You mean siphon money from USPS

> they'd build up an institutional set of training data for machine learning and could build the best handwriting recognizer on the planet.

You mean, they'd use private people's data to build up a closed dataset they'd sell to other companies for money.

Is it a closed dataset, or are they selling it to other companies? Your language evokes privacy concerns. I’m sure you know that, and I’m sure that that’s intentional. Using someone’s handwriting as training data for OCR, as part of a training dataset significantly large enough to be used as training data in the first place, does not pose a material additional privacy concern.

But it does. It so very much does, because it always does. Always. Every time.

Data gets anonymized, then other firms de-anonymize. Co-mingled anonymized, too.

And once the MBAs see it, it will be pried away, and sold, and processed, and snarfed, with hooks during training too.

Because it always happens. Always. Every time. It's like candy to a child, like food to a starving man.

At this point, we are insane to think data stored or even observed, is not being monetized against us.

So the parent's post is not alarmist, because it always, always happens.

USPS literally processes people's private addresses. I'm amazed how you can say "Does not pose privacy concern" with a straight face when the entirety of the past decade has shown us that if there's private data, companies will abuse it, sell it, leak it etc.

It'd get delivered, if it does, because rather than drop mail they'll send it to Gloucester and the delivery people in Gloucester might be able to fix that, so nothing to do with this facility.

The issues that we (as keyers) had with the software in development were virtually impossible to address, given that there were several corporate layers between the contacts. The one time that a load of developers came to our data centre to observe keyers in action, one of our quick-thinking lower managers managed to snag a contact detail and for about a month there was an easy option to get small niggles fixed.

oh, btw my tph was 2-3000 so I don't know why the USPS workers are so slow. Actually thats just boasting, I probably do - it is likely that the software is a whole lot better than it was, which means that the quality of the images that it cannot handle has gone right down and takes longer for a human to process as a result.

The work-load almost certainly has changed with technological improvement. Back then, my recollection is that we'd see more machine type than handwritten addresses, but only slightly. Also we were only filling in the gaps depending on how well the original OCR worked. Sometimes we'd have to key in the important portions of the whole address (house number, part of a street name, postal code)... sometimes just the postal code, and sometimes only a portion of the postal code. Some of these data entry modes were faster than others; the extended "zip code" data entry only required two numeric keystrokes whereas the full data entry mode required much more time.

One of my most valuable lessons working there as a green engineer - you can figure out more in an hour sitting down with the end user then months of meetings before then.

The article says, "Our keyers key about a thousand images an hour, so they're spending about 2 seconds on an image".

Are keyers typing whole addresses in that quickly?

At least in the UK, if we had a postcode then that was enough. Not even required to type the whole thing. The UK postcode is split into two parts, the first part (outcode) designates the sorting office to send to and for newly posted mail on the afternoon shift then thats the main thing they wanted immediately; the image has been taken and the mail item has been tagged and will be in a van for some hours. If no postcode then usually the postal town is enough to get it on its way. When I was there, the second part (incode) was not done but I think these days that is processed by the night shift before the posties came in at 5am to take it for delivery.

Looking at the pictures in the article, they only have one address on the screen at a time. We had two; the one you are keying and the next one stacked above. It doesn't take long to learn to pipeline, so you are looking at the top image all the time but typing the details from the bottom one which was previously up top. There was an indicator to the side saying what it wanted and you generally only had to type that or various rejects.

also, 70 hours a week of that is pretty insane. Our normal was 40 and sometimes (eg xmas) up to 20 hours overtime was available. We worked 60 mins for each 10 minute eye break, and one 30 min meal break during the shift.