👾 5,653 clips I was never going to watch

For years, one project sat at the bottom of my list because it was too big to ever actually start. Last month I finished it.

In this issue:

I work with editors on my videos. They’re always asking me for footage. Often the same clips, more than once.

Here’s the thing: some of the videos I could find just fine.

If a clip is tied to a place, I usually remember when I was there, and photo search handles locations well now. I type “Almaty,” I get my Almaty footage.

What I could never find was the abstract stuff.

Me at a cafe table. Walking down some street. A hostel dorm room.

Search isn’t great with those yet, and neither am I, short of opening clips one at a time until I got lucky.

And I have a lotttttt of footage.

5,653 clips across 135 folders, on Google Drive alone. Plus everything in Google Photos. And there’s more sitting on old computers and hard drives I haven’t even gotten to yet.

Just the footage I’ve counted would take over 70 hours to watch. Two work weeks of nonstop watching, not including taking notes.

So organizing my footage lived on the someday list. The one where things go to never happen.

I handed the watching to AI.

The short version: I set up something that goes through every clip, one by one, looks at it, and writes down what it is.

What’s happening, roughly where it was shot, what kind of shot it is. Each clip becomes a line I can search.

Now, instead of me digging, an editor can look up “hostel dorm,” “walking, street” or “waterfall” and find it in seconds.

I’ll spare you the mechanics. (If you want the actual technical version, how it’s wired together and what I’d change, reply and I’ll send it. That’s a whole other email.)

The build was the boring part. What surprised me was running the process.

Everything started on my laptop.

Then I realized I’d been downloading so much footage that I’d throttled my internet for two days and couldn’t do anything (including work). So I moved the whole thing onto a rented machine in the cloud.

Good fix, with one side effect: now it was running somewhere I couldn’t really watch.

That mattered, because the thing barely showed its work.

It would grind through an entire folder before recording that it had done anything. So on the big folders, the counter sat frozen for hours while it worked the whole time.

I spent a lot of it convinced something was broken and it had died. It hadn’t.

Learning to tell “stalled” from “just quiet” was most of the work.

But there were also real failures underneath the fake ones.

One time I restarted it and it lost its API key. It spent a day failing every clip without telling me.

Another night it ran out of the credits it needed and stopped.

Both easy to fix once I found them.

The hard part was telling which kind of broken I was looking at.

The biggest speedup came from realizing the machine itself was the bottleneck.

When the largest folders were crawling, I made the rented machine four times bigger, and a day’s worth of work finished in about ninety minutes.

The smart part was building it somewhere I could resize on demand in the first place.

The thing that had felt impossible for years turned out not to be hard. It was just big.

That distinction matters more than it sounds. Difficult and big are different problems.

Difficult means you don’t know how (yet). Big means you know exactly how, and there’s simply too much of it for one person and a finite number of evenings.

For most of my life, “too big” was a wall. You couldn’t throw more of yourself at it than you had.

That wall has been moving.

The projects you’ve shelved specifically because of their size might be the ones worth a second look now.

The work is still the work. You just don’t have to be the only one doing it.

— Caroline

P.S. If you’re into building with AI, here’s how I built myself an AI coach a while back. Same basic idea as this project, just simpler.

In 1990, after years of practicing medicine and reviewing Van Gogh’s case history via his hundreds of letters, Arenberg published a paper in JAMA diagnosing Van Gogh as suffering not from epilepsy, as the artist’s physician claimed a century earlier, but from Ménière’s disease, an inner-ear affliction that can cause vertigo, of which Van Gogh complained, and tinnitus, a persistent ringing in the ears. Ménière’s, to Arenberg, could better explain Van Gogh’s decision to slice off his ear. After retiring, in 2017, Arenberg recommitted himself to studying Van Gogh and became convinced that art historians had made an even more alarming mistake: Van Gogh had not committed suicide. He’d been murdered.

Read the article for free on Air Mail, a lively digital read for the world citizen, with stories both foreign and domestic that you won’t find anywhere else, written by some of the world’s finest journalists.

Read the Article

Forwarded this email? Subscribe — it's free →