How We Built the AI Trailer: What Worked, What Broke, and What We Learned
On this page
Congratulations, AGI Is Almost Here is a fictional, satirical trailer about Walter Sterling, the Bureau of Artificial Intelligence Preparedness and Domestic Continuity, and the Robinson family preparing for several possible futures. It is also a record of how the trailer was actually made: with AI-assisted generation, human editorial decisions, and a great deal of verification between the two.
From Lab posts to a video experiment
This trailer also served as a proof of concept for turning the Lab’s writing about artificial intelligence and artificial general intelligence (AGI) into a visual, narrative format. The goal was to make the research easier to experience as a story while testing a repeatable production workflow for a larger video series.
For now, the rest of the main feature film is paused. Video generation, continuity, audio, and editorial finishing require more time than the current production window allows, so this trailer stands as a completed first experiment and a record of what the larger project would require when it resumes.
What we were trying to make
The goal was not to make a collection of impressive AI clips. It was to make a coherent trailer with a distinct voice, recurring characters, readable props, deliberate pacing, and a clear handoff from the Bureau’s broadcast world to the Robinsons’ domestic story. The finished piece needed to feel like a short film rather than a technology demonstration.
What went smoothly
- The script became a real production control. The approved trailer wording was locked before the final finishing pass. That gave every later decision a reference point and prevented attractive generated material from quietly changing the story.
- Voice and picture were separated. ElevenLabs supplied the controlled narration, while DaVinci Resolve became the authority for timing, pacing, and the final mix. This was much more dependable than asking a video model to solve picture, dialogue, and editorial rhythm at the same time.
- Reference images made continuity possible. Character portraits, Bureau graphics, props, and location references created a visual vocabulary that could be reused across short generated clips.
- Technical checks caught problems early. Frame rate, frame count, duration, audio format, checksums, and duplicate candidates were inspected before assets were treated as production-ready.
- The archive is now findable. The source material, audio, Flow generations, Resolve media, review exports, and control documents were consolidated on the external AGI Video Assets drive without replacing the canonical project repository.
What required intervention
AI video generation was useful, but it was not a complete production department. Short clips often needed to be treated as coverage rather than finished shots. We compared alternate generations, preserved continuity frames, re-downloaded source media when needed, and chose the usable moments in Resolve.
The generated audio also needed a clear boundary. Flow could return useful environmental sound, but its speech and timing were not the canonical soundtrack. Generated audio was removed or retained selectively, while the approved narration and dialogue remained separate and editable in Resolve.
Text inside generated footage was another recurring weakness. Forecast papers, labels, and Bureau language were made legible with controlled stills, overlays, and editorial graphics rather than trusting a model to render exact copy consistently.
What did not work the first time
- Character continuity was not automatic. Regenerating a prominent Walter shot encountered a platform identity restriction. The practical answer was to preserve the approved continuity material, use compatible coverage, and avoid pretending that every generation belonged in the same visual sequence.
- Some early edits felt faster than intended. The issue was not a hidden Resolve retime. It came from timing, source audio, and shot transitions. Rebuilding the narration as independently movable sentence-level clips made the pacing controllable.
- Generated props needed editorial repair. The forecast-paper insert went through several iterations before the wording, orientation, and hold duration were readable. The final solution used a controlled graphic and a measured hold.
- Browser production surfaces were part of the workflow risk. Downloads, platform state, upload state, and credit or generation state all had to be verified independently. A visible form was never treated as proof that a generation or publication had completed.
What the production taught us
The strongest workflow was hybrid: use generative tools for exploration and coverage, use ElevenLabs when the spoken performance must be controlled, and use Resolve for continuity, timing, text, sound, and final authority. The human role was not removed. It moved upward into deciding what the film meant, what counted as evidence of a finished shot, and which imperfections belonged in the final cut.
The most important lesson was simple: an AI-generated clip is a proposal, not a finished scene. A production becomes dependable when every asset has a known source, every spoken line has an owner-approved version, every technical claim is measured, and the final edit can still be changed without rebuilding the entire project.
The finished trailer
The final trailer export is a 137-second, 1920×1080 H.264/AAC master. It is a work of fiction and an exploratory creative project, not a forecast or a statement of fact.
The trailer is embedded below so you can watch it without leaving this production record.
Open Congratulations, AGI Is Almost Here | Official Trailer directly on YouTube.
Built with Google Flow, ElevenLabs, DaVinci Resolve, and a human-directed production process.
