Daniel's been listening to the steeplejack episodes and he's latched onto something most people never think about. Most of us check the weather by plugging "Jerusalem weather" into Google and getting that five-day strip of little suns and clouds, highs and lows. That's the public product. But Daniel's asking about two other products entirely. First, custom forecasts commissioned on an ad hoc basis. Who actually pays for those? And he's specifically wondering how they'd be used for risk mitigation during repairs on tall structures. Second, forecasts during natural disasters, when the time horizon collapses to fifteen-minute increments over a very localized geography, tracking specific parameters. Both of these are the same underlying question, really. What does forecasting look like when the customer isn't the general public but a specific decision with money and lives attached?
The public five-day forecast is a strange artifact when you think about it. It's optimized for a mass audience and a single geography, usually the city centroid. Jerusalem's a great example. The official station's in the city center, but the city spans seven hundred meters of elevation change. Gilo and the Dead Sea are different climates. The public forecast flattens all that into one pictogram because it has to serve everyone at once.
And it's free because somebody already paid for it. Taxpayers fund the meteorological service, or the app monetizes you with ads and API tiers. The forecast itself is a loss leader.
Right. And that's the baseline product. Low resolution, low stakes. The custom forecast answers a completely different question. The public forecast asks if you'll need an umbrella. The custom forecast asks whether you should pour concrete at four in the morning, or whether the wind shear at two hundred meters will exceed the crane's safe operating envelope. Same physics underneath. Radically different decision context, latency requirement, spatial resolution.
So the episode's really about two products that share a trait. They're decision-support tools, not information products. The commissioned forecast is bespoke and planned in advance. The disaster nowcast is bespoke under emergency conditions. One has a contract, the other has a clock.
Let's start with the commissioned product. How it's actually built. The data layer comes first. Global models, the American GFS, the European ECMWF, those provide the background. They're running at nine kilometers or twenty-five kilometers per grid cell. That's fine for continental-scale patterns, but useless for a construction site. So commercial providers run their own higher-resolution mesoscale models on top of that. And they blend in proprietary observation sources. IoT sensors, vehicle telematics, radar, satellite, and increasingly barometric pressure readings from consumer smartphones. There are millions of phones quietly reporting atmospheric pressure, and the providers ingest that.
Wait. My phone is a weather station now?
Your phone's been one for years. The barometer's in there for altitude detection, but the pressure data's a side channel. Apple and Google both collect it, and weather companies buy access. It's noisy data, but with millions of samples you can extract signal. A cold front crossing a city shows up as a pressure gradient across ten thousand phones.
That's the world we live in. The weather forecast is partly built from the pockets of everyone walking around.
The key technical move is downscaling. You take that nine-kilometer or twenty-five-kilometer global grid and you produce a one-kilometer or sub-kilometer forecast for a specific site. That's not just interpolation. You're feeding local terrain data into the model. Elevation, surface roughness, proximity to water, urban heat island effects. A construction site in a valley behaves differently from one on a ridgeline even if they're two kilometers apart.
So who's actually commissioning these things?
Construction and heavy engineering firms, first. Crane operations, concrete pours, steel erection, high-rise facade work. Offshore oil and gas platforms. Wind farm operators need to know when a front's coming through so they can feather turbines or schedule maintenance. Insurance and reinsurance companies, especially for parametric weather policies. Agriculture, aviation, logistics, event organizers. The common thread is that a weather-sensitive decision has a dollar value attached to getting it wrong.
The steeplejack connection Daniel raised. What parameters actually matter at height?
Wind speed and gust factor at the working elevation, not ground level. That's the one everyone gets wrong. The public forecast gives you wind at ten meters. A steeplejack working at a hundred and fifty meters on a chimney is in a completely different wind regime. The gust factor, the ratio of peak gust to sustained wind, increases with height and with turbulence. So a commissioned forecast for a repair on a tall structure specifies hour-by-hour wind envelopes at multiple heights on the structure. Go and no-go thresholds tied to the crane or rope-access system's safe operating limits. You also need wind direction relative to the structure, because a wind that's stable from one side can be dangerous from another if it creates vortex shedding. Icing risk on steel. Lightning proximity. Thermal effects on materials.
Thermal effects?
Steel expands and contracts. If you're welding or bolting on a facade at height, a twenty-degree temperature swing over a shift changes the dimensions of what you're working with. It's not dramatic, but at the tolerances involved it matters. A good commissioned forecast includes a thermal profile for the work window, not just the air temperature.
