Embeddings in Bitcoin: what people hid in the chain after the genesis block.

On the morning of 3 January 2009 the front page of The Times carried the headline "Chancellor on brink of second bailout for banks". Some time that day Satoshi Nakamoto typed it, with the date, into the first block of Bitcoin, the block now called the genesis block. It sits in a field that only the person who mines a block gets to fill, a few spare bytes at the front of the transaction that pays the miner, and it has been read as a timestamp, a manifesto and a joke ever since. It is also the first thing anyone wrote into Bitcoin that was not money.

Bitcoin was never meant to hold writing. Every byte in a block is supposed to be a coin moving from one owner to another. But a transaction is only a list of numbers, and a few of those numbers are loose enough that a person can put whatever they like in them, as long as they pay the fee. Bitquery indexes the complete Bitcoin chain, so we read every block, from Satoshi's headline to the first days of September 2026, and pulled out every embedding, every place where the numbers turn out to be words. There are four such places. The oldest, the miner's field where the headline lives, is the one people talk about, and it turns out to be the quietest. The busiest by far is a slot Bitcoin added in 2014 specifically so that people would stop doing damage elsewhere. The most surprising is the damage itself, a trick with fake addresses that has been used to write whole books into the chain, one line at a time, in a form no computer running Bitcoin can ever delete.

Three things came out of the reading. The tradition Satoshi started, of miners leaving notes in their blocks, peaked in 2015 and is gone. The real library on the chain is written in fake addresses, tens of thousands of them, holding a papal encyclical, a human-rights statement on Tiananmen, an epic in Sanskrit and a tribute to a murdered cryptographer. And since Bitcoin Core lifted its 80-byte limit on data outputs in October 2025, the chain has become a billboard, with one five-day campaign of more than 13,000 transactions aimed at a single developer. One limit on what follows has to be said up front. Our index is missing the miner's field for about a fifth of all blocks, all of them between 2016 and 2021, so for those years the count of miners' notes is a floor and not a measurement.

Four places to hide a sentence

A Bitcoin transaction has inputs, which point at coins being spent, and outputs, which say who gets them. Each output carries a small program, usually one line long, that names the owner by a 20-byte fingerprint of their key. Each input carries the signature that satisfies the program on the coin it spends. Nothing in that design has room for a message, and yet every part of it leaks a little.

The first leak is the coinbase. The transaction that pays the miner has an input that spends nothing, so its signature field is free space, and the miner can put anything in it. That is where the headline lives, and where mining pools now stamp their names. Only the miner of a block can write there, and there are only about 144 blocks a day, so the field is scarce.

The second is a fake address. An output that pays a fingerprint pays whoever holds the matching key. If the fingerprint is not a fingerprint at all but 20 letters of text, nobody holds the key, the coins are gone for good, and the text is on the chain. This works in any year and anyone can do it. It is also the most expensive trick, though the fee does not show it. Every node on the network keeps a list of unspent outputs so it can check new spends, and a fake-address output sits in that list forever, because it can never be spent.

The third is a fake public key. Some older output types carry whole keys rather than fingerprints, and a key holds three times the text a fingerprint does. This was the trick of choice in Bitcoin's first years and it is how the WikiLeaks cables were pointed at from inside the chain.

The fourth is OP_RETURN. In 2014 Bitcoin's developers gave up fighting the fake-address trick and added an approved slot, an output type that announces itself as data, pays nobody, and is dropped from the unspent list the moment it arrives. In return they capped it, at 40 bytes and soon 80. The cap held for eleven years.

The place Satoshi used turned out to be a trickle of pool names with a few human years in the middle. The approved slot carries far more, but most of it is machine memos. The fake addresses are where the writing is.

The years the miners talked

For the first two years nobody followed Satoshi's example. The coinbase carried the nonce, the number a miner changes while hunting for a valid block, and nothing else. Then, in the summer of 2011, a pool called Eligius, run by the developer Luke Dashjr, began writing lines of Latin into its blocks: Benedictum Nomen Iesu, Benedicta excelsa Mater Dei, a cycle of Catholic blessings that ran across fifteen blocks that June. The reply came two months later, in the fake-key channel rather than the coinbase, and it was an obscene attack on Dashjr by name that ended "Stop polluting the blockchain with your nonsense." Every argument Bitcoin has had since about writing on the chain was already there in August 2011, including the irony that the complaint was itself written on the chain.

