How AI connects to Crypto
One unbroken line runs from the first inscriptions cut into clay to the blockchain, and the machinery that reads that line is the same machinery that became language models. Follow the arrows.
← scroll the schematic sideways on small screens →
The one idea in every node
Every link in the chain is about an inscription as a commitment: marks fixed in a medium so they can be preserved, read only by the right party, proved authentic, or agreed upon by strangers. Epigraphy studies inscriptions that survived because the medium was hard. A cipher makes reading depend on a key. Decipherment recovers a key from statistical structure. Cryptography makes the key mathematically unrecoverable. Cryptocurrency uses that math to produce a public inscription no authority needs to vouch for. Language models are the decipherer's statistical machinery turned into a general engine for reading and writing.
Epigraphy
The oldest inscriptions are not poetry. Sumerian cuneiform from about 3200 BCE is overwhelmingly accounting: grain, sheep, rations, debts. Schmandt-Besserat traced writing itself back to clay tokens sealed in bullae, which means the first writing system was a record of who owed what to whom. Coins joined the tradition in Lydia around 600 BCE. A coin's legend and image are an inscription whose job is to prove issuer and authenticity and to resist forgery. Numismatics is a branch of epigraphy, and money has been an inscribed claim from the start.
Ciphering
The earliest known ciphers are themselves epigraphic objects: altered hieroglyphs in the tomb of Khnumhotep II around 1900 BCE, a Mesopotamian tablet that encrypts a pottery-glaze recipe, Atbash in the Hebrew Bible, the Spartan scytale, the Caesar shift. Secret writing is ordinary inscription plus one design goal: only the holder of the key can read it.
Deciphering
Champollion's Rosetta Stone was a known-plaintext attack, with Greek as the crib. Grotefend cracked Old Persian cuneiform in 1802 by guessing the royal formula "X, great king, king of kings, son of Y". Alice Kober's grids of inflection triplets and Ventris's 1952 reading of Linear B were a purely distributional attack with no bilingual at all. Al-Kindi's ninth-century treatise on deciphering messages introduced frequency analysis, developed as much for textual scholarship of the Quran as for ciphers. The same statistics read a lost script and break a substitution cipher.
Cryptography
Alberti's disk, Vigenère, Kasiski's break, Enigma and Bletchley are the escalation. Kerckhoffs's principle in 1883 moved all security into the key. Shannon's 1949 paper on secrecy systems then made the field a branch of information theory: entropy, redundancy, unicity distance, perfect secrecy. His crucial observation is that a cipher is breakable exactly because the plaintext language is redundant. English is roughly half redundant, and the cryptanalyst lives on that redundancy.
Mathematics
Hardy boasted in 1940 that number theory could never be used for war. Within four decades it was the whole game: Diffie and Hellman on the discrete logarithm in 1976, RSA on factoring in 1977, Koblitz and Miller on elliptic curves in 1985, Merkle's hash trees in 1979, and hash functions that turn any inscription into a short fingerprint. A digital signature is an inscription that proves authorship without a mint or a king.
Cryptocurrency
Haber and Stornetta's 1991 hash-chained timestamping, Back's Hashcash, Wei Dai's b-money, Szabo's bit gold, then Bitcoin in 2008 assembling ECDSA on secp256k1, SHA-256, Merkle trees and proof of work. A blockchain is literally an epigraphic object: an append-only ledger cut into a medium engineered to be immutable, whose entire security is that nobody can recut the stone. It is a ledger of debts and balances, the same genre as the Uruk tablets. The genesis block even carries an epigraph, the Times headline of 3 January 2009. Mining is seigniorage, the modern equivalent of the die-cutter's labour that made a Roman coin costly to counterfeit. On-chain analysis is decipherment: address clustering heuristics are Kober's grids applied to a pseudonymous script that was left without a Rosetta Stone on purpose.
