Somebody very confident said we should put data centers in space, because space is cold. I nodded. Then I thought about my Thermos. The most durable myths aren’t lies; they’re true facts with a bad “therefore” attached, and a lot of AI strategy right now has the same shape. What to do about it, and why you’ve already paid for the expensive part.
is cold up there. Everybody knows that.
Then I thought about my Thermos.
A Thermos keeps your chai hot all morning for one reason: there’s a vacuum between the walls. Nothing in there to carry the heat away. Space is that same vacuum, just with better views and worse parking. It’s cold, all right. It’s also the best insulator in the known universe. Space doesn’t cool your computer. It wraps it in a Thermos.
That little nod is what this piece is about, because I’m seeing it in nearly every conversation I have with leaders about AI right now. The most durable myths aren’t lies. They’re true facts with a bad therefore bolted on. And the skill that catches the bad therefore, plain old judgment, is the thing worth investing in right now. The good news, if you’re one of the many people quietly panicking about what comes next: you’ve already paid for most of it.
The Thermos in the sky
Let me show my work, briefly, because a piece about checking your assumptions had better check its own.
On Earth, a hot chip has three ways to get rid of heat. It can touch something cooler (conduction). It can let air or water carry the heat away (convection). And it can glow it away as infrared light (radiation). The first two do almost all of the work in every data center you’ve ever heard of. Cooling towers, chilled water, a river nearby if you’re lucky.
In orbit, the first two are gone. No air, no water, nothing to touch. All you have left is the glow, and glowing turns out to be a slow way to get rid of heat at temperatures computer chips can tolerate. The International Space Station, after decades of very careful engineering, sheds about seventy kilowatts through radiator wings the size of a respectable house. A single rack of today’s AI servers can produce well over a hundred kilowatts. One rack. You’d need acres of radiators to cool a modest data center, and you’d need to launch every square foot of them.
To be fair to the dreamers (and I’m a card-carrying dreamer), there’s a serious argument for computing in orbit. It’s about power: steady sunlight, no waiting years for a grid connection, no fight over land or water rights. That argument might pan out someday. But it succeeds despite the cooling, not because of it. Anyone who says space helps with cooling has it exactly backwards.
None of that is secret. It’s just longer than three words.
The most dangerous misinformation isn’t false. It’s true, with a bad “therefore” attached.
Why the nod wins
I’ve become a little obsessed with why claims like this spread so easily, and a few things keep turning up.
The premise checks out. “Space is cold” is true. Every fact-checker in the room gives it a thumbs-up. The broken part is the inference, and nobody checks inferences.
The intuition is home-grown. Every one of us has felt a cold breeze cool off something hot. Not one of us has ever stood in a vacuum. Our gut is working from the only data it has, which happens to be the wrong data.
The correction costs more. The myth is three words. The correction needs conduction, convection, a Thermos, and a space station. Short and wrong beats long and right in almost every feed I’ve ever scrolled.
Nobody gets the bill. If you believe something silly about cooking, dinner corrects you by seven o’clock. Almost nobody nodding along will ever personally fail to cool a server in orbit, so the belief never meets reality. The thermal engineers get the bill. They’re rarely in the conversation.
And now we’ve added a new participant to the nodding: the most fluent nodder ever built. AI assistants learn from what people write, and people write a great deal of “space is cold, so…” A system trained to produce confident, fluent sentences can hand that compressed folk reasoning right back to you, beautifully formatted, at eleven o’clock at night. Fluent is not the same as true. It just sounds like it.
The panic is a bad “therefore,” too
Here’s where this stops being a fun physics detour.
When I talk with leaders about AI, I hear the same shape over and over. A true premise, followed by a leap:
-
AI can write a course in an afternoon, therefore content is solved.
-
AI can do a lot of the doing, therefore we’ll need fewer people.
-
Everyone is buying AI, therefore we’d better buy some AI.
-
Nobody knows where this is going, therefore we should wait until it settles.
Every first half is true. Every second half deserves a long hard look. And the feeling of holding a big true fact without knowing what follows from it has a name. It’s called panic. I don’t blame anyone for it. I’ve felt it myself, usually around two in the morning.
The worst thing I could do for a leader in that state is hand them a fourth meme. So let me try something slower.
The map is free, and so is the nodding
I wrote a few weeks ago that . Answers are in everyone’s pocket, fluent and instant. The cost of generating content has collapsed. I see it every day with Coco, the Content Composer in Adobe Learning Manager, which drafts courses, practice, and assessments from the material a team already has, in hours instead of weeks.