The coinbase's human years belong to the pools that let their miners type. BTC Guild, Eligius and then the Chinese pools HaoBTC and BTCChina put a text box in front of their users, and the users used it. The 2015 crop is the largest, with aphorisms in capitals, birthday wishes, Matrix quotes, "Long live Ukraine!" in three blocks on a single September day. In June 2016 a HaoBTC miner named Sun Chun Yu proposed to Zhuang Yuan in a block. Her answer, "YES!", is in another one, two hours later. Every one of those blocks was worth 25 bitcoin to whoever mined it, and the proposal cost nothing extra.

After 2016 the notes thin out to almost nothing. From 2017 the coinbase carries pool names, worker identifiers, and a few slogans that are politics of a different kind. Foundry's blocks say "#dropgold". Marathon's say "Made in USA", and for two weeks in May 2021 they said "OFAC Compliant Block", a tag that appears in two dozen blocks in our index and then stops. Antpool blocks still carry a customer's birthday wish now and then, which is why the bars do not quite reach zero. But the thing Satoshi did, a person leaving a line in a block they had won, belongs to the years when individuals still mined. The miners are companies now, and companies leave logos.

The library nobody can burn

Six weeks after the Eligius blessings, on 30 July 2011, the security researcher Dan Kaminsky built a transaction with dozens of fake addresses in it. Read in order, their fingerprints spell a tribute to Len Sassaman, a cryptographer and friend who had died that month, complete with a portrait drawn in text characters. The transaction paid a heavy fee and left most of a bitcoin sitting in outputs that can never be spent. It is the earliest document we found on the chain, and it set the pattern for everything that followed.

A document is cut into pieces the length of an address, each piece is dressed up as one, and a transaction pays a few hundred satoshis, hundred-millionths of a bitcoin, to each one. The reader collects the outputs and glues them back together. Nobody can spend the coins, so every piece stays in the unspent list on every node, permanently. The developers built OP_RETURN in 2014 as the alternative, a data output that nodes could forget. It did not stop the address trick, and 2014 and 2015 were the trick's biggest years.

What people chose to make permanent is the interesting part. In September 2014 someone uploaded long passages of the Mahabharata, in Sanskrit, across more than a hundred transactions in a single evening. On the fourth day of the Umbrella protests that October, six transactions carried news coverage of Hong Kong's occupied streets. The week Pope Francis published his encyclical on climate, Laudato Si', in June 2015, someone spent two hours putting the whole English text into the chain, sixty transactions long. In March 2017 a single transaction carried a Human Rights Watch statement demanding that China acknowledge the Tiananmen killings. Alongside these are the things you would expect from an archive built by whoever showed up: biographies of Gandhi, Einstein and Michael Jackson, Valentine's messages, a birthday greeting from "The Fantastic Four Gang", and a long Italian obscenity.

The WikiLeaks entry is the odd one out, and it is the one people have heard of. On 8 April 2013 a transaction wrote "Wikileaks Cablegate Backup" and the name of a compressed archive into fake public keys, with a pointer to a second transaction holding a download tool named after Satoshi. The archive itself is in the chain too, spread over the outputs of the transactions around it. It cost 0.048 BTC in fees, and whether it was a backup, a stunt or a way of putting the leak beyond seizure, it left a government's cables somewhere no court can order deleted.

The text outputs hold about three bitcoin between them, locked for good, which is a small bill. Each one is also a row that every full node has carried since the day it was written and will carry until the network stops, and the uploaders of 2014 and 2015 added tens of thousands of them in two years. Our sibling piece on what inscriptions did to Bitcoin measured the same kind of permanent state from a different decade; the address trick was the original version of it.

The sanctioned slot, and what actually goes in it

OP_RETURN was meant to be the compromise. Write your data here, pay for the bytes, and the network will forget it once the block is stored. People took the offer. There are a quarter of a billion of these outputs on the chain, and a few million of them decode as readable text. Almost none of that text was written for a person to read. Most of it is memos that a cross-chain swap service writes to itself, records from a name registry, hashes with a prefix, token mints and JSON. Filtered down to sentences of two or more words, the human part of OP_RETURN is about 15,000 distinct messages in twelve years.

