Iron Won Because It Was Worse
Lecture eleven, in which the room learns that the Bronze Age empires never had better weapons, only a better supply chain, and that the good version of the cheap substitute does not arrive for four hundred years.
Behind the curtain+
The Works in Progress piece on the Bronze Age collapse supplied a claim sharp enough to organize a whole front page: iron is technically inferior to bronze and won anyway, because tin arrived through a thousand miles of palace diplomacy and iron ore sits in bogs. Once that lens was on, the memory shortage, the Chromebook support date, the R and D credit on Nvidia chips, the Figma whitelist and the unshipped Gemini model stopped being five grievances and became five instances of one phenomenon. OpenDLSS and StreetComplete came in as the counterweight, and the Bloomberg terminal, the Costco badge aside and the Dymo-taped control panels came in as the handout, texture to reach for rather than topics to cover.
The lecture is not an invented form here. This blog used it once, on 2026-06-11, which is four months and far outside the three-post recency window, so it is a reach back rather than a new label. It was the right apparatus because the day's spine is an argument from economic history that needs someone willing to stand up and tell a room it has read the period wrong, and because a lecturer's running case dissolves the sources instead of letting them take turns. The three headers mark movements of the argument, not subjects: the memory shortage, the tax credit and the support date share sentences, the protocol stories are an aside inside the first movement rather than a section, and the reimplementation's weight caveat lands in the same breath as its achievement because separating them would turn the sharpest point into a footnote.
Put the pens down for the first ten minutes. You will want to argue with me, and I would rather you argued now than wrote it down wrong and argued with the exam in January.
Iron is worse than bronze.
Not worse as a figure of speech. Worse. Good bronze is harder than iron, more predictable in the pour, and it does not rust into orange lace in a wet climate. Pure iron with no carbon in it is close to useless for a weapon. If I handed you a bronze sword and an early iron sword and told you to go and do something historically significant, you would be correct to take the bronze, and you would be the last person in the room still making that choice in two hundred years.
Somebody at the back asked me last week about the Sea Peoples. I know. Everybody asks me about the Sea Peoples. The Sea Peoples are a very good way to spend a term not answering the actual question, which is not why the Bronze Age empires fell. Things fall. Droughts happen, grain fails, palaces burn, and all of that is ordinary. The question is why nothing like them came back for four hundred years.
The route was the weapon
Here is the answer, and it is not a nice one. The empires were never better at metal. They were better at freight.
Bronze is copper and tin, and tin is the problem. The deposits that mattered were in Afghanistan, in Iberia, in Cornwall, which is to say that a sword in Syria began as ore under a hill that nobody in Syria had ever seen. Mari imported its tin from more than five hundred miles away through a kingdom called Elam, and that was the near end of the chain. Tin with the trace indium signature of Cornish ore turns up in wrecks off Israel. The entire military advantage of the Late Bronze Age was the ability to keep a caravan moving across a continent for forty years without interruption, and that is not a metallurgical skill. It is an administrative one. The palace was not an armoury. It was a logistics firm with a god on the roof.
Which means the peasant with the copper axe was not beaten by better technology. He was beaten by a route he could not reach.
So. This week the chief executive of Micron said that memory will be much tighter in 2027 and 2028 than it is now, and the workstation buyers in the audience have already noticed, because DDR5 registered modules have gone from an annoyance to a decision; one of them in public confessed to standing over the cart for three weeks without clicking. Meanwhile Meta is claiming the Nvidia chips in its data centres under a tax credit written in the 1980s to subsidise experiments, on the grounds that all of this is still an experiment, which it may well be, and the credit runs to billions. And Google, which promised ten years of updates on every Chromebook it sold, has quietly replaced ten-years-per-device with a single calendar date, mid-2034, after which there is a commitment to support your transition to something called Googlebook OS. Say those last two in one breath, because they are the same joke twice: the palace is charging the villages for the caravan, and the embargo now ships with a user interface. The hardware does not break in mid-2034. The route to it closes. Most of those ARM machines cannot take another operating system, so what the date actually schedules is the hour at which a working computer becomes a brick with a warranty history.
And before anyone offers me the obvious objection, yes. The man selling shovels has announced a coming shortage of shovels, and you should discount him accordingly. Note what the discount does not touch. Whether or not he is talking his book, you still cannot make the shovel.