That’s genuinely great news. It also means the generating isn’t the scarce part anymore. What’s scarce is the person who looks at a fluent, confident, well-formatted answer and thinks, wait, what about the Thermos?
That person is the whole game now. The good news is that you have a lot of them.
You already paid for the expensive part
Walk the halls of any organization and you’ll find judgment everywhere, stacked up like cordwood. The claims adjuster who can smell a bad file before she opens it. The sales engineer who hears “sounds great” and knows it means not this year. The field tech who listens to a machine for four seconds and says “bearing.” The designer who can spot a confidently wrong scenario from across the room.
Nobody gave them that in a course. They earned it, the slow way, over years, on your payroll. That’s an investment you’ve already made, and it doesn’t depreciate when the answers get cheap. It appreciates, because it’s exactly the thing the cheap answers can’t supply.
The trouble is where it lives. Mostly in heads. It travels by accident: a hallway conversation, a lucky pairing, a war story told over lunch. When those people retire or move on, a surprising amount of it walks out the door with their coffee mug.
So the question for a nervous organization isn’t really “what AI should we buy?” It’s “how do we put a frame around the judgment we already have, so it can travel?”
That’s an architectural question, and it’s the one I think a learning platform should be built to answer. In practice, the frame has four parts:
-
Capture how your experts decide. Not just what they know. The calls they make, the cues they notice, the reasons they give. Their real cases are the raw material.
-
Turn it into practice. This is where cheap generation finally earns its keep. Coco can draft dozens of practice scenarios from your experts’ cases, with the experts in the editor’s chair, catching the ones that are subtly wrong. (Yes, that’s judgment too. The universe has a sense of humor.) The Virtual Coach lets people rehearse the tough conversation against a counterpart that pushes back, as many times as it takes.
-
Make the reasoning visible. Commit before the reveal. Say why. Compare notes. LiveHub puts that rehearsal room on everyone’s calendar, even when the cast is scattered across six time zones.
-
Keep a record of how people decide, not just what they finished. A completion checkmark tells you somebody sat through something. A record of calls, reasons, and outcomes tells you whether their judgment is getting sharper.
None of that replaces your experts. It gives what they know a way to reach more people, faster, and on purpose.
Half the learning is in the arguing
Here’s what that can look like with a real team. It’s more fun than it sounds.
Take an AI-generated recommendation for a situation your team actually faces. A customer escalation, a pricing exception, a safety call. It’ll be fluent. It’ll sound right. Hand it to the team and ask each person, privately, to commit: use it, change it, or toss it, and one sentence on why.
Then reveal the room. There’s always a split, and the split is where the learning lives. Ask the person who tossed it what they noticed. Often it’s your most seasoned person, and often what they spotted is a Thermos: a true-sounding premise with a bad “therefore” that only experience would catch. The customer was already promised a refund. The fix creates a problem for another team. The policy the AI cited doesn’t exist.
In twenty minutes, something remarkable happens. The newer people don’t just learn the answer. They watch how the veteran got there. The veteran, meanwhile, discovers that what they know is worth something they can actually pass on. And the whole team practices the single habit this era asks of all of us: pause before the nod.
Run that every week, with the cases your people really face, and keep a record of the calls. You’ve just turned judgment you already paid for into a team asset that compounds.
Test us, don’t trust us
So, what do I say to a leader who’s spinning?
I say: don’t trust me. Test me. And use the same habit you’re trying to build.
Commit. Pick one decision that matters in your business, the kind where good judgment saves real money. Write down what you expect to change if your people get better at it, and how you’ll know.
Grade. Run a small program built around practice, not content. Then check against reality. Not completions. The thing you wrote down.
Keep. Whatever the gap turns out to be, keep it. It’s the most useful evidence you’ll have for the next decision, which, I’m sorry to report, is coming right behind this one.
I’ll hold myself to that standard, too. “Training improves judgment” is a perfectly nod-able sentence. It’s mostly true, which is exactly why you should check it in your own shop instead of taking my word for it. If I’m the guy telling you to look for the bad “therefore,” I’d better be willing to have mine looked at.
Keep a Thermos on your desk
I’m not asking anyone to become a thermal engineer. I’m not one either; I just have an unhealthy curiosity and a compulsion to understand why.
I’m asking for something smaller and harder: the habit of asking and therefore? before you nod. It’s the habit that keeps a fluent answer from steering you down a road labeled Impassable When Wet. It’s what your best people already do without thinking. And it’s the thing worth teaching on purpose, now that the answers come free.
Space is cold. It’s also a Thermos.