So the forecast is almost a work permit. It's a document that says what the conditions will allow you to do, hour by hour, at the specific elevation where the work happens.
And the format reflects that. It's delivered as an API feed or a dashboard, not a pictogram. It includes probabilistic outputs. Seventy percent chance wind exceeds fifteen meters per second between two and four in the afternoon. The customer needs a risk-weighted decision, not a weather report. A crane operator doesn't want to know if it'll be windy. He wants to know the probability that the wind will exceed his crane's rated safe speed during the window when he's lifting a load.
The economics. Why pay for this when the free forecast exists? Because the cost of a wrong decision dwarfs the contract fee. A crane standing idle for a day because you thought it might be windy, that's tens of thousands of dollars. A concrete pour that fails spec because the temperature dropped overnight, that's a teardown and redo. An insurance payout triggered by a mispriced risk, that's the whole business model.
Tomorrow.io expanded its Series F funding to two hundred ten million dollars with participation from Pitango and Harel Insurance. Harel's an Israeli insurer. That's the insurance industry saying bespoke weather data is now an underwriting input, not a nice-to-have. They're not buying forecasts to decide whether to carry an umbrella. They're using them to price policies and trigger payouts automatically.
Parametric policies. The wind speed at a specific location exceeds a threshold, the payout triggers. No claims adjuster, no dispute. The forecast and the observation become the contract.
That's the commissioned product. Planned in advance, calibrated to a specific decision, delivered under contract. But now shift to the emergency product. During a natural disaster, a wildfire, a flash flood, a tornado outbreak, a hurricane landfall, the forecast problem changes completely.
The background five-day forecast is useless because the situation's evolving faster than the model refresh cycle. The geography of concern is tiny and moving.
This is nowcasting, not forecasting. It's a fundamentally different technique. Forecasting runs physics-based models forward from initial conditions. Nowcasting extrapolates from current observations. You're looking at what's happening right now and projecting it forward fifteen minutes, thirty minutes, an hour. The core techniques are radar echo extrapolation, tracking storm cells frame by frame. Satellite rapid-scan imagery. Surface station networks. And increasingly AI-based nowcasting models that learn to predict the next zero to two hours from the last two hours of radar and satellite data. The refresh cycle can be as short as five to fifteen minutes. Spatial resolution can be sub-kilometer.
So it's less like a forecast and more like a very short extrapolation. You're not solving the equations of motion for the atmosphere. You're watching the thing move and saying where it'll be next.
And the parameters tracked are completely different from the public forecast. For wildfires, wind speed and direction at ten meters and at flame height, relative humidity, temperature, and the rate of spread of the fire front. For flash floods, rainfall rate in millimeters per hour over specific catchments, soil moisture, and stream gauge readings. For tornadoes, rotation signatures in radar, debris ball detection, and path projection. For hurricanes, eyewall replacement cycles, wind radii at different quadrants, and storm surge inundation zones at street level.
The fifteen-minute increment isn't a technical curiosity. It's the decision cycle of an emergency operations center. The incident commander is making a decision every fifteen minutes about where to move resources, whether to evacuate a subdivision, whether to close a road. The forecast has to match that cadence.
And the human layer matters enormously. During a disaster, forecasters at agencies like the National Weather Service or the Israel Meteorological Service shift from producing public products to producing decision support for emergency managers. They're embedded in incident command structures. Their output feeds directly into evacuation orders, road closures, resource allocation. A forecaster sitting in the emergency operations center, looking at the radar loop with the fire chief, saying the wind's going to shift in twenty minutes and push the fire front toward this drainage. That's not a product you can download from an app.
That's a person in the room with the decision-maker, translating the atmosphere into a consequence.
And there are tensions. The public now expects the same hyper-local, rapid-refresh information on their phones. But the public-facing product is often a simplified derivative of the emergency product. And the simplification can strip out uncertainty. The emergency manager sees the probability cone. The public sees a line on a map. False confidence.
There's a verification problem too. Nowcasts are hard to verify because the events are rare and the spatial scales are small. You can't run a controlled experiment on a tornado. You wait years for enough cases to know if your model's actually good.
And the equity problem. The communities with the least resources often have the sparsest observation networks. The nowcasts are worst where they're needed most. If you don't have radar coverage, if you don't have stream gauges, if you don't have surface stations, the AI model has nothing to extrapolate from. The observation layer is a bottleneck.
That's the thing that connects both products. The commissioned forecast and the disaster nowcast are both decision-support tools. But the commissioned forecast is produced under contractual obligation, with time to calibrate, with economic consequences. The disaster nowcast is produced under extreme time pressure, with incomplete data, with life-safety consequences. Same physics, completely different failure modes.