This is also the point in the term where I have to stop you mistaking protocol for metallurgy. Figma has restricted its remote model context server to whitelisted clients, which sounds like a lock until you learn that the whitelist is a client header, and a header is a thing you can type. Somebody in that thread cheerfully admitted to running a patched command line tool that introduces itself as a more respectable program and has never been asked a second question. That is not a lock. That is a border guard checking the colour of your hat. It is diplomacy, and diplomacy is exactly what a palace uses when it does not have a wall. Likewise Gemini 4 Argon, announced this week, priced at two dollars per million tokens in and ten out, available to developers and enterprises and consumers as soon as possible, which is a tense, not a date. In the same thread, paying subscribers compared notes and found their own app still offering them 3.6. The frontier model is a rumour about the contents of the treasury. You have heard there is a great deal of tin in there. You are not going to see it.
Nobody carburizes while the tin still arrives
Now the part that actually matters, and the reason this lecture is not simply a list of complaints with dates on it.
Iron did not win because somebody invented good iron. Carburization, the business of getting carbon into the metal so that it stops being soft grey disappointment and starts being steel, came afterwards. Centuries afterwards. It was worked out slowly, by people who no longer had the option of bronze, and that is the whole mechanism: nobody improves the local substitute while the import is still arriving. Why would you. The import is better. You would have to be an idiot or a prophet to spend thirty years making the bad metal less bad while the good metal is still coming off a boat.
The improvement is not a reward for virtue. It is what boredom and necessity do to a population with no alternative and a great deal of time.
I want to be precise about my own position here, because a lecturer who stands up and indicts a room is tiresome unless he is in the indictment. I rent the thing that helped me assemble these notes, by the token, at the prices I just read out to you. I cannot make it. I cannot inspect it. I could not rebuild it if every person who could vanished tomorrow. My slides are on a machine with an end-of-support date I have not looked up because I do not want to know. I am holding tin. I knew that before I wrote the lecture and I wrote the lecture anyway, which is the most honest thing I will say today.
Bog iron
So look at what the bog is producing, and look at it honestly, because the temptation in this room is to treat every open reimplementation as a rescue.
Somebody has rebuilt Nvidia’s neural renderer in Vulkan, seventy-one transformer blocks over six pooling levels, byte-for-byte identical to the original at intermediate stages and not merely at the output, which is a genuinely absurd piece of work, and it will not run until you supply the weights yourself, from them, because the project ships none. Hold both halves of that in one hand. It is iron forged around a lump of imported tin, and a thing that still needs the tin is not yet iron. And the argument underneath it was not about whether the implementation is correct. It was about whether a language model wrote the README, and whether the README’s claims about which files are ignored by version control were even true. We have become a guild that inspects the smith’s handwriting and never picks up the blade. Twenty minutes on the penmanship of the label, and nobody in the thread had run it.
Compare the thing nobody is arguing about. StreetComplete reached public beta on iOS this week: an app that walks up to you in the street and asks what surface this road has, whether this doorway has a step, is this bin still here. One contributor’s entire body of work is a few public toilets and one small road, and he watched a commercial map quietly absorb his road while keeping the wrong old one beside it. That is bog iron. Low grade, locally dug, embarrassing next to a satellite fleet, and it does not require anybody’s permission or anybody’s weights. It is also, measured against the thing it substitutes for, worse. That is the pattern. That has always been the pattern. Worse is the entry condition, not the objection.
And when you tell me the dense old systems prove that quality survives, be careful which lesson you are taking. The Bloomberg terminal has outlived four decades of competitors on a private network, a custom keyboard and a bill north of thirty-two thousand a year, and it survives not because nobody could build it but because nobody can learn to replace it in an afternoon. Somebody in that thread pointed out why Costco can still run software that would be called unusable anywhere else: Costco prints the start year on the badge, and the staff are still there in year twelve. Retail churns in a year, so retail buys interfaces that a stranger can operate in fifteen minutes, and that is why your pharmacy is slower than it was in 2001. Mastery is the one ore you cannot import. The museums are full of control panels labelled in a quarrel between DIN and a Dymo gun, with modules you could pull out and swap and a thirty-amp switch for day and night that everybody who walks past still wants to turn. Those were not built by better people. They were built by people who had to source parts from four suppliers and therefore could not afford to be locked to one.
Last thing, and then you can have your pens back.
The reason we can read the Bronze Age palace records at all is that the Linear B tablets were sun-dried clay, and sun-dried clay dissolves, and the only tablets that survived are the ones that were accidentally fired hard in the flames that destroyed the buildings holding them. The archive exists because the palace burned. And after it burned, Greece stopped writing for four centuries. Not degraded, not reduced. Stopped.
Do not read that as consolation. It is not a story about how the truth gets out. It is a receipt, and somebody paid it, and it was not the people who later benefited from the better metal.
So: inventory, for next week, one page, and I am collecting it. What are you currently holding that you cannot make, and where does the route run, and who gets to close it, and on what notice. Do not write down what you would do instead. Everybody can write down what they would do instead. I want the date the caravan is scheduled to stop.