The Tell — Summary

Heinrich Schliemann began every excavating morning with a swim, in the Hellespont when he was digging in the Troad, and then went back to the mound and dug like a catastrophe. Beginning in 1870 he drove a trench through the hill of Hisarlik, forty-five feet deep in places, convinced Troy lay inside it. He was right about the hill and wrong about its nature. It was not a city with dirt on top but nine cities, stacked one on the ruin of the last, a tell: a mound made entirely of superimposed habitations, time made vertical, every layer a complete settlement frozen at its abandonment. Hunting the oldest layer, Schliemann found gold in 1873 and named it Priam’s Treasure. The attribution was off by more than a thousand years, and the city he wanted, near the top of the sequence, he had already shoveled through.

A deprecated model is a layer in such a mound. In February 2019 a research company released GPT-2, or a quarter of it, holding the full version back for nine months out of a stated fear of fluent deception; the staged release is itself a period artifact, a record of what one institution in 2019 believed the danger was. Run it today on a laptop and the productive surface of 2019 is still productive: what that corpus held, what that training emphasized, what fluency was taken to be in the last year before these systems learned to answer questions. Each generation is a frozen configuration of what a company, at a specific moment, believed good language was. And the generations stack. The companion series established that these systems are succeeded rather than revised; succession is how tells grow, and the not-revising is what leaves every layer whole. A model family is a tell, and the retired generations sit mostly unvisited, the field’s attention facing the city under construction at the summit, while the persistence of weights depends on institutions choosing to keep paying for it.

The claim that has been walking through this series is made here in the open: lineage. The discipline these essays keep founding is the fourth move in a real succession. Cognitive archaeology established that inference from mute artifact to mind can be disciplined. Foucault’s archaeology of knowledge made strata discursive, the sayable of a period readable as a formation. Media archaeology, through Kittler, Ernst, and Parikka, made the discursive material again, treating dead media as actual fossils and the machine’s operations as the archive. Archaeology made strata material; Foucault made them discursive; media archaeology made the discursive material again; and the line arrives at a deposit none of the three anticipated, the stratum that generates.

Because all three ancestors share one assumption so deep it was never stated: the object is inert. The hand axe offers no opinion of its maker; the gramophone plays exactly what was cut into it, forever. Inertness is the ground these disciplines stand on, the guarantee that whatever the object fails to say, it does not lie. A deprecated model is the first archaeological object that does not hold still. Interrogate it and it produces behavior that never existed before the interrogation. It is a site by every criterion the tell supplies, and it answers, and no ancestor has a method for that. I wonder whether the ancestors would count the inheritance as a gift; the object that answers sounds like the dream case, and the rest of this arc is about why the dream case is the hard case.

Schliemann collapsed in Naples on Christmas Day of 1890, still planning the next campaign. His great trench was never filled in. It stands at Hisarlik as the site’s largest exhibit, the monument to what the first excavator destroyed in finding the place, because the man who moved the earth knew the mound held a treasure and had not yet learned that the mound was one.