The category is moving fast enough to break its own courses
3 min liveBefore anything else, a warning about the ground we are standing on. Most material teaching this subject is teaching a tool landscape that has already changed, and the clearest example is sitting 29 days away from us.
OpenAI has left video. The Sora consumer app shut on 26 April 2026, and the Sora 2 API sunsets on 24 September 2026. OpenAI's own deprecations page lists every Sora model and the Videos API itself, and the column headed "recommended replacement" contains a dash. There is no successor. No reason was given. The only public statement was "We're saying goodbye to the Sora app."
Every article, tutorial and course that teaches "Sora, Veo and Runway" as the three pillars of this field is already wrong. That is not a footnote about one product. It is the reason this course teaches judgment rather than button positions, because the buttons have a half-life of about a year and the judgment does not.
LiveTwo more that break the standard outline2 min▶
If you go looking for the obvious next thing after Veo 3, you will not find it.
- There is no Veo 4. Google broke its own naming sequence and launched Gemini Omni instead, described as a model that "can create anything from any input, starting with video". Any roadmap slide predicting Veo 4 is predicting something that does not exist.
- Every Google video model is still marked Preview, including Veo 3.1 and Gemini Omni Flash. No stable video model is listed at all. Preview is not a formality: Veo 2 and Veo 3 were announced deprecated on 15 June 2026 and shut down on 30 June, fifteen days later.
So the working assumption for the whole course is this. Anything you learn about a specific model is perishable. Anything you learn about how attention behaves is not, and that is what tonight is about.
The ladder, and which rung is free
8 min liveBelow is the shape of the thing you are going to build tonight. Five changes to the same sixty seconds of footage. Read the size of the first jump before anything else.
Look at the gap between the first two bars. The largest improvement available to you costs nothing, needs no software, and is available to somebody with a phone. It is simply saying what the video is about before you say hello. Every rung above it is smaller, and every rung above it needs a tool.
That ordering is the argument of this entire course. The tools are genuinely useful, and sessions 2 and 3 go deep into what they can and cannot do. But if you spend your budget on generation and skip the first three seconds, you have bought the small lever and skipped the big one.
Self-studyWhat the numbers on this page are, exactly4 min▶
Worth being precise, because most content in this field is not.
The silence detection is real. The playground below holds a sixty-second audio envelope. When you press detect, it genuinely scans that envelope for runs below 0.085 amplitude lasting at least 0.4 seconds, and reports what it finds. The four gaps and the 12.4 seconds are computed at the moment you click, not stored. So is the 47.6 second runtime, which is arithmetic on the cut points it found.
The retention curve is a model. It is a teaching model of how attention behaves, built to be directionally honest, and it is not a measurement of your audience or anybody else's. The page tells you that on the widget itself rather than burying it here.
The distinction matters more than it sounds. A course that presents a model as a measurement is doing the thing this whole field does badly, and you should hold the rest of the material to the same line. When session 5 gives you physics benchmark scores, those are measurements from published papers. When it gives you an iteration cost, you will get an explicit warning that no independent audit of that number exists.
You are a bad judge of your own pauses
6 min liveEvery speaker leaves gaps. Almost nobody hears their own, because when you were recording you were thinking, and thinking feels like a moment rather than four seconds. A detector does not have that problem.
Cutting these is not an editorial decision and it costs you nothing. You said the same words. The viewer receives them in 47.6 seconds instead of 60, and the version they receive feels like it was made by somebody who valued their time. That perception is doing more work than the twelve seconds.
This is also the one place in the whole course where the tools are unambiguously worth it. Finding these by hand means scrubbing a timeline. Every serious editor now does it automatically, and session 3 covers what each one will and will not do, including the constraints their marketing pages do not mention.
Self-studyThe constraint nobody puts on the feature page3 min▶
A preview of session 3, because it changes whether you should rely on this.
Descript's filler-word removal, the best-known version of this feature, "currently detects filler words in English transcripts only". That is from its own documentation. If you work in any other language, the feature does not exist for you, and nothing on the marketing page says so.
There is a second one that catches people out. The option called "avoid harsh cuts" deliberately skips filler words that "can't be removed without clipping into nearby words or leaving behind an awkward cut". This is good behaviour and it is also why the number of fillers removed will never match the number detected. People report this as a bug. It is the feature working correctly.
The pattern generalises, and it is the spine of session 3: across every vendor in this category, the advertised feature is never the real constraint. The real constraint is English-only, or single-face-only, or two-seconds-maximum, and it is always three clicks deeper than the page that sold it to you.
