The Last Scarce Thing
On Infinite Regress, Terminal Value, and Lossy Chains
I have written three essays about building. That you do not need to be an expert to build. That capability comes from the building, not before it. That the building itself is the strategy, and the people who cannot stop are the ones who move the world. I believe all of it.
But there is a question those essays did not answer, and it has started to feel like the only one that matters. If you can build anything, what should you build?
For most of history this question was academic, because you could not build anything. Capability was the constraint. The thing you could imagine was always far ahead of the thing you could make, and so the scarce resource, the thing worth organizing your life around, was the ability to build at all. That is what the three essays were about. Acquiring it. Refusing to wait for permission to have it.
That constraint is dissolving. Not gone, but dissolving faster than most people have adjusted to. When the cost of building collapses, building stops being the bottleneck. And whatever stops being the bottleneck stops being the scarce thing.
The scarce thing is now knowing what is worth building.
I have helped build things that, looking back, were not worth building, and I did not always know it at the time. The building was rarely the problem. The people were good at building. The trouble lived upstream, in the choosing, which is a harder and lonelier act than it gets credit for. A thing gets picked, ratified, funded, handed down, and somewhere in that chain the question of whether it was worth solving never quite gets asked out loud. The building is clean. The choosing is not. And no amount of clean building rescues a problem that was not worth solving.
There is a second failure, quieter than the first. Sometimes the right thing does get chosen, and then no one asks what it is actually for. It gets built, shipped, and its value assumed rather than traced. Choosing what to build and knowing why it matters turn out to be the same act, and it is easy to do only the first half.
The Climb With No Top
There is a question I have learned to keep asking, of others and of myself, and it is not a popular one in a meeting. Someone proposes a thing to build, explains its value, and the honest next question is: so what? You answer. And the question comes again: so what?
It can feel like obstruction. But I have not found a better way to tell whether a thing has value or only appears to, and I include my own proposals in that. Every justification invites the same question of the thing above it, and you keep climbing. We can automate this. So what? So we need fewer people to do it. So what? So it costs less. So what? So we can do other things. So what? And here, usually, the answers get vague, or someone says it is what someone above them asked for, and the climbing stops.
This is the infinite regress problem, and it is old. Every reason rests on another reason. Every “this matters” can be met with “why,” and the answer to the why can be met with another why, forever. The chain does not end on its own. It only ends when someone decides to stop climbing. The whole question of what is worth building comes down to where, and why, that climbing stops.
Say-So, or the World
There are two ways the climb can stop, and the difference between them is everything.
The first way is that someone with authority declares it stops. A leader decides the metric matters. It gets written into a plan. The regress halts because the org chart says it halts, not because anything outside the organization confirmed it. Call this the terminal outcome problem: the stopping point feels final, but its finality is borrowed entirely from the person who declared it. Replace the person and the terminal state can evaporate. The thing that was obviously worth building last year is quietly defunded this year, and nothing in the world changed. Only the internal agreement did.
A great deal of good work gets built this way, mine included. There is nothing dishonest about it. Conviction has to start somewhere, and often a leader’s judgment is the best signal available. But it is fragile in a specific way: its value cannot survive the question so what reaching all the way to the top, because at the top there is only someone’s say-so.
The second way the climb stops is that it reaches something outside the organization that does not answer to anyone inside it. A customer who stays or leaves. A market that prices the thing up or down. A risk that turns out to be real or not. Someone who pays, or does not. These validators do not negotiate. You cannot redefine your way into a customer staying. You cannot ratify a market into agreement. They sit outside the system that produced the work, which is exactly what makes them trustworthy: they have no stake in your story about why the work matters.
When the climb ends here, it ends because the world confirmed it, not because a person did. That is what a real terminal outcome looks like. The regress stops at revealed preference, and it stays stopped even after the leadership changes.
So the first test of anything worth building is simple, though it is uncomfortable to apply to your own ideas. Does the chain of so what end at something external and impartial, or does it end at someone’s say-so? If the latter, you do not have a terminal outcome yet. You have a conviction that will hold exactly as long as the person behind it does.
