How an AI Actually Decides Which 8 Seconds of Your Footage Are the Good Part

Here is how it works, in plain terms. An AI finds the best moments in footage by measuring what is actually on screen — how much motion there is, how the sound rises and falls, where the beat lands, and how well the shot is framed. It scores each moment on those signals, ranks them, and hands you the strongest ones first. It does not read a script or guess at what you meant. Sisyphos does this on your Mac and explains every pick in plain words, so you can see why a clip earned its place — or why it didn't.

How does an AI find the best moments in footage?

An AI finds the best moments in footage by turning what happens on screen into numbers it can compare. It splits the footage into shots, then measures each one: how much motion there is, how the sound rises and falls, where the beat sits, whether the frame is sharp and well-composed. Those measurements — not the words in a transcript — are what it scores. The strongest moments rise to the top of a ranked list; the weak ones fall out, each with a reason attached. Sisyphos is a local-first desktop app that watches your raw footage, finds the best moments, and explains every pick by what's on screen — motion, sound, beat, and framing. Nothing leaves your Mac. What you get back is a first pass at your selects that you can read, question, and change before it ever reaches your video editing timeline.

How does AI video editing work, step by step?

Most AI tools you have seen work on speech. They read a transcript and cut around the words. That is one kind of AI video editing, and it is fine for a podcast or a talking-head video. It falls apart the moment nobody is talking — a first dance, a downhill run, a drone pass over a ridge with no dialogue in it at all.

Finding the best moments in footage that has no speech takes a different approach. Here is the shape of it.

1. It breaks your footage into shots. Every time the camera cuts or the scene changes, that is a new shot to look at on its own.
2. It watches each shot and measures it. Not the words — the picture and the sound. Motion, sound energy, beat, sharpness, framing.
3. It notices the repeats. When two clips are near-duplicates — the same moment from nearly the same angle, a few seconds apart — it flags the weaker one, so your list is not padded with the same shot twice.
4. It rates the moments and writes down why. This is the AI judge — it rates them and writes down why, in plain words.
5. It hands you a ranked list. The strongest picks sit at the top. Everything it set aside is still there too, with the reason it lost.

None of those steps needs a single spoken word. That is the whole point.

What does the AI actually measure?

Four kinds of thing you would notice yourself, if you had the hours to scrub every clip by hand:

• Motion. A still tripod shot and a fast pan carry different weight. A sudden surge of movement — a jump, an impact, a run breaking into the open — reads as a moment worth a look.
• Sound. Not the words, the shape of the audio. A cheer, a laugh, a swell in the music, a sharp rise after a quiet stretch.
• Beat. Where the rhythm lands, so a cut can sit on a real beat instead of an arbitrary spot on the timeline.
• Sharpness and framing. Whether the shot is in focus, well exposed, and composed — not a blurry throwaway between two good takes.

Each of those is a number the tool can measure and compare. Put them together and you have a picture of where the good part of a clip actually is — which is how a two-minute clip gets narrowed to the eight seconds worth keeping.

Does the AI read the transcript?

Only to help you search — never to make the cut. Speech, if there is any, is a search convenience only — a way to jump to a clip, never the reason it's chosen. The cut always comes from what's measured on screen, not from words.

But the transcript is never the basis for the cut. The cut comes from the signals on screen. Speech-based editors do it the other way around: no words, no moments, nothing to work with. Measure what is on screen instead, and footage where nobody speaks — sport, weddings, B-roll, a drone pass — becomes something an AI can actually rank.

What does "it explains every pick" mean?

This is where a tool that finds moments splits from a black box that just spits out a cut. Every pick comes with a reason. Every reject, too.

Sisyphos keeps a record of every clip it considered: whether it made the cut, where it ranked, and — when it was dropped — the reason in plain words. A clip falls out because it is "too similar to a better take," or because there was "too little happening." Those are causes an editor recognizes, not codes. Ask "why isn't this moment in my cut?" and you get an answer, not a shrug.

Say you shot a two-minute clip of a mountain-bike run. The middle drags — the rider is just coasting. Then there is a jump, a hard landing, and a whoop from the person filming. The tool measures the motion spike, the impact, the sound, and marks the eight seconds around the landing as the keeper. It trims toward that best second, never just cropping to the middle of the clip. And it tells you that is why — so if you disagree, you can override it. This is the difference between an AI you can question and an auto-editor that decides for you.

Where this fits in your edit

Finding the moments is the culling pass — the unpaid first half of the job, where you scrub hours of raw footage looking for the parts worth using. That is the part an AI can take off your hands. It is not the same as the edit itself, and it helps to keep the two ideas separate: the tool proposes the selects, you make the edit.

Once you have approved the ranked list, the selects travel to your editor. Sisyphos pushes them into a running DaVinci Resolve timeline with one click, or exports standard FCPXML and EDL (CMX3600), so the selects reach Premiere, Final Cut, or anything that reads them. What lands is a rough cut you can open and change, with the reasons riding along — so a week later you can still see why each clip earned its place.

It runs on your Mac, and only a few still frames plus the measurements go to the AI judge, under your own key, at cents per clip with a hard cap. One license is €129 at launch, perpetual, with 12 months of updates.

Sisyphos is in development. The way it finds and explains the best moments — on the record, in plain words — is what it is being built to do. Join the waitlist and you hear it first when it ships.

You stay the editor.