Where AI and language modeling enter
- Shannon's 1951 guessing game measured the entropy of English by having people predict the next letter. Next-token prediction is that experiment at scale. The language model learns precisely the redundancy the cryptanalyst exploits, so the two are one computation seen from opposite sides.
- Bletchley's smoothing became NLP's smoothing. Turing's Banburismus scored evidence in decibans, and I. J. Good's estimator for unseen events, built to attack Enigma, became Good-Turing smoothing in every n-gram language model. That is a direct genealogical link, not an analogy.
- Warren Weaver's 1949 memo founded machine translation as codebreaking. A Russian article, he wrote, "is really written in English, but it has been coded in some strange symbols. I will now proceed to decode." IBM's statistical translation systems in the 1990s formalized that as the noisy channel, and that frame produced modern NLP.
- Decipherment is now an AI task. Snyder, Barzilay and Knight deciphered Ugaritic automatically in 2010 using Hebrew as a cognate. Knight's group cracked the Copiale cipher in 2011. Luo, Cao and Barzilay read Linear B and Ugaritic with neural models in 2019. DeepMind's Ithaca in 2022 restores, dates and geolocates damaged Greek inscriptions, and its successor Aeneas in 2025 does the same for Latin, which is epigraphy performed by a model. In August 2026 Claude Fable 5.1 solved Thomas Urquhart's Cyphral Distich, a cryptogram printed in 1653 and unsolved for 370 years, in 44 minutes, by noticing that the key was the book itself: each number indexes a word in the 32 numbered wishes printed just before it. That is Kober's move, structure found inside the text rather than a key brought from outside.
- Embeddings are Kober's grid made continuous. Distributional structure, Harris in 1954 and Firth's "company it keeps", is what both the decipherer and the vector model exploit. Unsupervised cross-lingual embedding alignment maps two languages onto each other with no dictionary at all: decipherment without a bilingual, done by geometry. Vector representations of relationships, patented at Lawrence Berkeley National Laboratory in 2005 ahead of the Word2Vec era, sit on this same line.
- Compression ties the chain together. Byte-pair encoding, the tokenizer under today's models, is a 1994 compression algorithm. Shannon showed compression, secrecy and prediction are the same quantity, and recent work states it plainly as "language modeling is compression". Everything in AI is loss minimization on vectors and tensors.
- Cryptocurrency gives models what coins gave states. Signatures let an agent prove identity and commit to an inscription; consensus lets machines agree on a shared ledger without trusting each other. The GPUs built for proof of work and for training are the same hardware lineage.
References and citations
Recent breakthroughs in epigraphy, cryptology and formal mathematics
- Vals AI, Claude Fable 5.1 Solves the Cyphral Distich, 31 August 2026. Urquhart's 1653 two-line cryptogram, number 28 on Klaus Schmeh's list of famous unsolved messages, decrypted in 44 minutes; the key was internal to the text. Coverage: Schneier on Security.
- Anthropic and Columbia University, first fully machine-checked formal proof of Fermat's Last Theorem in Lean, September 2026: 13 million lines, more than 30,000 supporting theorems, 11 days, parallel Claude agents; confirmed by Kevin Buzzard as proving the theorem from the axioms alone. Decrypt, Yahoo Tech. Wiles proved the theorem in 1995; the 2026 result is the formal verification of that proof.
- OpenAI, proposed resolution of the Navier-Stokes existence and smoothness problem, 8 September 2026: an internal agent system, on the order of 10,000 concurrent agents, produced an analytical finite-time-singularity argument in about 88 hours, then GPT-6 Astra formalized and verified it in Lean over a further 17 hours. The Rundown, DEV Community. A claim under community review at the time of writing; see evidence and dispute.
- Google DeepMind, Aeneas, Nature, July 2025: contextualizes, restores, dates and attributes Latin inscriptions, dating to within about 13 years. Smithsonian, Phys.org.
