A model that understands buildings

Hi Reader!

Last week the big news was that Demis Hassabis has stepped back from day-to-day responsibility for Gemini towards long term strategy and scientific work. Read it one way, and it's a step back from a race that Google is losing. I don't buy that. I believe that Demis Hassabis is getting off the CEO treadmill to chase an even bigger prize: a world model, one that understands, simulates and acts in the physical world.

What a world model actually is

We're all familiar with large language models (LLMs). Language models predict the next word. A world model holds a representation of how an environment behaves, predicts what happens next, and estimates the consequences of an action before anyone takes it. It is the requirement for robotics, for scientific discovery, and for anything that has to survive contact with physics. Hassabis didn't win a Nobel Prize for nothing. He talks constantly about using this intelligence to solve real problems, and solving real problems means modelling the physical world rather than describing it.

Look at what Google is already shipping. Nano Banana and Veo are winning on images and video, and you cannot generate convincing video without some internal grasp of how objects move, fall, occlude and persist. Genie 3 goes further and generates interactive environments a person or a machine can move through in real time. Gemini Robotics takes the same thinking into spatial reasoning and multi step action in real spaces. Four products, one direction of travel.

What else is Google up to?

You've also got Street View and Maps. There's Waymo, which has spent years simulating physical environments because the alternative was crashing real cars. They've got immense TPU capacity and Isomorphic Labs is proof that they can take this approach into a hard scientific domain and get results.

Google is not to be underestimated just because it isn't throwing resource at outcompeting the latest ChatGPT model.

What world models will mean for our sector

Buildings are physical, spatial and temporal systems. Documents describe them. They are not the thing itself.

Today AI helps you organise an analysis. A mature world model would simulate the consequences. Design teams could test how a building behaves under different occupancy, daylight and wind conditions before a line is drawn. Site teams could sequence logistics, crane movements and temporary works in simulation. Drone and scan data could be compared against the intended model automatically, with deviations flagged. Facilities managers could model refurbishment and evacuation options across a forty year life.

An LLM can help you structure that question today.
A world model would answer it, and show you what changes when you move one variable.

The caveat (there's always one!)

Genie can currently only hold a world together for a few minutes and a building has to hold together for forty years, through refurbishments, ownership changes and regulatory updates, with every assumption traceable back to a source. Materials have to behave the way materials actually behave, under load, over time, in northern European weather.

Here's what to watch out for: at some point a simulation will be accurate enough that a chartered engineer signs their name under its output and accepts the liability.

Once that happens. Everything is going to change.

I believe that moment is closer than most people in our industry think. Inside twelve months, is my guess.
Which brings me to the part that actually matters for your firm...

The window is open now, and it's not limited to world models

You cannot buy a world model today. You can buy, and should already be using, the AI that handles everything wrapped around the design work.

Bids and tender responses. Prequalification questionnaires you answer twenty times a year with the same content in a different order. Estimating and the assembly of cost data. Compliance checks against regulations and planning conditions. Monthly reports. Safety documentation. Meeting notes that turn into actions that turn into disputes when nobody wrote them down properly.

Repeatable, high volume, low creativity, and exactly where your margin leaks. The return on automating all of this is available now, and with tools that already exist. But you've got to show your people how, and actually by law you've got to provide AI training if your people are using AI in the workplace.

When world models land, they are going to need structured data and people who are fluent in working with AI.
Spend the next year putting AI to work on the surrounding processes and you will be well placed for when the world models land; you won't be adapting to something new.

This week on ChattingGPT: Endra's Niklas Lindgren

Niklas Lindgren is the founder and CEO of Endra, a Stockholm-based AI platform built for mechanical, electrical and plumbing engineers, which just raised a $50m Series A led by Andreessen Horowitz. While we don't normally have vendors on the podcast, we made an exception in this case because Endra looks like the Revit killer everyone's been waiting for after all these years.

In this episode, Maryrose and Niklas get into:

Why MEP engineering has barely changed since the 1990s, and what finally shifts it
Why jobs are going nowhere, even as 70 to 80% of project hours change
Why the outsourced, commoditised work is the most exposed

Niklas makes the case that this is a people game that takes real time, not a quick pilot, and shares the one practical step every engineering firm should take this week. Take a listen.

🎧 Apple Podcasts

🎧 Spotify

Soon we'll be adding Quboz. Here's why. 🍉

Article 4 of the EU AI Act: what it asks of you, and how to prove you've done it.

Join us for a free lunchtime webinar on Thurs 17 September.

Most firms we speak to fall into one of two camps. Either they haven't heard of Article 4, or they've heard of it and assume someone else is handling it.

Since February 2025, any organisation using AI has a legal obligation to ensure its people are AI literate. Not a recommendation. An obligation, sitting in Article 4 of the EU AI Act, with no grace period and no small company exemption.

Join us for an hour on what the obligation actually requires, what counts as evidence you've met it, and how to build AI literacy across a team without stopping the work.

We'll cover:

What Article 4 says, in plain English
Who it applies to (deployers as well as developers, which catches nearly everyone)
What regulators are likely to look for
What good AI literacy looks like in a working business
Where firms are getting this wrong

Thursday 17 September, 1pm to 1.30pm. Free to join. Bring your questions.

How would you like your Fridays back?

Our live courses are built on real built environment workflows: tenders, drawing reviews, programme tracking. No marketing examples or generic 'one size fits all'.

We are delighted to offer 9 very specific courses ranging from Nano Banana for Architects, to AI for Bid Teams, Copilot for Finance teams, and everything else in between.

One thing we have learned from almost 3 years of training professionals is that you're all special, and you like to learn with your own crew. So finance people get to learn with finance people. Architects together. That's what we offer, with our first courses starting on 15 September.