8.6.23

Machine Learning by Matt Dube

The summer after freshman year, I worked for the USPS alongside my friend Kent and his younger sister Nora. We were hired to review the performance of the MLOCR-B, a next level mail sorting machine. The evident enthusiasm of the folks who trained us, gray adults with big empty smiles, was suspicious to us, as was the fact that we were never invited to see the machine. Even in the training video tapes we watched which celebrated the service of the humble five digit zip code before hyping the dawn of a new four-digit suffix, nine digits total, that the MLOCR-B could recognize, there wasn’t so much as a single still photograph of the machine itself. 

We learned the machine from its output, the flat single-ply cardstock slips trimmed to the size of a standard business envelope that we compared to the readout that came wrapped around them with a rubber band, each bundle one hundred forty-four slips. For each slip, there was a corresponding address and we marked each as successful or misdirected. The MLOCR-B did an OK job, as good as any ordinary person. Some slips came to us dinged, nicked, folded at the corner, and it was hard to tell if this was to test the machine or if it was a legit accident; it frustrated the sorting capabilities of the MLOCR-B and us, trying to reconcile the laser jet address and barcode. “Anyone would struggle with this,” we said of the marred slips. “It’s not a fair test.” We were promised that future tests would handle actual mail, but we never saw that in our ten-week trial. Driving home after our six-hour shifts, we didn’t so much argue as try to talk through, in the way we were learning to do at college, whether what we were doing was taking away jobs from someone else we’d never meet or if they would be, like our trainers, excited to have this burden lifted from them. We didn’t believe in the future, really, but thought of it as a place alot like here where someday we’d be delivered.

Matt Dube