When AI starts to understand space, predict change, and act inside both virtual and real environments, the thing we're designing may no longer be a product. It may be a world that keeps running.
01What comes after the chat box?
I've spent the last two years deep in this. It started with using ChatGPT to help me write, then building actual products with Claude Code and Codex, then eventually running several AI agents that divide work between them. Somewhere in there, the way I make things changed completely.
Shipping a reasonably complex digital product used to mean juggling design, development, five tools that don't talk to each other, and a long gap between the idea and anything real. Now a lot of that work is just… close at hand.
But lately a different question has been nagging at me. If AI stops sitting in a chat window answering our questions, where does it go next?
Almost everything we use today talks to us through text, images or voice. Even agents, the ones that actually do things, mostly operate inside digital systems. They search, write code, move files, click around software.
The real world isn't a chat window. It has space and time, relationships between objects, and things that keep changing whether or not anyone asked. If AI is going to participate in our lives and our work in any serious way, it probably needs to understand more than what we said. It needs some grasp of what's happening around it, and what its own actions might cause.
Which is how I ended up reading about world models.
02From understanding language to understanding worlds
Here's the simplest way I've found to hold the three ideas apart. Ask the same question three times, to three different kinds of system:
A world model cares about one thing above all: if this action happens, what does the environment do next?
Think about a kid with wooden blocks. Recognizing a block is the easy part. What they're really learning, slowly and by knocking a lot of towers over, is how things relate: why some structures hold and some don't, what happens to everything else when you put one block here.
That intuition about environment, cause and consequence is a huge part of what we mean by intelligence, and it's the part language models were never really built for. Try it: add blocks and see where confidence goes:
Three blocks, centred. Nothing interesting is happening yet.
Today's world models are nowhere near understanding the real world. I want to be clear about that, because the demos are seductive. But the progress in the last year is genuinely strange to watch.
Reading all three back to back made something click for me:
And I've become much more interested in the second one.
The demos are doing a lot of work
I want to slow down here, because I've now caught myself getting excited by a good launch video three separate times, and I'd rather be useful than enthusiastic.
A benchmark published this year called World Models' Last Exam in Physics ran eight image-to-video models through 40 controlled tasks across nine categories: collisions, rolling and statics, pendulums, optics, hydrostatics, phase change, electromagnetism, granular flow, surface tension. 1,280 generated videos, scored against what the physics actually predicts rather than against a reference clip. The results are sobering:
The highest score any model reached. The scale is not a curve; it's measured against physics.
The individual failures are more revealing than the averages. Pick a task:
The finding I keep coming back to is the free-fall one. Every single free-fall video passed the temporal-consistency screen. Nothing flickered or teleported. It all looked right, and the average physics score was still 26.6. The objects fell convincingly along the wrong trajectory.
There's a sharper structural version of this argument in a paper bluntly titled Critiques of World Models. Its claim: a video generator isn't a world model at all, because it produces one fixed trajectory and can't answer "what if I'd done something else instead." No explicit state, no counterfactual control. By that standard, the 3D scene models are beautiful static worlds without physics or agents, and the interactive ones hold coherence for a minute or two. Which is roughly where I'd honestly place all of it today.
None of that makes the direction wrong. It means the gap between generates a convincing world and understands a world well enough to act in it is the actual work, and it's wider than the launch footage suggests. I'd rather know that going in.
03Looking back, why did I keep building 3D spaces?
I've always liked designs with a sense of place. Given the choice between a normal webpage and something that makes you feel like you're entering somewhere, I'll take the second one every time. A building. A corridor. A door that opens. A scene that shifts as you move through it.
None of that is decoration, or at least it doesn't have to be. Space gives you orientation. It carries story. It changes how you relate to the thing you're using.
When I built Yuya Film Studio, I didn't want a portfolio site with videos on it. I wanted people to walk into a film studio: through the entrance, into the hall, around the production areas, into the Previs Lab to look at shot planning, then into the shooting space.
Somewhere in the middle of building that with Three.js and WebGL, I noticed the part I enjoyed most had stopped being "make this look good." It had become a set of harder questions. Why would someone walk over here? What should this room make you feel? How does one interaction hand you naturally to the next? Can the space itself be doing the explaining?
Then I started building TeamBot HQ, and that one got more interesting, because it isn't a space you visit. It's an attempt to make AI agents' collaboration visible. Each agent has a role, a work area, a task state. Instead of reading a log to figure out what everyone's doing, you look at a studio and understand how the team is running.
Which led me somewhere I didn't expect: what if a 3D space stopped displaying information and started reflecting a system that's actually running?
An agent starts working and something in the room responds. A task moves from Research to Builder and you watch the work move instead of reading a status change. And then, further out: what if the system could simulate a few different ways to execute, given the tasks and constraints and resources it has, and let you pick?
That stops being a fun 3D dashboard. It starts being a way to see, understand and steer an intelligent system.
I should be honest here: none of what I've built is a world model. It's spatial interaction, live system visualization, AI collaboration. But doing it is exactly what got me interested in spatial intelligence in the first place.
04We may not just use AI. We may share an environment with it.
