A thermostat sounds like the simplest device in the world. You've got a sensor, a set point, and a switch. That's the whole thing, right?
That's the pitch. The reality is a lot stranger.
Daniel's got a whole thing about this. He wrote in about the episode we did on thermostat algorithms, and he wants to push it somewhere new. His point is that the bleeping battle you hear between an infrared blaster and an air conditioner, that constant back-and-forth, is really a lesson in how basic algorithms work. The failure mode, to borrow a phrase, is an overzealous algorithm chasing the set point, adjusting fan speed, set point, whatever it can reach, trying to hold the temperature in a stable band. Part of why that fails is thermal mass. The air in the room doesn't drop the instant you apply cooling.
Right. The room is a reservoir.
So he wants to move from that technical observation to something bigger. How do we build buildings for greater efficiency? He points out that well-intentioned sustainability mandates produce strict thermostat rules that would work nicely if the world weren't getting hotter, and the outcome is overheated airports that proclaim a sustainability award while the passengers inside are miserable. He brings up the ancient world, Stonehenge, and the creative HVAC solutions here in Israel. And he's asking at two layers. The architectural layer, where high rises go up with a wall of glass facing south and nobody puts solar blocking or temperature-blocking coatings on the windows because it's expensive and cost avoidance is the operative concern in high-rise development. And then the older buildings, where people designed with humility around the climate and a good measure of common sense. His actual question: can ancient building practices, in hot, cold, or variable climates, teach us to run modern cooling more efficiently without making life miserable for the people inside? And he wants the direct emissions angle too, not just the electricity.
That's a lot of layers. Let's take them one at a time.
Let's start with the algorithm, because the failure pattern is more interesting than the fix.
A naive thermostat is what control theory calls a bang-bang controller with hysteresis. Cooling comes on when the temperature rises above the set point plus a deadband, and it goes off when the temperature drops below the set point minus that deadband. The deadband is the whole trick. It exists precisely to stop the short-cycling, the bleeping battle Daniel's describing. Without it, the compressor toggles on and off every few seconds, which wears out the equipment and wastes a shocking amount of energy on startup surges.
So the deadband is the thing keeping the peace.
It's the thing keeping the compressor alive. Now here's where it gets interesting. The room's temperature responds to cooling with a delay and a lag. A controller that doesn't model that delay keeps applying cooling because the sensor hasn't caught up yet. Then it overshoots. Then it over-corrects the other way. In PID control that's called integral windup, and it's the classic failure of a controller that thinks the world responds instantly when it doesn't.
The algorithm is confident and wrong.
The algorithm has a model of the world, and the model is missing a variable. There's a paper from July, ThermoForce, and the finding is almost funny. They built a grey-box thermal model, the kind that's supposed to be accurate, and it had the best passive accuracy of anything they tested. Best at forecasting temperature. And it predicted the effect of a cooling action with the wrong sign.
Wrong sign. You cool the room and the model says it gets warmer.
Not quite that dramatic in practice, but functionally yes. A model that looks accurate at forecasting can be actively wrong about what a control action will do. The fix was to build a physics-structured model that's monotone in the control input by construction. More cooling can never produce more heat. That sounds trivial until you realize the accurate model didn't guarantee it. They embedded it in model predictive control on the BOPTEST benchmark and cut thermal discomfort by thirty-three to eighty-four percent versus the native controller, while also reducing energy.
So the model's shape mattered more than its accuracy.
The model's understanding of the world mattered more than its score on a forecasting test. That's the whole episode in one finding.
What about the district-scale work?
BuildSys this year, twenty-five buildings in one district. Centralized model predictive control tracked the load well, about eight point eight percent bias, but it concentrated all its actuation on a few buildings, and that caused twenty-four point eight percent comfort violations. Decentralized reinforcement learning spread the effort evenly but tracked poorly. The hybrid won. The lesson is that who the algorithm optimizes for changes the comfort outcome. You can hit your energy target perfectly and make a quarter of the buildings miserable.
And the language model experiment?
Multi-zone variable air volume system, a frontier model got a six point two percent electricity reduction, and it did it by reducing the ventilation margin. Efficiency gains quietly trading against air quality. Nobody put that on the sustainability award.
So now we get to thermal mass, which is the thing the algorithm keeps forgetting.
The room's air is a small thermal reservoir. The walls, the floor, the furniture, those are a much larger one. When you start cooling, you're removing heat from the air first, but the mass re-radiates stored heat back into the air. The air temperature doesn't drop until the mass itself has been cooled. That produces the time lag, and it produces what's called the decrement factor, the damping of the temperature swing. A controller that ignores mass is always behind the room. It's chasing something that already moved.