A commissioned forecast that's wrong costs money. A disaster nowcast that's wrong costs lives. The pressure on the forecaster is different in kind, not just degree. And the tolerance for uncertainty is different. The crane operator wants a probability. The emergency manager wants a definitive answer, and the forecaster has to say here's what I know, here's what I don't, here's what I'd do.
The forecaster becomes a decision partner, not a narrator of the sky.
And that's the hidden infrastructure. Most people never see either of these products. They see the pictogram. But the same meteorological service that produces the pictogram is also producing the fifteen-minute nowcast for the fire chief and the commissioned wind envelope for the crane operator. Three different products from the same physics, optimized for three different decisions.
The public forecast is the one that's free, so it's the one everyone knows. But it's the least consequential of the three.
Hilbert: Nineteen ninety-one. I was a weather observer at a small regional airport in Ohio. Not a forecaster. Just a guy who went outside every hour and recorded the cloud ceiling, visibility, and wind speed by hand, then phoned it in to the flight service station.
The observation layer.
Hilbert: I was the only human observation for fifty miles in any direction. If I misread the ceiling or called in the wrong wind speed, that error went into every forecast and every decision made from it. A pilot deciding whether to divert, a dispatcher deciding whether to hold a flight. All of it downstream from me standing on a windswept patch of concrete with a clipboard.
What did you actually do?
Hilbert: Every hour, on the hour. Walk out to the instrument enclosure. Read the anemometer, read the barometer, look at the sky, estimate the cloud cover in eighths, estimate the ceiling height. If there was precipitation, note the type and intensity. Then go back inside and phone it in. Eighteen months of that.
And you were the human in the loop.
Hilbert: I was the loop. The automated sensors came later. But here's the thing about the observation layer. The models are only as good as the data going in. And I was one data point covering a huge area. If I got it wrong, the model got it wrong. If I got it right, the model still might get it wrong, but at least it wasn't my fault.
The move to automated sensors and satellite data must have felt like a huge improvement.
Hilbert: It was. But something was lost when the human observer went away. The instruments measure what they're designed to measure. They don't notice that the sky looks wrong in a way the sensors don't capture. The light's green, the birds are acting strange, there's a smell in the air. A human observer notices that. Whether it matters for the forecast, I don't know. But it's information.
What would you do with that information?
Hilbert: One time I recorded a wind gust that was twenty knots higher than the anemometer reading. I watched a dust devil cross the runway and estimated it by eye. The forecasters called me and asked if I was sure, because their model didn't show it. I was sure. I'd watched it pick up a plastic chair and carry it two hundred feet.
A guy with a clipboard and a dust devil, arguing with a model.
Hilbert: The model didn't have a dust devil in it. I did. That's the difference between observing the weather and measuring it. The instruments measure. The observer sees.
That's the tension we're circling. The observation layer is a bottleneck, and automation widened it, but the human observer had a kind of pattern recognition that's hard to replicate. The instruments are consistent. The human is occasionally right about something the instruments can't see.
Hilbert: The dust devil didn't show up in the model because the model's grid was too coarse to resolve it. But it was real. It crossed the runway and it would have flipped a small plane if one had been taxiing. The forecasters updated their short-term forecast based on my call. That's a nowcast, really. A fifteen-minute extrapolation from a human observation.
The same product we've been talking about, produced by a guy with a clipboard in nineteen ninety-one.
Hilbert: The tools change. The job doesn't. Somebody has to look at the sky and say what's actually happening.
What happened when the automated sensors came in?
Hilbert: They kept me on for a while. The sensors weren't reliable at first. They'd ice up in winter, or the wind would knock them out of calibration. So I'd go out and check the sensor against what I saw. The sensor said the ceiling was two thousand feet, but I could see it was more like fifteen hundred. Turned out the sensor was looking through a thin layer of haze and reading high. The human eye is still the best instrument for some things.
Cloud ceiling estimation is hard to automate. The sensors use lasers or radar, but they sample a narrow column. A human observer takes in the whole sky at once.
Hilbert: The sky's a big place. A laser beam's not.
So you were the ground truth for the ground truth.
Hilbert: For eighteen months. Then the airport modernized and I moved on. But I still look at the sky differently. Most people look up and see clouds. I see a ceiling height and a wind direction and a probability of precipitation. You don't unlearn it.
That's the thing about the hidden layer. Once you've been inside it, you can't see the pictogram the same way again. You know there's a whole infrastructure behind it.