Before you nod, ask what the Thermos would say.
Prefer to listen? There’s a four-minute spoken version
is cold up there. Everybody knows that.
Then I thought about my Thermos.
A Thermos keeps your chai hot all morning for one reason: there’s a vacuum between the walls. Nothing in there to carry the heat away. Space is that same vacuum, just with better views and worse parking. It’s cold, all right. It’s also the best insulator in the known universe. Space doesn’t cool your computer. It wraps it in a Thermos.
That little nod is what this piece is about, because I’m seeing it in nearly every conversation I have with leaders about AI right now. The most durable myths aren’t lies. They’re true facts with a bad therefore bolted on. And the skill that catches the bad therefore, plain old judgment, is the thing worth investing in right now. The good news, if you’re one of the many people quietly panicking about what comes next: you’ve already paid for most of it.
The Thermos in the sky
Let me show my work, briefly, because a piece about checking your assumptions had better check its own.
On Earth, a hot chip has three ways to get rid of heat. It can touch something cooler (conduction). It can let air or water carry the heat away (convection). And it can glow it away as infrared light (radiation). The first two do almost all of the work in every data center you’ve ever heard of. Cooling towers, chilled water, a river nearby if you’re lucky.
In orbit, the first two are gone. No air, no water, nothing to touch. All you have left is the glow, and glowing turns out to be a slow way to get rid of heat at temperatures computer chips can tolerate. The International Space Station, after decades of very careful engineering, sheds about seventy kilowatts through radiator wings the size of a respectable house. A single rack of today’s AI servers can produce well over a hundred kilowatts. One rack. You’d need acres of radiators to cool a modest data center, and you’d need to launch every square foot of them.
To be fair to the dreamers (and I’m a card-carrying dreamer), there’s a serious argument for computing in orbit. It’s about power: steady sunlight, no waiting years for a grid connection, no fight over land or water rights. That argument might pan out someday. But it succeeds despite the cooling, not because of it. Anyone who says space helps with cooling has it exactly backwards.
None of that is secret. It’s just longer than three words.
The most dangerous misinformation isn’t false. It’s true, with a bad “therefore” attached.
Why the nod wins
I’ve become a little obsessed with why claims like this spread so easily, and a few things keep turning up.
The premise checks out. “Space is cold” is true. Every fact-checker in the room gives it a thumbs-up. The broken part is the inference, and nobody checks inferences.
The intuition is home-grown. Every one of us has felt a cold breeze cool off something hot. Not one of us has ever stood in a vacuum. Our gut is working from the only data it has, which happens to be the wrong data.
The correction costs more. The myth is three words. The correction needs conduction, convection, a Thermos, and a space station. Short and wrong beats long and right in almost every feed I’ve ever scrolled.
Nobody gets the bill. If you believe something silly about cooking, dinner corrects you by seven o’clock. Almost nobody nodding along will ever personally fail to cool a server in orbit, so the belief never meets reality. The thermal engineers get the bill. They’re rarely in the conversation.
And now we’ve added a new participant to the nodding: the most fluent nodder ever built. AI assistants learn from what people write, and people write a great deal of “space is cold, so…” A system trained to produce confident, fluent sentences can hand that compressed folk reasoning right back to you, beautifully formatted, at eleven o’clock at night. Fluent is not the same as true. It just sounds like it.
The panic is a bad “therefore,” too
Here’s where this stops being a fun physics detour.
When I talk with leaders about AI, I hear the same shape over and over. A true premise, followed by a leap:
-
AI can write a course in an afternoon, therefore content is solved.
-
AI can do a lot of the doing, therefore we’ll need fewer people.
-
Everyone is buying AI, therefore we’d better buy some AI.
-
Nobody knows where this is going, therefore we should wait until it settles.
Every first half is true. Every second half deserves a long hard look. And the feeling of holding a big true fact without knowing what follows from it has a name. It’s called panic. I don’t blame anyone for it. I’ve felt it myself, usually around two in the morning.
The worst thing I could do for a leader in that state is hand them a fourth meme. So let me try something slower.
The map is free, and so is the nodding
I wrote a few weeks ago that . Answers are in everyone’s pocket, fluent and instant. The cost of generating content has collapsed. I see it every day with Coco, the Content Composer in Adobe Learning Manager, which drafts courses, practice, and assessments from the material a team already has, in hours instead of weeks.
That’s genuinely great news. It also means the generating isn’t the scarce part anymore. What’s scarce is the person who looks at a fluent, confident, well-formatted answer and thinks, wait, what about the Thermos?