The sentences themselves are what you would find on any wall. Love notes in German and Portuguese, birth announcements, "Carving my name into the blockchain", a good deal of Satoshi impersonation, the odd confession about faxing a seed phrase to customer support. A few are history. On the anniversary of the genesis block in 2022, someone wrote "dedicated to Hal Finney on the 3rd of January, 2009. I'll be next", in one transaction. The day after Alexei Navalny died in 2024, another carried "RIP Navalny. Someday Russia will be free" and, in Russian, "Stop the war with Ukraine".

The repeated messages tell a different story, about who pays to say the same thing thousands of times. The biggest single campaign before 2026 was a protest in the block-size fight of March 2016, "503: Bitcoin over capacity!", repeated across a hundred blocks. The next was a buyer's advertisement with a Protonmail address, then a wallet-recovery service, then a birthday message from the exchange Gemini for the pizza-day block.

The cap comes off

In October 2025 Bitcoin Core shipped version 30, and with it the removal of the 80-byte limit that nodes had applied to OP_RETURN outputs before passing them on to other nodes. The rules every node enforces did not change, and a miner could always include a bigger output if someone handed it to them directly. The change was that ordinary nodes would now pass such transactions along. A rival version of the software, Bitcoin Knots, maintained by Luke Dashjr, kept the cap and tightened it, and the two camps spent the winter arguing in public about what a node should agree to carry.

The chain shows what the change did. Before v30, a text message over 80 bytes appeared a few dozen times in most months, and each one had to be handed straight to a miner. By February 2026 there were thousands. The big single payloads arrived too, the full text of the United States Constitution in one output that May, and Satoshi Nakamoto's original announcement email in June. The largest image we found, nearly a hundred kilobytes of it, is older than the change. Marathon mined it in May 2025, and at that size it could not have travelled through ordinary nodes, so someone handed it to the miner directly.

What the new room was used for first was a fight about the room itself. On 26 February 2026, between one in the morning and half past four in the afternoon, 4,446 transactions from some fifteen hundred addresses put thousands of distinct rants against inscriptions, the images and tokens written into blocks since 2023, and against the miners who accept them, into seventeen blocks. "Fee market manipulation 101", "spam protesting spam", "miners enabling spam for short-term fees are traitors": the phrasing has the cadence of text generated by a model and posted by a script, and the whole day cost a few hundredths of a bitcoin in fees, a few hundred satoshis a message. On every measure that the chain records, it was the thing it was complaining about.

Three weeks later came something more pointed. Over five days in late March 2026, 13,062 transactions carried the same two sentences: "Luke Dash Jr went to the OP_NET House. He's on the OP_NET List". Each came from a separate address of the newest format, Taproot, one address per message, which is the signature of a script working through a list of freshly made keys. It cost its sender a tenth of a bitcoin in fees, a few hundred satoshis a message, about $7,000 at the price of the day, to have the maintainer of the client that wanted to limit OP_RETURN taunted by name in a place that cannot be edited. The chain records the message and the cost. It does not record who paid.

Cheap to write, expensive to keep

Put the campaigns next to the documents and the economics of writing on Bitcoin come out backwards. Kaminsky's tribute to Sassaman cost close to a whole bitcoin in 2011, when a bitcoin was worth about $13, and it will occupy its rows in every node's unspent list for as long as there are nodes. The encyclical cost less than a tenth of one in 2015. The OP_NET taunt, thirteen thousand messages that every node was free to discard, also cost about a tenth of one, at 2026 prices. Counted in bitcoin per line, the taunt paid about one fifteen-hundredth of what the tribute paid. Counted in the dollars of the day, it paid about three times as much, because the coin had gone from about $13 to about $70,000 in between. The 80-byte cap had priced one kind of speech, the long kind, out of the approved slot and into the permanent one. Lifting it made long writing cheap and disposable at the same time.

Satoshi's headline was possible because a block has a field nobody polices. Seventeen years on, the same chain holds an encyclical that no node can drop, a human-rights statement addressed to a government that bans it, and thirteen thousand copies of a taunt that any node may discard. Which of the four places gets used depends on what the software will carry that year. Who paid for the taunt, the chain does not say.

Blocks and transactions behind this piece