Most AI products still follow one familiar pattern: an input box, a conversation, a wait, a result. It works. It isn't going away. But it isn't right for everything.
Take managing a complicated project. "How's it going?" is rarely the real question. The real questions are: what's running in parallel right now, where is it stuck, what breaks if I change priorities, and what do the alternatives actually cost me? Try both ways of answering that:
So here's the thing I keep imagining. You enter a digital workspace. The agents aren't names in a sidebar. They're present, showing what they're working on, responding as tasks change. You can watch the system, and you can reach into it. And before anything expensive runs, the system plays out a few versions of what might happen so you can see the trade-offs for yourself.
For creative work, that's exploring shots and scenes and story structure before production. For a company, it's finally understanding a process that nobody can hold in their head. For robotics and physical AI, it's somewhere safe to train and fail.
Those need different technology, and most of them don't need a giant generative world model. But they point the same direction: AI is moving from processing information toward understanding environments, taking action, and working alongside people.
Which means designers may not only be asking how an interface gets used. We'll be asking how an intelligent environment gets understood, trusted, and controlled.
05This also changed how I think about my own career
Recently I've been paying attention to something called Forward Deployed Engineering. Roughly: engineers who go deep into real business environments. They build, but they also have to understand the customer's actual problem, connect AI capability to a real workflow, and own whether the thing creates value.
It's not a new invention. Palantir coined the role in the early 2010s, working with agencies that couldn't fully specify what they needed or hand over their data, so engineers went and sat with them. They called them Deltas, and until around 2016 the company had more of them than it had regular software engineers. The title is having a second life right now for an obvious reason:
A consultant hands over recommendations. An FDE builds the thing inside the customer's workflow and stays until it runs reliably, then feeds what they learned back to the product team.
MIT's NANDA study found that 95% of enterprise generative-AI pilots showed no measurable business impact. That's not a model problem. That's a deployment problem.
Half engineering, half reading the room. The interviews test both.
Both OpenAI and Anthropic stood up dedicated deployment arms this year.
That 95% number is the whole argument, really. We are not short of capability. We're short of people who can put capability somewhere it matters, which is a design problem as much as an engineering one.
That resonates more than I expected. Because I've become convinced the valuable work isn't building more AI features. It's knowing where to point them.
A model can be state of the art. A 3D world can be gorgeous. An agent system can demonstrate impressively complicated coordination. If none of it solves a human problem, it stays an impressive demo.
I'd rather be at the intersection than in the center of any one of these.
The intersection I mean: AI systems, spatial design, intelligent experiences, and actually deploying AI inside real businesses. I still love designing and making: spaces with story in them, interfaces with imagination, things people want to poke at. I just increasingly want what I build to leave my laptop and land in someone's real workflow.
Not telling people what AI can do. Letting them feel what it does for them.
World models point at a longer-range version of the same thing. As AI gets better at spatial understanding, environment simulation and action, somebody will need to design and deploy those environments: people who understand humans and products and systems. I have no idea what that job is called. Spatial AI engineer? AI systems architect? Something that doesn't exist yet? I just know it's the direction I want to walk in.
06A small prediction for the next three to five years
I don't think all software becomes 3D. I don't think every agent needs a body. Sometimes a button and a clear table is still the best design anyone's going to do.
But I think technologies that are separate today start merging. Agents plan and execute. World models simulate environments and possible change. Spatial computing offers more natural ways to interact. And real business systems connect all of it to work people actually do.
Three things I'm specifically watching. Tap each one:
None of this is a clean road. The real world is extremely complicated. Simulations will be wrong. AI will misjudge things. So the systems that matter won't just be more capable. They'll need better verification, clearer feedback, and a human holding the decision at the moments that count.
07What I actually want to make
Looking back at UX/UI work, interactive sites, Creator OS, Video OS, the automation workflows, the 3D film studio and TeamBot HQ, projects that look unrelated, there's one thread running through all of it. I like turning abstract ideas into things you can see, enter and use.
Once that was a design file. Then an interactive website. Now it's an AI system that collaborates and gets things done. Next, I hope, it's a digital environment with real intelligence, spatial sense and interaction in it, one that doesn't just display information but helps people understand a problem, explore their options, and act.
I don't know how far the technology gets in five years, or what form world models eventually take in ordinary life. But I'm enjoying the process of building and watching a new technology quietly change what products, spaces and interaction even mean.
I've also stopped trying to fit inside one job title. The boundaries between design, product, engineering, AI and spatial experience are moving. For me this isn't a designer becoming an engineer. It's taking the creativity, the user instinct and the spatial thinking I already have, and carrying them into a more intelligent world.
From designing interfaces, to designing systems, to exploring environments that can understand and respond to us. That's the direction I'm most excited about.
One last thought
We used to design how people use software. Now we're designing how people work with AI. Next, maybe, we design how people, agents and intelligent environments work together.
From agents to world models isn't only a technical upgrade to me. It feels more like a change in our relationship with computers. We stop handing a machine instructions and waiting for output, and start creating alongside systems that understand an environment, simulate change, and take part in the work.
I'd like to be more than a spectator for that.
The future isn't just about building smarter AI. It's about designing a better world for humans and intelligence to work together.