1.5: Data in the Dirt

Our partners, the Afghan Local Police (ALP). Author’s photo. Afghanistan, 2011.

Google Earth wasn’t the only free software tool I was able to leverage in Afghanistan. In fact, the first time I dabbled with machine learning was all the way back in 2010, over a decade before the launch of ChatGPT and DALL-E, when I used it to help keep track of faces.¹

The VSO mission had one of the densest information handovers I’ve ever had to manage during a RIP.² There was the normal military things like equipment to be signed for an target decks to be briefed. But most of what the outgoing team leader tried desperately to pour into my head was unclassified information that was much more vital to the success of our mission.

It was who was who in the valley. Not just names, though even that was complicated. Every person’s identity was name, father’s name, tribe, village. There were six different guys named ‘Palawan’ just among the ones I knew.³ I had to know who hated who? Who worked with who? Where did one tribe’s boundary stop and another start? This wasn’t easy when the tribes didn’t even agree amongst themselves. Which tribe had started a war in the valley five years ago? Which tribe did it a hundred years ago? Where were the mosques? Where were the wells? Where was it safe for me to rip around in man-jams and a motorcycle?⁴ Where should I never go without a gun truck?

Handover documents are typically PowerPoint slides in the Army, which means instead of georectified maps, you’re just getting arts and crafts. Colorful boxes traced over screen grabs. In the case of VSO, so much of the VSO handover was about the people it meant there were dozens of PowerPoint slides with small headshots on them. Except back in 2010 the bandwidth of outstations was abysmal. Trying to email (because it was always email) files would often choke out our system, even those just over just three megabytes, or roughly the file size of a short song in mp3. To try and help the sharing of files, all the PowerPoint images were compacted via a program called NXPowerlite, which reduced the images resolution to save file size. It worked exactly as it was supposed to. Except after one or two saves, every picture of every person was basically so pixelated you couldn’t tell anyone apart anymore. It didn’t help the handover much.

As we got started working in the valley, we began taking our own pictures. To prevent viruses from getting onto the DoD network we had a standalone unclassified laptop where all the files would get uploaded and scanned before burning them to CDs to transfer to the secure network. On that laptop was a folder which was almost immediately overflowing with hundreds of photos in it. The upside was a ton of data, but the downside was it quickly became unwieldy. One of the soldiers wanted to delete the old pictures, but I have a hoarder’s instinct when it comes to data, and I am almost always loathe to delete it. Storage is cheaper than data recovery. We just needed a way to give the data structure. We tried folders, but that ended up being a lot of additional work. So I went googling and found Picasa.

Distributed for free by Google, Picasa made organizing photos a lot easier because it leveraged the metadata already embedded in the picture. No need to name the file anything specific or put it into any folder tree, you could find the pics you wanted by filtering the metadata your camera adds to the file for you by default. No additional work, doing the action does the work. What I didn’t know at the time was Picasa also had facial recognition built in. I found out when it grouped a bunch of photos without me even asking.

When I was dumping another batch of pictures onto the standalone, Picasa opened its own tab on the screen and generated a Hollywood Squares of every face it found on its own in the pictures. It then asked me who they were. I clicked on the first one and typed in ‘Haji Amir’, the name of the local whose headshot was center screen. Immediately the squares reshuffled and suddenly I was looking at five more pics. The app was asking me, is this also Haji Amir? Four were, so I just clicked the check marks. One wasn’t so I instead typed in ‘Palawan’. Another shuffle and three more new pictures appeared. This was my first encounter with machine learning.

Tagging people in pictures became a thing you did at night to help you stay awake while pulling guard in the operations center. In our Picasa library you had every single key player in the valley, typically with up to 100 or more photos of each, at full resolution. Even better, when you clicked on the baseball card image, it brought up the whole picture, so it was easier to see who hung out with whom, as the same people did or didn’t come up in the shared pictures. We’d inadvertently built a poor man’s yearbook for the valley.

Back over on Google Earth, I had a single KMZ file that had every inch of data I could get ahold of about our valley. Every tribal map, check point, school, bridge; every historical report we could find for the entire valley, going back years. It was interactive, easily editable, and you could turn on or off the layers you wanted with a click of a mouse. All in 3mbs. It was hilariously small. Just one of the five different PowerPoints we’d inherited, already NXpowerlited to irrelevance, was over 20mbs. Once a month I emailed a copy of the KMZ file to myself. That way, if I was on the road and I needed to explain something to another leader, I’d just download it and open it up on Google Earth in minutes.⁵

My team was able to get roads built, resupplies delivered, and a local economy tied into the district, all with a 3mb file in their email. We could walk into the offices of coalition partners, Department of State, and NGOs and in minutes brief them on our valley, all on Google Earth.⁶ After about a month we realized the old PowerPoint handover files we’d been maintaining from the previous team were all but irrelevant. Instead, we focused on keeping the KMZ up to date and the Picasa library running. When it came time for us to RIP with the next team, I knew we’d given them more than just pixelated bowling pins and what they could hastily scribble into a green notebook. We gave them a living database of information they could interact with and easily build upon.

Leaving Afghanistan also meant leaving my first command. This meant a pause on my autonomy to drive data by fiat; my days as a data warlord were over. Instead, I’d need to start focusing on changing the organizational processes around me.

1. If you don't know what ChatGPT & DALL-E are, you are part of the problem.

2. Relief in Place. When two units hand over, typically done at the end of a deployment. A new unit comes in and you try to cram over nine months of experience into their brains in one week before you leave. Also called RIP/TOA (Transfer of Authority)

3. That wasn’t actually any of their real name as I learned later. It’s a nickname. It means ‘wrestler’ and is given to big guys or, because every culture loves irony, to small ones. This convention is the same at Michigan State where every dorm had its own ‘Moose’.

4. Really called a ‘shalwar kameez’, they were nick-named ‘man-jams’ because they are as comfy to wear as pajamas.

5. It took just a few more minutes if I had to download Google Earth first, which thanks to COIC was only a few clicks away.

6. Non-Governmental Organizations. They are typically, but not exclusively, nonprofit entities often focused on public good and humanitarian issues.