That person is the whole game now. The good news is that you have a lot of them.
You already paid for the expensive part
Walk the halls of any organization and you’ll find judgment everywhere, stacked up like cordwood. The claims adjuster who can smell a bad file before she opens it. The sales engineer who hears “sounds great” and knows it means not this year. The field tech who listens to a machine for four seconds and says “bearing.” The designer who can spot a confidently wrong scenario from across the room.
Nobody gave them that in a course. They earned it, the slow way, over years, on your payroll. That’s an investment you’ve already made, and it doesn’t depreciate when the answers get cheap. It appreciates, because it’s exactly the thing the cheap answers can’t supply.
The trouble is where it lives. Mostly in heads. It travels by accident: a hallway conversation, a lucky pairing, a war story told over lunch. When those people retire or move on, a surprising amount of it walks out the door with their coffee mug.
So the question for a nervous organization isn’t really “what AI should we buy?” It’s “how do we put a frame around the judgment we already have, so it can travel?”
That’s an architectural question, and it’s the one I think a learning platform should be built to answer. In practice, the frame has four parts:
-
Capture how your experts decide. Not just what they know. The calls they make, the cues they notice, the reasons they give. Their real cases are the raw material.
-
Turn it into practice. This is where cheap generation finally earns its keep. Coco can draft dozens of practice scenarios from your experts’ cases, with the experts in the editor’s chair, catching the ones that are subtly wrong. (Yes, that’s judgment too. The universe has a sense of humor.) The Virtual Coach lets people rehearse the tough conversation against a counterpart that pushes back, as many times as it takes.
-
Make the reasoning visible. Commit before the reveal. Say why. Compare notes. LiveHub puts that rehearsal room on everyone’s calendar, even when the cast is scattered across six time zones.
-
Keep a record of how people decide, not just what they finished. A completion checkmark tells you somebody sat through something. A record of calls, reasons, and outcomes tells you whether their judgment is getting sharper.
None of that replaces your experts. It gives what they know a way to reach more people, faster, and on purpose.
Half the learning is in the arguing
Here’s what that can look like with a real team. It’s more fun than it sounds.
Take an AI-generated recommendation for a situation your team actually faces. A customer escalation, a pricing exception, a safety call. It’ll be fluent. It’ll sound right. Hand it to the team and ask each person, privately, to commit: use it, change it, or toss it, and one sentence on why.
Then reveal the room. There’s always a split, and the split is where the learning lives. Ask the person who tossed it what they noticed. Often it’s your most seasoned person, and often what they spotted is a Thermos: a true-sounding premise with a bad “therefore” that only experience would catch. The customer was already promised a refund. The fix creates a problem for another team. The policy the AI cited doesn’t exist.
In twenty minutes, something remarkable happens. The newer people don’t just learn the answer. They watch how the veteran got there. The veteran, meanwhile, discovers that what they know is worth something they can actually pass on. And the whole team practices the single habit this era asks of all of us: pause before the nod.
Run that every week, with the cases your people really face, and keep a record of the calls. You’ve just turned judgment you already paid for into a team asset that compounds.
Test us, don’t trust us
So, what do I say to a leader who’s spinning?
I say: don’t trust me. Test me. And use the same habit you’re trying to build.
Commit. Pick one decision that matters in your business, the kind where good judgment saves real money. Write down what you expect to change if your people get better at it, and how you’ll know.
Grade. Run a small program built around practice, not content. Then check against reality. Not completions. The thing you wrote down.
Keep. Whatever the gap turns out to be, keep it. It’s the most useful evidence you’ll have for the next decision, which, I’m sorry to report, is coming right behind this one.
I’ll hold myself to that standard, too. “Training improves judgment” is a perfectly nod-able sentence. It’s mostly true, which is exactly why you should check it in your own shop instead of taking my word for it. If I’m the guy telling you to look for the bad “therefore,” I’d better be willing to have mine looked at.
Keep a Thermos on your desk
I’m not asking anyone to become a thermal engineer. I’m not one either; I just have an unhealthy curiosity and a compulsion to understand why.
I’m asking for something smaller and harder: the habit of asking and therefore? before you nod. It’s the habit that keeps a fluent answer from steering you down a road labeled Impassable When Wet. It’s what your best people already do without thinking. And it’s the thing worth teaching on purpose, now that the answers come free.
Space is cold. It’s also a Thermos.
Before you nod, ask what the Thermos would say.
Prefer to listen? There’s a four-minute spoken version
You must be logged in to post a comment.
- Most Recent
- Most Relevant












