Original video in Chinese.
From now on, I’ll hand all my videos over to AI for editing.
Actually, videos like mine, where I talk to the camera, are a natural fit for AI editing. Because I’m basically just reading from a script. Since I can read from a script, AI can also edit from a script.
I roughly divide the editing work for my talking-head videos into two parts: A Roll and B Roll. Let’s start with A Roll.
A Roll is the part where I’m speaking to the camera. During recording, even if there’s a script, there will still be times when I read something wrong or don’t like how it sounds. So the key to A Roll editing is to delete the incorrect parts and keep only the correct ones. And you have to leave natural pauses between the clips, otherwise it will sound very awkward.
To get AI to edit A Roll, I specifically made a Skill: newtype Cut. Its working principle can be summed up in one sentence:
First understand the video, then do a rough cut based on the content.
Under newtype Cut’s guidance, the Agent first uses speech recognition to convert the entire video into text with time coordinates. That way, it not only knows what you said, but also where each sentence appears in the video.
Next, it compares the recognized content sentence by sentence with the script to determine which sentence each segment of the talking corresponds to.
For example, when I’m filming, I might say one sentence wrong three times, and only get it right on the fourth try. In that case, newtype Cut will determine that the first few attempts were incomplete tries, and the final one is the formal version that should be kept.
After finding the content that needs to be kept, it converts those decisions back into video cut points, while leaving a bit of breathing room before and after each sentence, which is the natural pause.
After the initial cut, it also reassembles the retained audio and checks it again. Once it determines there are no problems, it generates an XML file. After importing it into DaVinci, it becomes a timeline.
So the value of newtype Cut is that it can guide the Agent to complete the most time-consuming part for me: listening to the material, finding repetitions, selecting complete expressions, and organizing the raw footage into a timeline that can then be further refined.
In the past, A Roll editing would take me at least an hour and a half. Now all of that time is completely saved. It wasn’t a waste either to spend so many Tokens making and repeatedly optimizing this Skill.
I’ve already put newtype Cut on GitHub, and the link has been posted in Knowledge Planet. You can grab it yourselves. Next is B Roll.
B Roll is the part used to support the expression of A Roll. In my videos, B Roll comes in two types:
One is screen recording. I often introduce AI tools, so I use OBS a lot to record the screen, and then overlay the recorded clips on top of A Roll.
The other is animation effects. When I’m talking about certain abstract concepts, I need animation effects to help people understand them.
For screen recording, I don’t need AI to do it for me. Because I recorded everything myself, only I know which clip corresponds to which content, and the operation is very simple. What I really need AI to do is generate animation effects based on what I’m talking about. I use the Remotion plugin to do this.
First, I give the Agent the entire video script for analysis, and let it pick out the parts suitable for turning into animation effects.
Next, based on the selection results, I let it call the Remotion plugin to generate the corresponding animation effects. Some of these animation effects are full-screen, and some have transparent backgrounds, so they can be displayed together with me.
Finally, the Agent exports the animation effect for each clip as an MP4 file. I just overlay them in the corresponding positions in A Roll.
To get this done before, I had to go into editing software to find suitable effect plugins and then tweak parameters and so on. Now, just letting AI understand the semantics and generate MP4s directly really saves time.
And I believe this will also become the norm for content in the AI era: generated instantly, used once, and discarded.
At this point, AI has basically taken over 80% of my editing work. But the work before editing—that is, script writing and video shooting—I won’t let AI do for me.
From a technical perspective, this isn’t hard at all. There’s no difficulty in letting AI mass-generate content in my style, and then having a digital human that looks like me do the talking in front of the camera.
But the reason I absolutely won’t do that is: the more things AI generates, the greater the value of real people becomes.
I previously posted an article in Knowledge Planet called “The Authenticity Economy.” I think:
In the old era, “real” was the default, and “creation” was scarce. We paid for good ideas and polished production.
In the new era, which is today’s AIGC era, “creation” is abundant, and “real” is scarce. We will be more willing to pay for content that can be proven to be “real human original,” “real human experience,” and “real human emotion.”
Actually, today there are already a lot, a lot of self-media accounts making videos entirely with AI. And public account articles are also a major disaster area. If you look at those so-called ten-thousand-word long reads, many of them are written by AI, going back and forth saying the same circular stuff, with nothing in them at all, a complete waste of time.
So sooner or later, we’ll reach a critical point—when everyone has had enough. That’s when the authenticity economy will take off.
Finally, a reminder for everyone: what AI is good at and what it isn’t good at is one dividing line; what AI should do and what it shouldn’t do is another dividing line. Don’t just stare at the first one.
OK, that’s all for this episode. If you want to learn about AI, want to reclaim individual sovereignty, and want to find like-minded people, come join our newtype community. See you next time!