- Vesuvius Challenge, Herculaneum scrolls: first words read from an unopened carbonized scroll in 2024; a scroll virtually unwrapped in full, roughly 1.5 metres across 20 columns, announced June 2026. CNN, Scientific American.
- Assael, Sommerschield et al., Ithaca, Nature 603, 2022: restoring and attributing damaged Greek inscriptions with a deep network. Nature.
- Oranchak, Van Eycke and Blake, the Zodiac 340 cipher, December 2020: solved after 51 years with software-driven search over transposition and homophonic substitution, confirmed by the FBI. Wikipedia.
- Kambhatla, Bird and Knight, Solving Historical Dictionary Codes with a Neural Language Model, 2020. arXiv 2010.04746.
- Luo, Cao and Barzilay, Neural Decipherment via Minimum-Cost Flow: from Ugaritic to Linear B, ACL 2019. arXiv 1906.06718.
- Knight, Megyesi and Schaefer, The Copiale Cipher, 2011: an 18th-century 105-page encrypted manuscript broken with statistical machine translation methods. ACL Anthology.
- Snyder, Barzilay and Knight, A Statistical Model for Lost Language Decipherment, ACL 2010: Ugaritic deciphered automatically using Hebrew as the known relative. ACL Anthology.
- Conneau, Lample et al., Word Translation Without Parallel Data, ICLR 2018: aligning two embedding spaces with no dictionary, decipherment by geometry. arXiv 1710.04087.
- Delétang et al., Language Modeling Is Compression, 2023. arXiv 2309.10668.
Foundational sources behind each dot
- Schmandt-Besserat, Before Writing: From Counting to Cuneiform, 1992. Clay tokens, bullae and the accounting origin of script.
- Mrozek, Coins, Money and Epigraphy, 1st to 3rd c., Moneta series; announced on the epigraphy list by Georges Depeyrot, 20 September 2003, and cited in the founder's note below.
- Al-Kindi, A Manuscript on Deciphering Cryptographic Messages, 9th century. Frequency analysis.
- Kerckhoffs, La cryptographie militaire, 1883. Security must live in the key.
- Shannon, Communication Theory of Secrecy Systems, Bell System Technical Journal, 1949. IEEE.
- Weaver, Translation, memorandum, 1949: machine translation framed as decoding. ACL Anthology.
- Shannon, Prediction and Entropy of Printed English, Bell System Technical Journal, 1951. The first next-letter language model.
- Good, The Population Frequencies of Species and the Estimation of Population Parameters, Biometrika, 1953. Good-Turing estimation, born at Bletchley.
- Harris, Distributional Structure, Word, 1954; Firth, A Synopsis of Linguistic Theory, 1957.
- Chadwick, The Decipherment of Linear B, 1958. Kober's grids and Ventris's solution.
- Diffie and Hellman, New Directions in Cryptography, IEEE Transactions on Information Theory, 1976.
- Rivest, Shamir and Adleman, A Method for Obtaining Digital Signatures and Public-Key Cryptosystems, Communications of the ACM, 1978.
- Merkle, A Certified Digital Signature, 1979. Hash trees.
- Haber and Stornetta, How to Time-Stamp a Digital Document, Journal of Cryptology, 1991. The hash chain.
- Nakamoto, Bitcoin: A Peer-to-Peer Electronic Cash System, 2008. bitcoin.org.
- Kahn, The Codebreakers, 1967; Singh, The Code Book, 1999. The general histories this page compresses.
- Franks et al., US Patent 7,987,191, System and Method for Constructing Relationship Networks, filed 2005: vector representations for detecting relationships between entities. Google Patents.
- Franks, note on Hacker News, 13 December 2018: NLP has its roots in epigraphy; epigraphy, ciphers, information security and cryptography are all related. news.ycombinator.com.
Expanded from a 2018 note by Cymetica founder Kasian Franks on Hacker News: "Epigraphy, ciphers/deciphering, info security, cryptography are all related." See also Vector · Matrix · Tensor.