🎬 My partner asked me a couple of weeks back, which of the 12 AI-generated cuts was “the right one”. I didn’t have a formula for that. I just knew 😀
That moment stuck with me more than I expected. Tried few POC/Research around it and came across a story about Jensen Huang being asked who the smartest person he’d ever met was. He refused to answer — said the question itself was outdated.
I get that now, in a very unglamorous way.
For 20 years, my job was basically to know things — kernel drivers, DRM, distributed systems, things most people never think about. That depth was the whole value proposition.
Now I run an AI video platform, and I watch it generate more variations in a minute than a full production team used to make in a week. If I’m honest, there’s a small ego hit in that — the thing I spent two decades getting good at is exactly what’s getting automated first.
But sitting with clients and support team, I keep noticing the same gap: the model can nail lip-sync, match a brand’s color palette, hit every technical spec in the brief — and still produce a cut that’s wrong. Not broken. Wrong. The pacing feels a beat too fast for the product, the avatar’s tone reads slightly too corporate for a Gen-Z audience. None of that shows up as an error in any QA pipeline. It only shows up when you watch it.
That’s the part I couldn’t fully explain — I just knew it when I saw it, the same way you know a take is off before you can say why.
Models are getting excellent at optimizing for the spec. Knowing when the spec itself is missing the point — that’s still entirely on us.
Anyone else running into this in their own work?
The Spec Is Right, the Cut Is Wrong: Why AI Still Needs Taste
