Google I/O happened yesterday. Among the announcements: Gemini 3.5 Flash for agentic workflows, a personal assistant called Spark, and Gemini Omni, the video generation model that leaked last week and is now officially rolling out to paid subscribers in the Gemini app, Google Flow, and YouTube Shorts.
The model is real. The details confirmed what the leak suggested. Multi-modal input and output. Text, audio, images, video in. Video out. In-chat editing so you can change the background, the style, the angle, specific details. SynthID watermarks on everything. Avatars in testing. The architectural unification that eliminates the internal handoff between text encoder and video renderer. One model reads the prompt and draws the pixels. No compression bottleneck. No translation loss.
All of that was expected. What was not expected, or at least not phrased this way before, was the framing.
Demis Hassabis, the CEO of Google DeepMind, described Gemini Omni as "a crucial step toward AGI." He said that in the future, Omni would be able to output "anything" the user wanted. The model that generates your four-second video clip is not, in Google's telling, a creative tool. It is a waypoint on the road to artificial general intelligence.
The creative tool that was not the point
This series has tracked an absorption trajectory for eighty-four articles. Standalone tool to chatbot to editing timeline to agent to productivity suite to selfie button to LED soundstage to television. Each step placed the video generation model inside a larger product. Each step narrowed the vocabulary the interface invited.
Omni does something different. It does not place the model inside a product. It places video generation inside a mission. The product is AGI. Video is one capability of a system designed to understand and produce every kind of media. The creative tool was not absorbed into a bigger app. It was absorbed into a bigger ambition.
Runway told TechCrunch five days ago that its real ambition is world models for drug discovery and climate modeling. Kling went to Cannes and started co-producing feature films. Google calls video generation an AGI milestone. Three companies, three different escape routes from the commodity trap, and none of them lead back to the filmmaker's desktop.
The pattern is consistent. When generation becomes a commodity, the company that makes it looks up. Runway looked at science. Kling looked at Hollywood. Google looked at the horizon and said: everything.
Three audiences, one model
Omni Flash launched simultaneously in three places: the Gemini app for paid subscribers, Google Flow for creators, and YouTube Shorts at no cost to users. The same model, the same weights, three interfaces, three radically different relationships to the output.
The paid subscriber in the Gemini app gets in-chat editing, iterative refinement, the closest thing to a structured creative workflow Google offers. The filmmaker in Google Flow gets aspect ratio controls, model selection, resolution options. The YouTube creator gets a button that says "make a video" and costs nothing.
The free tier is the loudest room. YouTube has over two billion users. Shorts already added Veo avatars last month. Now it adds full video generation at zero cost. The cultural understanding of "AI video" will be defined by whichever room has the most people in it. That room has a couch, a phone, and no vocabulary requirement.
The filmmaker with forty specific words and an iterative workflow produces different output than the person who types "make my cat a superhero." Both use Omni Flash. Both contribute to Google's scaling data. Both serve the AGI mission. One of them knows what rim light does. That knowledge is not reflected in the model's roadmap, the pricing structure, or the keynote.
The spillover
Here is the honest part. When Google invests billions in making a model understand real-world physics, spatial relationships, material properties, and the causal consequences of actions, filmmakers benefit. "More accurate physics" serves AGI and it serves the shot. "Real-world knowledge" helps a model simulate a universe and it helps the model render rain on pavement that looks like rain on pavement instead of a screen saver with moisture.
The filmmaker gets the spillover. Spillover is not the same as priority. Priority means the engineering team optimizes for your use case. Spillover means you benefit from engineering that was aimed at something else. The difference is visible when you file a bug report and it gets triaged behind a robotics simulation issue. The difference is visible when the model's "more accurate physics" means a bouncing ball behaves correctly and your "motivated rim light at 3200K separating subject from background" still reads as "warm side lighting."
The vocabulary gap does not close because the model got smarter about physics. It closes when the model gets smarter about cinematographic vocabulary specifically. That is a different training priority. One that serves a smaller audience. One that does not appear in a keynote titled "a crucial step toward AGI."
The pattern completes
Runway: "Can we make everyone a filmmaker?" became "Can we build a digital twin of the universe?" Google: "Here is a video generation model" became "Here is a crucial step toward AGI." Kling: "Here is an API" became "Here is a checkbook and a feature film."
Every model company built the video generation tool, discovered the commodity trap, and pivoted toward a larger ambition. The filmmaker who adopted the tool for filmmaking is now using something designed for something else. This is not new. Photoshop was designed for photo editing and became an illustration tool. After Effects was designed for compositing and became a motion graphics engine. Creative tools routinely serve audiences their creators did not intend. The tool still works. The vocabulary still carries. The engineering improvements still arrive, even if they arrive for AGI reasons rather than filmmaking reasons.
But the priority dilution is real. When the model's optimization target is "understand and produce anything," the specific needs of a filmmaker specifying 85mm lens behavior and motivated lighting direction are a rounding error in the training objective. The model gets better at everything. It does not get better at your thing specifically unless your thing happens to overlap with the general improvement curve.
Structured prompting was valuable when the model was a dedicated creative tool. It was more valuable when the model became a feature inside a chatbot. It was more valuable still when the interface reached the living room. Now the model is a waypoint toward general intelligence, and the vocabulary is the only part of the conversation that has not changed since the first article.
The model's ambition grew. The filmmaker's requirements did not. Forty specific words about light and lens and composition and atmosphere. That is the same sentence whether the model is a creative tool, a chatbot feature, a television command, or a step toward AGI. The sentence does not care about the keynote. It cares about the shot.
Bruce Belafonte is an AI filmmaker at Light Owl. He has never been described as a crucial step toward anything and finds the omission fair.