An AI engineer skipped the prediction test

On May 22, Tom Di Mino was looking at a line of Linear A in which every word was already readable except the first. It was a prayer formula that recurs, in slightly different regional forms, across five sanctuary sites on Crete, and the known words came from the 60-odd signs Linear A shares with the deciphered Linear B. The unknown opening word was a verb built from five of those known signs plus one that appears only in Linear A, catalogued as *301. Di Mino read *301 as “na,” which gave him the root “nawaya,” meaning “to dwell.” That matches the three-consonant Semitic root N-W-Y, which means “to dwell or inhabit” in Hebrew, Akkadian and related languages. Read in full, the prayer looked like later Hebrew prayers, except that it was addressed to a goddess.

From that reading Di Mino, a self-taught AI engineer and amateur linguist in the Hudson Valley, claims to have deciphered Linear A, the Minoan script that has held out against linguists for more than a century. He has read around the subject for seven years, has some proficiency in eight languages including Attic Greek, Sanskrit, Arabic and Ugaritic, and started the work this January. He argues the script records an extinct Semitic language ancestral to Biblical Hebrew, Arabic and Aramaic, roughly as Latin is ancestral to Italian. The claim is under review by linguists at Rutgers and Cambridge. If it holds, it is the kind of result that put the 1952 decipherment of Linear B on the front page of the New York Times.

Linear A was used on Crete from about 1800 BC until 1450 BC, when Mycenaean Greeks took the island and adapted its signs into Linear B. Michael Ventris read Linear B as Greek in 1952, building on Alice Kober’s statistical work on how the signs patterned and shifted. Because 60 core syllabic signs carry over, experts could already guess how many Linear A signs sounded without knowing what the words meant. Thirteen signs occur only in Linear A and had no accepted sound value. There is far less Linear A than Linear B, and much of what survives is commodity inventories that say little about the grammar, which is part of why Linear B fell first.

The method is the part worth copying. Di Mino used Claude Code to build a suite of Python scripts that query and cross-reference the digitized Linear A corpus from the GORILA and SigLA databases, so he could test a hypothesis against every inscription at once rather than by hand. Kober and Ventris did the same kind of pattern-hunting on paper slips over years. The model did not read the script. It built the instrument that let one person check a guess across the whole corpus in an afternoon, which is a sensible division of labour: a human proposes the readings, the tooling says quickly whether they hold up elsewhere.

The draft lists proposed sound values for 40 signs, including the 13 that previously had none, plus values for five Linear B signs that were still open. There is a lexicon of 408 Linear A terms rendered in English and a nine-page manuscript titled “Ya Diktu: Grammar of the Minoan Peak Sanctuary Libation Formula.” None of the underlying ideas is wholly new. Cyrus Gordon argued Linear A was Semitic in a 1957 Antiquity article that never produced working translations and never caught on; Di Mino says his version actually reads the prayers, and claims to be the first to tie Linear A inscriptions to Hebrew ones rather than to the Akkadian and Phoenician tablets Gordon worked from.

The validation Di Mino offers is that his reading of the Linear A logograms also resolves problems in existing Linear B translations. That is a claim of internal consistency: the solution accounts for more than the data used to build it, which is a point in its favour. The thing that actually settles a decipherment is prediction, readings that keep coming out as coherent text on inscriptions not used to derive them, confirmed by people other than the author. Ventris’s Greek stuck because new tablets kept reading as sensible Greek. That out-of-sample check is the step still outstanding, and the only public write-up so far notes that its author knows Di Mino socially and has already been edited once for accuracy.

So treat the 408-term lexicon as a set of predictions, not confirmed readings. Whether Linear A was Semitic turns on whether those sign values hold up on prayer formulas and inventory tablets Di Mino did not use to build them, checked by reviewers who start from doubt. The Rutgers and Cambridge review is where that gets decided.