And the ancient builders didn't have an algorithm. They had mass.
They had mass, orientation, and ventilation, encoded in the structure itself. The building was the model.
Which is where the phase-change materials come in, because that's the modern engineering version of the thick wall.
You embed a material in the wall or the roof that absorbs heat by changing phase instead of by raising its temperature. Spherical phase-change modules in a concrete roof cut indoor surface temperature by ten point two degrees Celsius on average and reduced cooling load by up to sixty-nine percent in hot climates. That's a simulation study, so treat the number as an upper bound, but the direction is real.
Sixty-nine percent is not a rounding error.
There's a newer one, phase-change material infused into brick with metal foam to help it conduct. Under forty-five degree peak outdoor conditions, it cut peak indoor surface temperature by more than three degrees, peak heat gain through the wall by more than sixty-six percent, and required cooling energy by nearly half. And the internal rate of return on that frequently exceeds sixty percent. The economics work.
So the payback is there and we still build the glass box.
We build the glass box because the glass box is cheap to build and the cost lands on someone else. Which is Daniel's point exactly.
Let's go to the review, because that's the best single document on this.
Manshour and Lehmann, a systematic review, thirty studies from nineteen eighty to twenty twenty-five, spanning the Middle East and North Africa out to South Asia. Strong consensus around three passive principles. Solar control, natural ventilation, and thermal mass. The vernacular solutions that still work: courtyards, wind towers, the badgir and the malqaf, and thick masonry walls. The modern innovations that extend the toolkit: cool roofs, phase-change materials, parametric optimization. And the conclusion is that context-sensitive hybrid solutions, traditional knowledge plus modern technology, hold the greatest potential. The barriers they list are climate variability, cultural shifts, regulations, and economics. Not technical feasibility.
The technology works. The incentives don't.
That's the sentence.
Give me the courtyard evidence, because that one's been tested with actual people in it.
Seville social housing. The courtyards worked, and I'll use their word, faultlessly, as a highly effective passive cooling system. Up to twenty point five percent improvement in energy efficiency, and it was confirmed by occupant surveys. Not modeled. Surveyed. People said it was better.
And the Mediterranean study, the one about orientation.
Low-pitched roofs and a top chimney improved summer thermal comfort by twelve point six and five percent respectively. Thirteen and six point eight in winter. And here's the part I love. Pergolas on the north façade barely helped at all. Orientation matters. You can do the right thing on the wrong wall and get nothing.
Now the cost-avoidance objection. Daniel says solar blocking is expensive and it's a cost that can be avoided. What's the counter?
A static rooftop shading system. Install and forget. It blocks summer sun and admits winter sun, and it delivered up to thirty point three eight percent annual HVAC reduction, over five megawatt hours, and a seventy-one point three percent relative increase in thermal comfort. No moving parts. No maintenance. That's the counter to expensive. That's a fixed piece of geometry.
Thirty percent for something with no motors in it.
And there's a naturally ventilated office building in France with six summers of data. Solar control, natural ventilation, thermal inertia, a green atrium with automated openings, pedestal fans. No air conditioning at all. And they have occupant comfort surveys to go with it. Six summers.
So the building does the work.
The building does the work, and the equipment shrinks to a backup.
Now the mandate problem, because that's where Daniel's airport observation lives.
Royal Berkshire Hospital. Most rooms were already breaching overheating thresholds. They put in local air conditioning to cut the afternoon peaks, and the authors note it was in contradiction to the National Health Service's sustainability plans. The rule and the requirement pull in opposite directions. That's the mechanism. And I should flag, Daniel's airport example is illustrative. I couldn't find a specific named airport case in the research, so treat it as the pattern rather than one documented building.
The pattern is well documented though.
The pattern is the pattern. And it gets worse as the climate warms. Osaka office building, under twenty-nineties projections. Unconditioned zones exceed thirty-six degrees Celsius in summer. Cooling demand rises twenty percent, from three thousand three hundred and twenty-three kilowatts to three thousand nine hundred and eighty-three. Passive strategies, cross-ventilation, dynamic shading, high-performance insulation, cut cooling loads by fifteen to twenty-five percent. A thermostat rule that worked in nineteen ninety fails in twenty ninety.
And this isn't just comfort. You've got the medical training, so tell me why this is a health variable.