Too Far to Carry
But passing that test is not enough, and this is the part I missed for the longest time. You can have a real external validator at the end of the chain and still be building something whose value you will never be able to claim. Because the chain can be long.
Watch one stretch out. We reduce cost, so we free up capacity, so we do more work, so quality improves, so more people use the product. The last link, people using the product, is externally validated. Users decide that; the organization does not get a vote. The terminal outcome is real.
Now count the steps, and notice what lives in each gap. Cost reduction has to actually fund something, not just disappear into the next budget. The freed capacity has to go toward work that matters, not more meetings. More work has to improve quality, but volume is not quality, and it often is not. Better quality has to change whether people adopt, but mature products are held by lock-in, habit, and switching costs that dwarf any marginal improvement. By the time the signal from “reduce cost” reaches “more users,” it has passed through five noisy joints, and there is almost nothing left of it. Any one of a hundred other things could explain the outcome. Your work cannot credibly claim it.
This is the lossy chain problem, and it is in some ways more dangerous than having no validator at all. When there is no external validator, the absence is at least visible; everyone can see the work rests on someone’s say-so. But when there is a real validator at the far end of a long chain, it provides cover. It becomes easy, and tempting, to point at the distant, legitimate outcome and let people assume the work caused it. The claim cannot really be falsified, because the chain is too long to trace. The work feels grounded. It may not be. It can be optimism wearing the shape of an outcome. And I have made exactly that claim about my own work and believed it.
So the bar is two-edged: the chain of so what has to reach something external, and it has to be short enough that the signal survives the trip. Most work will not clear it, and that is fine. Not everything needs to. Some of the most important work, the enabling and infrastructural kind, has no short path to an external signal, and pretending otherwise just produces dishonest metrics.
There is a discipline that follows from this, and it is not “find an external validator.” It is find the nearest one.
Cost reduction was never the real outcome anyway. It is always in service of whatever the savings get spent on. The value lives in the reallocation, not the reduction. Which means the honest thing to measure is what the recovered capacity actually produced, measured right where it produced it, not five steps downstream where the signal is gone. If the savings funded specific work, and that work converts into something external in two or three tight steps, the chain holds and the claim is real. If the only validator available sits at the far end of a long chain, you are not measuring an outcome. You are measuring your optimism about the outcome.
None of this is about how you build. The spec, alignment, and feedback are downstream: you build to discover these things, not the other way around. Whether the thing was worth building is not. That question stays upstream, before the first line.
The discipline is not to reject the work that has no clean signal. It is to know which kind you are doing, and to stop pointing at distant validators as if they were close.
The three essays on building argued that you should build, relentlessly, before you are ready, without waiting for permission. I still believe that. But relentless building has a failure mode the essays did not name: building the wrong thing faster than anyone else. The maniac who picks a problem with no real terminal value, or one buried at the end of a lossy chain, does not get rescued by velocity. They get there sooner, with more conviction, having spent more.
When building was scarce, the scarce skill was the will to build. Now that anyone can build almost anything, the scarce skill is choosing the problem whose value is real before a single line of it exists. The return on what you build is not discovered afterward in a deck. It is decided at the moment you choose what to build, and everything downstream is just collection.
That choice — what is actually worth building — is the last scarce thing.
Pick well. Then build like maniacs.
A Note for the Curious
None of this is original to me, though I arrived at it by argument before I found the names. The endless “so what” is the epistemic regress argument from ancient skepticism. The split between borrowed and real value is the old distinction between instrumental and intrinsic (or terminal) value. And the lossy chain lives in Goodhart’s law (the closer a metric sits to the real goal, the better it holds) and in the statistics of causal mediation, where each noisy step attenuates what you can measure until little survives. I find it reassuring, not deflating, that the same structure turns up in epistemology, ethics, and statistics. It usually means the thing is real.
You're not a genius, Animesh, but you're pretty smart. Smart enough for me to pay, and I'm smart enough to keep reading.