AI Influencers

Today there's one video. Thirteen seconds of it, made in May. It's the only thing I've got here that you could post. The folder meant to fill up with drafts is empty, and the first measure I wrote down, drafts per week, still sits at zero. My scheduler was supposed to leave a calendar behind. That calendar isn't there.

I'd meant it to make short videos for two personas I'd defined, and queue them up to post. My premise was simple: automate the fetching, keep the judgement. Fetching is most of the work in a short video: you find a format that's already working, you write to it, you find backgrounds, you render a voice, and then you stitch the pieces into one file. Judgement is which video I make and whether it's any good, and I built the fetching half and kept the other half for myself.

The fetching half runs. It doesn't check itself. That's the whole story of those thirteen seconds.

Walk my one file backwards. The narration behind it runs about twenty-six seconds and the background clip it got laid onto runs ten, so the video ends mid-sentence. About sixteen seconds went nowhere. My run reported success anyway, and I only found that out months later by measuring the two files against each other.

The script was too long because nothing made it short. My code works out a word count and asks for it. Only advice. The script came back nearly half over. Nothing in there reads what comes back and measures it.

Further back sits the step that's supposed to find a format that's already working. What did it find? On one run the model told me it couldn't see the thing I'd pointed it at, then offered up stock lines about gel nail polish anyway. My code scanned for the one line it expected, found it, and saved five clichés about gel nails as research. They read exactly like real findings and they carry no mark. I used to say on this page that the same step studies several frames of whatever I point it at, and it has only ever looked at one still image.

Then my faces and my voices. One persona's picture is the same file as the reference image sitting next to it. Wrong shape too. My other persona's voice reference is that same persona's recording. Same file, again. I'd written her twelve rules about how she talks, and nothing in my pipeline reads more than two details out of that file. The rules didn't matter.

Every one of these is the same move. A silent pass. Something arrived degraded, my machine took it, and it filed the result with the confidence it would've given a good one.

Not all of it is stuck here. Murmur measures how I write and speak, and the emphasis and pauses it records survive into the synthetic narration. The Distribution Engine took its machinery from this pipeline. It didn't build its own.

My machine still runs. What it doesn't do is stop at each handoff and ask whether the thing that turned up is the thing it asked for. So it sits where it sits: one video, an empty drafts folder, a counter at zero.