should Actually Be Designed For
When I think about the cities I have lived in and the neighborhoods that work, the design principle is not narrative. The design principle is resource access. Can residents reach the things they need without having to travel very far? Is the post office, the supermarket, the pharmacy, the hospital, the dentist, the mailing service all within a reasonable walk? If you need a car, is there parking? If you do not need a car, can you actually live there without one?
The neighborhoods that get this right have a different character from the ones that do not. Walkable neighborhoods feel alive because residents are crossing paths with their neighbors as a side effect of running their actual errands. Car-dependent neighborhoods feel sterile because the only place anyone meets anyone is the parking lot. The "narrative" of a place is largely a downstream consequence of whether the resource access works. Get the function right and the story builds itself.
I think Shigimatsu would not strongly disagree with this — his persona-driven methodology is partly a way to surface resource-access questions that floor-area maximization ignores. But the framing of "design cities like movies" puts the emphasis on the story when the emphasis probably should be on the day. What does this person have to do today, and can they do it from here without spending two hours in traffic?
On Gray Zones — Observation Beats Intention
The panel made a beautiful argument for protecting "gray zones" — the regulatory ambiguity where Tokyo's yokochō alleyways, informal chess corners, and guerrilla plant installations generate the texture that makes the city feel alive. The argument is correct. Over-regulation does kill culture. Sanitized, over-optimized cities feel interchangeable because they are.
Where I want to push back a little is on the framing of designing gray zones. I do not think designed gray zones work, because the moment you intentionally allocate a zone for ambiguity it stops being a gray zone and becomes a programmed authenticity exhibit. The gray zones that actually generate culture are the ones nobody planned — the moments where citizens did something unexpected with a space that the planners had not anticipated.
The operator version of this principle is closer to observe than design. The best products I have seen leave room for the user to surprise the creator. Early Instagram was a photo-filter app for photographers. The use case that actually mattered — daily life documentation — was something users invented after the product shipped. The team that built Instagram did not "design a gray zone for unexpected use." They built a tool with enough flexibility, watched how people actually used it, and adapted.
A cup is an even simpler example. The designer of a cup intends it for drinking water. The user, depending on their life, might use it to hold paintbrushes, to organize coins, to start a sourdough starter, to anchor a stack of papers, or to keep a houseplant alive. The cup did not need a "gray zone for unexpected use" written into its specification. The cup needed to be useful enough to hold things and simple enough that the user could decide what to put in it.
Cities are the same. The gray zones cannot be designed in. They emerge when residents are given the latitude to use space for purposes nobody specified, and when the regulatory environment is permissive enough to let those purposes survive without being made illegal. The design move is to not over-specify, not to allocate a specific block for designed ambiguity.
The Disaster Use Case — The Underrated Application
The piece of the panel that I think has the clearest near-term commercial path was almost a side comment from Aki Take: Space Data's digital-twin work for disaster management. The pitch is that simulating an evacuation — actually visualizing where 1000 people would move during a flood, which routes get blocked, which shelters fill up first — produces a different kind of decision than an abstract flood-zone classification map.
I find this version of the work much more credible than the cinematic-urbanism application. The disaster simulation does not need the synthetic population to be a perfect model of an individual citizen's psychology — it just needs to model the physical bottlenecks that emerge when many people try to move through specific spaces. That is much closer to fluid dynamics than to human behavioral prediction. And the buyer is clear: municipal disaster-management offices, national agencies, the UN. The product fit is concrete.
If I were prioritizing inside Space Data, I would push the disaster-simulation business hard while keeping the cinematic-urbanism work as a research bet that may or may not mature into something deployable in five years. The disaster work feels like it could be a business in twelve to eighteen months. The narrative-driven city planning feels like a thesis that needs to be tested on several actual completed projects before it can claim to be a methodology.
What This Means for Operators
For founders, builders, and operators reading this, here are the takeaways I would carry from the panel.
On synthetic populations. Use them to narrow your hypothesis space, not to replace your real-user research. The synthetic population can tell you which questions to bring to real people. It cannot tell you what the real people will actually do, because the synthetic population does not know what it does not know. The brick-house family is invisible to the LLM until a real person tells you about them.
On product design. Resource-access logic and observable behavior almost always beat narrative-driven design. The story is what users tell each other about a product that already works. The story does not, by itself, make the product work. Get the function right, watch how users behave, adapt to what they do, and the story takes care of itself.
On gray zones. You do not design ambiguity into a product. You design the product simply enough that users can find their own ambiguity, and you watch for the unexpected uses that emerge. The good product manager spends as much time studying how users surprise the team as on shipping the features the team already planned.
On the deep-tech bet itself. Space Data's LLM-agent simulation is the kind of work that may produce a transformative tool five years from now and may just as easily turn out to be a smart-looking dead end. The honest read is that it is too early to know. The work is worth tracking. It is not yet worth basing your urban policy on.
A Final Word on the Panel
I want to be clear that the panel was thoughtful and the panelists are doing serious work. Shigimatsu, Take, Yamada, and Pulpi are all clearly grappling with the right questions — what is the development industry getting wrong, can we replace floor-area economics with something more humane, what role do new tools like LLM agents play in city planning. The questions are correct. Some of the answers are still being worked out, and that is exactly what a conference like SusHi Tech is for.
The honest operator move, when you watch a panel like this, is to take the diagnostic seriously and stay skeptical of the prescription until it has shipped. The diagnostic — that current development economics is breaking — is correct. The prescription — that the fix is to design cities like movies, validated by synthetic populations — is a hypothesis that deserves more testing before it gets adopted as orthodoxy.
In the meantime, watch how people actually use the spaces you build. Talk to the brick-house family. Leave the cup simple enough that someone can put a paintbrush in it. The good cities, like the good products, are mostly the ones where the residents got to write the story.