What I Learned at NeurIPS 2025, Why Buildings Are Still Black Boxes, and How Semantic Architecture Finally Lets Them Speak

NeurIPS 2025: The Scale of AI Today

NeurIPS, the Conference on Neural Information Processing Systems, is the world’s largest and most influential AI research conference.

Earlier this year at an ASHRAE conference, I met Judah Goldfeder (Google / Columbia University). After discussing the challenges of making buildings machine-readable, he invited me to present this work at the UrbanAI Workshop at NeurIPS 2025.

NeurIPS 2025 brought together more than 22,000 attendees and over 5,800 papers and posters, the largest gathering of AI researchers anywhere in the world. The exhibit floor was packed with an impressive recruiting and industry presence from Google, Apple, Meta, NVIDIA, Tesla, Netflix, Microsoft, Amazon, Wells Fargo, the Abu Dhabi Investment Authority, and many other global technology leaders.

What struck me most was the sheer scale and speed of activity: thousands of researchers in every direction, entire walls filled with 800 posters at once, and massive global participation, especially from China,  making it clear that AI is now a planetary effort with extraordinary momentum.

But one sector was almost absent: the built environment.

This gap set the stage for my UrbanAI presentation:

What does it take for buildings and cities to join the AI ecosystem instead of being left behind?

Jack Dangermond: GIS, AI, and the Nervous System of Cities

Following my invitation, Jack Dangermond, founder of Esri, recorded this message for the UrbanAI Workshop at NeurIPS. Jack describes how GIS has become a global nervous system for understanding cities: how we map them, analyze them, predict them, and increasingly, how we connect them to AI.

Jack’s vision sets the stage for the deeper dive that follows: How do we make the inside of buildings as intelligent, accessible, and computable as the maps that surround them?

Buildings Are Robots Locked Behind Walls

Buildings behave like cyber-physical systems, sensing, reacting, consuming energy,  yet unlike robots, they hide their internal state. Their intelligence is trapped behind vendor systems, proprietary data, and fragmented workflows.

Urban AI cannot function if buildings refuse to reveal themselves.

Digital Firewalls / The Isolation Problem

I use the term “firewalls” in multiple senses here, intentionally. In technology, firewalls block access to information. In architecture, physical firewalls protect occupants by containing hazards.

As an architect, I’ve spent decades watching how the AEC industry resists delivering intelligence alongside the physical building. We hand over drawings, PDFs, and static models. Still, the AEC industry rarely provides the meaning of the building, the identity of its systems, the relationships between components, or the data structures that make it intelligible to machines.

The surprise, and the point of this talk, is that building data is trapped behind all of these walls.
The physical walls that define a building isolate its internal systems. The digital walls created by proprietary platforms isolate access to data. And the semantic walls, missing identity, missing structure, missing meaning, isolate understanding itself.

Until we break through all three layers of firewalls, physical, digital, and semantic, Urban AI cannot see or understand the built world. Buildings will remain silent systems in a world that expects them to speak.

Why the Built Environment Is Still “Dumb”

The myth of “smart buildings” hides a simple truth:
Most buildings have systems but no shared semantics, sensors but no identity, models but no interoperability.

AI sees buildings as black boxes.

Real-World Failed Firewalls 

When buildings cannot express their materials, systems, pathways, and risks, emergencies worsen.
Urban AI cannot compensate for missing semantics. We need ground truth first.

Hidden Scale of Signals

Modern buildings generate tens of thousands of daily events, temperatures, flows, occupancy, faults.
AI researchers assume this data exists and is usable.
Operators rarely see more than a fraction.

35 Years of Semantic Architecture

My career evolved from architecture to semantic architecture, designing the meaning structures that make the built world intelligible to machines.

Architects design walls; semantic architects remove invisible ones.

The Object Genome Project

Years before AI was mainstream, we began assigning identity and attributes to every building component, a “genetic code” for buildings.

We weren’t preparing buildings for BIM.

We were preparing them for AI.

When thousands of buildings share a semantic backbone, the portfolio becomes computable.
Campuses become semantic networks, not isolated projects.

PAE Living Building

PAE shows what happens when you break through the semantic wall.
We decomposed every system and signal into a living knowledge graph.

Now, electrical, plumbing, HVAC, structural, and programmatic share the same meaning system. To become an AI-ready building.

Real-Time Signals and Causal Understanding

When semantics and signals are aligned, the building becomes self-explanatory. A battery drop on a panel can propagate through the semantic graph to illuminate downstream risk, not through algorithms, but through meaning. AI becomes more accurate because the building is no longer guessing about itself.

Talking to the Building

With a semantic foundation, you can finally ask the building questions in natural language:
“What failed?” “Where is it?” “What will happen next?” And it answers without hallucination, because its knowledge is grounded in geometry, identity, and system relationships.

This is Urban AI without the guesswork.

C4SB + Linux Foundation

No single vendor or discipline can build Urban AI. We need open semantics, shared identity, and standardized access to the built world. The Coalition for Smarter Buildings (C4SB) and the Linux Foundation are establishing the open architecture required for cities that can finally speak to humans, to machines, and to AI.

SESSION RECORDING

Watch the NeurIPS 2025 presentation from the UrbanAI Workshop.

Japanese Translation Coming Up…