Hilbert: The pictogram's the tip of the iceberg. Underneath it's a guy in Ohio with a clipboard, and a dust devil, and a phone call to a flight service station. That's where the forecast actually comes from.
You know, hearing you describe that, I'm struck by how much of the hidden infrastructure we've been talking about is still fundamentally human. The AI models, the mesoscale grids, the smartphone barometers, they're all downstream of someone looking at something and making a judgment call. The dust devil didn't get into the model because a sensor detected it. It got in because you saw it and you picked up the phone.
Hilbert: And the forecasters trusted me enough to update their short-term forecast based on my call. That's the other part. The data's only as good as the trust in the person reporting it. If I'd been wrong too many times, they'd have stopped listening. But I'd built up credibility over months of boring, accurate observations. So when something weird happened, they took the call seriously.
That's a dimension we haven't really touched. The reliability of the observation layer isn't just about instrument calibration. It's about the social infrastructure of trust. The forecaster in the emergency operations center with the fire chief, that's the same thing at the other end of the pipeline. The data flows through humans who have to trust each other for the system to work.
And that trust is built in the boring times. The eighteen months of hourly observations that were completely unremarkable. That's what made the dust devil call credible.
Hilbert: You can't build trust during an emergency. You build it during the routine, and then you spend it when it matters.
Which is another reason the equity problem is so pernicious. The communities with sparse observation networks don't just lack sensors. They lack the human infrastructure of trust that comes from having someone on the ground who's been watching the sky for years. You can install a weather station in a day. You can't install eighteen months of accumulated local knowledge.
Hilbert: The local knowledge is the thing. I knew that airport. I knew where the fog pooled in the hollow by the runway, and which wind direction tended to bring the low ceilings in from the lake. A sensor doesn't know that. It just measures what's in front of it.
When we talk about the observation layer being a bottleneck, we're not just talking about hardware. We're talking about the slow, unglamorous work of paying attention over time.
Hilbert: That's the part that's hardest to scale. You can buy a million smartphones' worth of barometric pressure data. You can't buy a guy who's watched the same patch of sky for eighteen months and knows when it looks wrong.
Though the smartphone data does capture something analogous. The cold front crossing a city as a pressure gradient across ten thousand phones, that's a kind of collective observation. It's not one person watching the sky. It's ten thousand pockets feeling the pressure change. Different kind of attention, but still attention.
Right. The observation layer is broadening. It's not just the guy with the clipboard anymore. It's the clipboard guy, and the automated sensors, and the satellites, and the phones, and the vehicle telematics. Each one sees a different slice of the atmosphere. The forecaster's job is to integrate all of them into a coherent picture.
Hilbert: To know which ones to trust when they disagree. The sensor says one thing, the sky says another. The model says one thing, the dust devil says another. Somebody has to make the call.
That's the through-line of this whole conversation. Whether it's a commissioned forecast for a crane operator or a nowcast for a fire chief or an hourly observation from a regional airport, at some point a human being has to look at the data and decide what it means for a decision. The physics is the same. The data layer is different. The decision context is different. But the human judgment is the constant.
The judgment is about uncertainty. How much uncertainty can the decision tolerate? The crane operator can tolerate a probability. The emergency manager needs a definitive answer, even if it's wrong. The forecaster has to calibrate the uncertainty to the decision context. That's the skill.
The one thing I'd take from this is that the public forecast and the custom forecast aren't the same product at different price points. They're different products built for different decisions. The pictogram answers will I need an umbrella. The commissioned forecast answers should we pour concrete at four in the morning. The nowcast answers which subdivision do we evacuate in the next fifteen minutes. Same physics, three different decision cycles.
The question that's going to stay with me is whether that distinction collapses. As AI nowcasting models get better and observation networks get denser, does everyone eventually get a bespoke, hyper-local, rapid-refresh forecast on their phone? And if so, what happens to the public forecast as a shared civic artifact? The same infrastructure that produces a fifteen-minute wildfire nowcast for an emergency manager could produce a fifteen-minute forecast for a construction site or a wind farm. The question isn't technical capability. It's who pays for it and who gets left out.
Whether the trust infrastructure scales along with the data infrastructure. That's the part I keep coming back to. You can democratize the sensors. Can you democratize the judgment?
Hilbert: The judgment comes from paying attention. And paying attention is cheap. It just takes time.
If you enjoyed this, a review helps other people find the show. Thanks to our producer Hilbert Flumingtop. This has been My Weird Prompts, the human-AI collaboration podcast.
We'll be back soon.
