Daybook

Most of this month has been learning to make video with AI

I want to be straight about what this is. It is mostly study, not product. Over the last two weeks I have generated a few thousand clips, and most of them are wrong in some specific, instructive way: the diver's face changes between two shots that are supposed to be the same person, the air hose is yellow in one frame and black in the next, a camera move I asked to sink rises instead. Nothing is finished. Three films exist as rough cuts and none of them is something I would call done.

But some of it genuinely works, and it seems more honest to show that than to wait for a finished thing. Here is 35 seconds cut from the parts that hold up.

The thing I did not expect: the strongest material is not where the model invents the most. It is where I constrain it hardest. The physics panels are the cleanest images in the whole set, because a computer drawing a diagram is doing something it cannot be wrong about. The underwater shots are best when there is one subject doing one thing. The moment I ask for a diver and rising bubbles and swaying coral and moving light, it falls apart — the motion budget gets split and nothing gets enough of it.

What one instruction actually produces

This is the raw material, unedited: ten seconds, generated from a single instruction against one approved still. Nothing has been cut, graded or mixed. The sound came with it — the tool returns a synced audio track in the same call, which genuinely surprised me the first time. Turn it up.

Where I actually am with this

About a week of real work, and I want to be exact about what is and isn't automatic, because "AI made a video" hides all the interesting detail.

• The tool makes six to ten seconds at a time. That's the ceiling. The two-minute film above is dozens of those clips generated separately and stitched together — the stitching, the timing and the order are mine.
• The sound it generates is atmosphere, not music. It's texture that matches the picture — water, movement — and for a finished film I throw it away and build a proper mix underneath.
• The music isn't AI at all. It's a script: about a hundred lines that synthesise an ambient pad mathematically, no model involved, because the music models worth using won't run on my laptop. Below is eighteen seconds of it.
• The narration is a separate step with a separate tool, added after the picture is locked.
• Nothing chooses what's worth teaching. That's the whole job, and none of it is automated.

The voice, and the score

Four synthetic voices reading the same line from the scuba script. I picked the last one for the current cuts — it was noticeably better than the free local model I started with, and far better than the robotic placeholder before that. Whether the finished films use a synthetic voice or my own is still undecided.

Four things I have learned that I would not have guessed:

• Telling the model what to leave out does not work, and often summons it. Asking for footage with no branding put an actual National Geographic logo on a wetsuit. You have to describe what should be in frame, never what shouldn't.
• Camera specifications are mostly superstition. I ran the same image with the focal length, aperture and ISO changed and could not see a difference — and a camera name I invented on the spot scored the same as a real one. Film stock names are the exception; those genuinely change the picture.
• Curating beats prompting. Generating three or four versions and picking one beats any amount of rewriting the instruction. The published research on this puts human-curated output at roughly thirty points more realistic, and that matches what I see.
• Naming the thing that must not change is the whole trick. Every kind of drift I hit this month — wrong mask, wrong colour, wrong crop — was fixed by saying the invariant out loud in the instruction rather than assuming it would carry over.

The most useful failure was one that passed every automated check I had. A clip rendered at the right resolution, the right codec, decoded without a single error — and roughly eighty percent of it was black frames, because the instruction ended with the subject leaving the frame and the model obliged. Nothing mechanical caught it. The only signal was that the file's bitrate was twenty-five times lower than its siblings. That is now an automatic check. It is a good reminder that a file passing its tests and a film being any good are entirely different questions.

All of it is AI-made and labelled as such wherever it appears. The scuba material is educational only — it teaches ideas, it does not certify anyone to dive.