Three seconds is the entire budget
5 min liveNot the opening. The budget. Whatever occupies the first three seconds is what you spent your one guaranteed piece of attention on, and most video spends it on throat-clearing.
The objection people raise here is that it feels abrupt, and it does, for about a week. What it actually is is respectful: it assumes the viewer has somewhere else to be, which they do. The greeting is not deleted, it is demoted to second four, where it is heard by an audience that has already decided to stay.
Self-studyWhy this is harder than it looks3 min▶
The mechanical part takes ten minutes. The hard part is that you have to know what your video is about in one sentence, and a surprising number of videos do not.
If you cannot write that sentence, the problem was never the edit. The video does not have a point yet, and no amount of generation, captioning or cropping will give it one. Every rung above the first on that ladder is amplification, and amplification of nothing is still nothing.
This is also why the trap lever exists in the simulator. "It feels thin, add more" is the response of somebody who has sensed the missing sentence and reached for volume instead. The number goes down because the viewer can tell.
Climb the ladder, then break it
★ 18 min · everyone buildsThe detector below runs for real on a real envelope. The retention curve is a teaching model and says so on the widget. Work through the steps in order, and spend the most time on the last two.
Press detect before touching anything else. Watch it scan the waveform and mark four regions. It reports 12.4 seconds across four gaps. That is computed from the envelope at the moment you click, so change the threshold and the answer genuinely changes with it.
Turn on "say the point in the first three seconds" alone. Retention goes from 16 to 38 percent. Sit with the size of that. It is bigger than the next three levers combined, and it is the only one that needs no tool, no subscription and no skill.
Add "cut the dead air". Retention reaches 52 percent and the runtime readout drops from 60 to 47.6 seconds, because the cut points the detector actually found are being applied.
Add captions, then the crop. 64, then 72 percent. Both are real production choices with a real cost in time, and both are worth less than the free one you did first. That ordering is the whole lesson.
Now, from 72, switch on "add more, it feels thin". It falls to 56. This is the instinct that produces most bad video: the piece feels slight, so you pad it. The padding is what people leave during.
Finally, from 72, switch on the branded intro animation. It falls to 59. Nobody in the building will tell you this, because the intro was expensive and somebody signed it off. It spends the only three seconds you are guaranteed on something the viewer did not come for.
Why both traps are in the simulator rather than in a bullet list. You will be asked for both of these by name, usually by somebody senior, and usually with a good reason attached. Having watched the number go down yourself is a different kind of knowledge from having read that it goes down. Take the screenshot.
Before session 2
30 min- Find a video you published and watch the first three seconds only. Then stop it. Write down what a stranger now knows about what the video is for. If the answer is nothing, you have found your biggest available gain.
- Re-cut those three seconds and nothing else. No new footage. Move the sentence that says the point to the front, and delete whatever was there. This is the exercise, and it takes about ten minutes.
- Run any filler-and-silence tool over the same video. Note two numbers: how much dead air it found, and how much of it you had noticed. The gap between those is the point.
- Check one claim from tonight yourself. Open OpenAI's deprecations page and find the Sora rows. Confirm the September date and the empty replacement column with your own eyes. Getting into that habit is worth more than any single fact in this course.
- Optional, and it will annoy you: find any course or article published before mid-2026 that teaches AI video, and count how many of its named tools still exist in the form described.
Three questions
Clear all three and session 1 gets stamped in your passport.
1 · Of the five changes in the ladder, which produced the largest single gain, and what did it cost?
16 to 38 percent, larger than the next three levers combined. Captions and cropping are both real and both worth doing. They are worth less than the free one, which is why the ordering matters more than the list.
2 · A colleague wants a five-second branded animation at the top of every video. What does the simulator show, and why?
The three seconds are not free real estate. They are the entire budget of attention you are given before the decision to stay or leave is made. Anything occupying them has to be the reason someone stays.
3 · Why does this course open with the Sora retirement rather than with a tool tutorial?
No reason for the shutdown was ever given, so B and C are both speculation. What is verifiable is the pattern: the app closed in April, the API closes in September, the replacement column is empty, and Google shut two Veo models with fifteen days' notice. Perishability is the lesson.
What this session teaches, and where it came from
Every fact on this page, and its verification tier, is recorded in the course's source map., including a quarantine list of what is deliberately not stated.
Session 1 cheat sheet · pin this
Tonight was about attention, which does not change when the tools do. Session 2 is about the tools themselves: what Veo, Runway, ElevenLabs and Descript actually produce, what each second genuinely costs, and the clip-length and resolution limits that decide what you can make before you have written a word.