Indoor overheating and older adults. A two-hour cooling-centre intervention reduced cumulative thermal strain, but the symptoms and the mood disturbance persisted after they went back into the heat. The damage doesn't reset when you leave the room. There's a study on gut barrier damage, and at thirty-six degrees indoor temperature the markers of enterocyte damage rose sharply. The authors support the twenty-six degree indoor temperature recommendation for heat-vulnerable people. And housing factors predict heat illness. Housing age, crowding, and roof condition predict heat-related emergency visits and mortality. Building quality is literally a health variable.
The roof condition is a health variable.
That's the sentence I'd put on the poster.
Let's do the direct emissions, because Daniel specifically asked for it and it's the part people skip.
Refrigerants are potent greenhouse gases. HFCs like R-134a and R-125 have global warming potentials thousands of times that of carbon dioxide. The Kigali Amendment to the Montreal Protocol, twenty sixteen, phases them down, and it's expected to avoid nearly half a degree Celsius of warming by the end of the century. Half a degree from a refrigerant treaty.
And the replacement story is messier than the marketing.
The HFOs, the low-GWP substitutes, degrade atmospherically into trifluoroacetic acid, TFA, which is a persistent pollutant. And they can form HFC-23, which is a potent greenhouse gas, as a breakdown product. So their effective global warming potential may exceed the regulatory thresholds they were designed to clear. That's not a reason to keep the old ones, but it's not the clean story either.
Leakage.
Leakage matters as much as energy. Direct emissions from leakage, incorrect charging, and disposal are a major share of HVAC climate impact. Which is why life cycle climate performance is the right metric, not just the efficiency rating on the box. A unit that's efficient and leaks is worse than a unit that's less efficient and sealed.
And the zero-refrigerant future?
Elastocaloric cooling. Two hundred and sixty watts of cooling power, a twenty-two point five kelvin span. Ionocaloric refrigeration, about thirty percent of Carnot efficiency, a twenty-five degree lift. And passive radiative cooling coatings on the building envelope can cut annual cooling energy by more than twenty percent. The architectural point is that better buildings shrink the equipment, and shrinking the equipment shrinks the refrigerant problem. You don't have to solve the refrigerant chemistry if you don't need the refrigerant.
Which brings the whole thing back to mass.
The algorithm's model of the world is missing it. The building's physical model of the world is made of it. Same variable, two ends of the problem.
Hilbert: The deadband is usually not the algorithm's fault. It's the installer's.
Say that again.
Hilbert: I spent a summer installing and servicing thermostats and zoning dampers for a small HVAC contractor. That's where I learned what short-cycling sounds like. Compressor clicks on, clicks off, every ninety seconds. The homeowner had it set to sixty-eight and the room's mass simply would not cooperate. But the algorithm wasn't being overzealous. The installer had set the deadband to half a degree Fahrenheit, because the homeowner called and complained the temperature drifted too far from the set point. Half a degree. That's what kills the compressor. The algorithm was doing exactly what it was told. The human tuned it wrong.
So the failure is in the human's model of the algorithm, not the algorithm's model of the room.
Hilbert: The installer is optimizing for the phone call. He narrows the deadband, the complaint stops, the compressor dies two years later, and nobody connects the two. The houses that stayed comfortable in summer without the air conditioning running constantly were the old ones. Thick brick walls, deep porches. The new builds with the big south-facing windows were the ones I got called back to.
The installer narrowing the deadband to stop a complaint is doing exactly what the sustainability mandate does at building scale. Optimizing for the visible complaint and ignoring the physics.
Hilbert: Same shape. Smaller stakes. I have to collect something from a place that's about to close, and I'm not sure I'm going to make it.
The deadband detail is the whole episode in miniature. The model that matters isn't the one in the controller. It's the one in the head of the person who set it up.
The same is true at the scale of a building. The architect, the developer, the code writer, each of them has a model of how the building will behave, and the ones that ignore mass produce buildings that hit the target on paper and fail the occupant in practice.
Here's the thing I keep circling. Every serious review lands on the same conclusion. Hybrid. Traditional passive wisdom plus modern materials plus adaptive codes. And yet the incentive structure of high-rise development still rewards cost avoidance over thermal performance. If the consensus is that clear, why doesn't the money follow it?
Because the cost of a bad building is paid by the occupant, and the occupant isn't in the room when the budget is set. As the climate warms, the gap between buildings that buffer temperature swings and buildings that fight them in real time stops being an energy gap and becomes a health gap. The roof condition becomes a health variable.
That's the one to sit with. Thanks to Hilbert Flumingtop for producing. This has been My Weird Prompts, the human-AI collaboration podcast. If you want to send us something, email us at show at my weird prompts dot com.
We'll be back soon.
