@hocyzon: Zhang Tao @hidecloud mentioned in a talk at Peking University that when video generation models like Seedance bring the latency between entering a prompt and receiving visual feedback below one second, they could become the next generation of OS. I’ve been thinking about this over the past couple of days. When many people hear this, their first reaction is that video generation will get faster, Hongguo short dramas will update more quickly…

X AI KOLs Timeline News

Summary

In a talk at Peking University, Manus founder Zhang Tao suggested that when the latency of video generation models such as Seedance drops below one second, video generation will shift from one-way production to two-way, real-time interaction, potentially becoming a new generation of operating systems and reshaping the form of products in industries such as gaming and software.

During a talk at Peking University, Zhang Tao @hidecloud mentioned that when the latency between entering a prompt and receiving visual feedback from video generation models like Seedance falls below one second, they could become a new generation of OS. I've been thinking about this over the past couple of days. When many people hear this, their first reaction is that video generation will get faster and short dramas on Hongguo will update more quickly. That's still thinking within today's approach to video production. During the Q&A afterward, Zhang Tao explained that many people used to see video generation as something a few people produce and most people consume, so faster generation wouldn't bring a corresponding increase in demand. But the key is the mode of interaction. Once latency falls below one second, video generation is no longer a one-way process of entering a prompt, waiting for a result, watching the content, and generating it again. Instead, it becomes a two-way interaction with continuous control and immediate feedback. He gave a fascinating example: in the future, a livestreamer might propose a set of game rules and a setting, such as playing GTA in Chongqing. Every viewer in the livestream could enter that world and use a controller to control their own character. Someone else who finds the gameplay interesting could keep modifying the rules, switch the setting to Wuhan or Xi'an, and create their own version. At that point, the way content is produced changes. In the past, one person made content and others watched; in the future, everyone could enter the content, interact with it, and even create new ways to play. What a creator produces might no longer be a video, but an experience that others can enter, modify, and build upon. I think gaming may be the field where breakthroughs come most readily. Traditional games require professional teams to build maps, levels, and environments in advance, with the supply of content constrained by development costs. If future models can generate environments in real time based on user instructions, creators could focus more on gameplay while the models handle the actual environments. A gameplay concept could be repeatedly copied, modified, and shared, giving ordinary users the chance to become creators of game content too. This will also affect the software industry. Zhang Tao also mentioned an idea that had made a deep impression on him: in the future, everything may gradually become visual. Humans ultimately receive information through vision—even when a language model generates text, we still need to read it with our eyes. So why must content be generated through a language model before people consume it? Could visual models handle the entire process directly in the future? He even said, "Software is gonna die. Games are gonna die." If we extend this prediction to software products, today, when we open an application, we have to adapt to interfaces and workflows that developers have designed in advance. In the future, users might instead encounter a dynamically generated visual environment, interacting continuously with a model through natural actions. Fields such as gaming, education, and design could all see product forms completely different from those we know today.
Original Article
View Cached Full Text

Cached at: 11/21/26, 10:28 PM

Zhang Tao @hidecloud mentioned in a talk at Peking University that when video generation models like Seedance bring the latency between a prompt and visual feedback below one second, they could become a new generation of OS.

I’ve been thinking about this over the past couple of days.

When many people hear this, their first reaction is that video generation will get faster and short dramas on Hongguo will be updated more quickly. That still approaches the question through today’s model of video production.

During the Q&A afterward, Zhang Tao explained that many people previously assumed video generation was simply a matter of a few people producing and the majority consuming, so increasing speed wouldn’t bring a corresponding increase in demand. But the key is the mode of interaction. Once latency drops below one second, video generation is no longer a one-way process of entering a prompt, waiting for the result, watching the content, and regenerating it. Instead, it becomes a two-way interaction with continuous control and immediate feedback.

He gave a very interesting example: in the future, a livestreamer might propose a set of game rules and a setting, such as playing GTA in Chongqing. Every viewer in the livestream could enter that world and use a controller to control their own character. Someone else who finds the gameplay interesting could further modify the rules, switch the setting to Wuhan or Xi’an, and create their own version.

At that point, the way content is produced changes. In the past, one person created content and others watched; in the future, everyone could enter the content, interact with it, and even go on to create new ways to play. What creators produce might no longer be a video, but an experience that others can enter, modify, and build upon.

I think gaming may be the field most likely to see a breakthrough. Traditional games require professional teams to build maps, levels, and environments in advance, with the supply of content constrained by development costs. If future models can generate environments in real time based on user instructions, creators could focus more on gameplay while leaving the specific environments to the model. A gameplay concept could be continually copied, modified, and shared, giving ordinary users the chance to become creators of game content as well.

This will also affect the software industry.

Zhang Tao also mentioned an idea that had left a deep impression on him: in the future, everything may gradually become visual. Humans ultimately receive information through vision—even when a language model generates text, we still need to read it with our eyes. So why must content be generated by a language model and then passed on to humans to consume? Could a visual model handle the entire process directly in the future?

He even said, “Software gonna die, Game gonna die.”

If we extend that prediction to software products, today, when we open an application, we have to adapt to the interface and workflows its developers designed in advance. In the future, users might instead face a dynamically generated visual environment, continuously interacting with the model through natural actions. Gaming, education, design, and other fields could all see entirely new forms of products.

Hallucination Engineer (@hocyzon): This is from Manus founder Zhang Tao’s talk at Peking University yesterday. I put this together based on a transcript a friend sent me:

Similar Articles

@XAMTO_AI: ControlNet author Min Shen has come up with something new! The newly open-sourced FramePack directly lowers the barrier for video generation — runs on just 6GB VRAM, generates a 1-minute 30fps video with a 13B model, and on an RTX 4090 it takes only 1.5 seconds per frame. Such configuration requirements were unimaginable before. The core idea is frame-by-frame…

X AI KOLs Timeline

ControlNet author Min Shen has open-sourced the FramePack video generation model, which requires only 6GB of VRAM to run a 13B model, generates a 1-minute 30fps video, takes 1.5 seconds per frame on an RTX 4090, and comes with a one-click Windows package.

@Saccc_c: Learned from an open-source video prompt by a certain AIGC blogger (Fang Demu) on TikTok, and directly produced the following chase scene video using Seedance 2.5. Indeed, the more specific the description of the visuals and camera movements, the more stably AI can deliver quality output. (Specific prompt sent to the comments)

X AI KOLs Timeline

By referencing an open-source video prompt from AIGC blogger Fang Demu on TikTok, the user generated a chase scene video in one go using Seedance 2.5, proving that more specific descriptions can enhance the stability and commercial-grade quality of AI video generation.

@Zesee: https://x.com/Zesee/status/2072417062165635433

X AI KOLs Timeline

This article introduces how to combine the Codex AI programming agent and the Remotion video rendering library to code and modularize the video production process, enabling daily updates for self-media creators. The author shares a complete methodology from script structure and template design to daily workflow, helping creators free themselves from repetitive labor and build a video production pipeline.

@sukie234: As a small player in the downstream ecosystem of the Seedance model, many of my clients are in the Chinese short drama and advertising industry. I truly believe Seedance has unlimited potential. Currently, I spend most of my time researching how to best leverage the Seedance tokens I hold, and I have already developed four products…

X AI KOLs Following

The author, as a downstream developer of the Seedance model, shares their views on the model's potential. They have already developed four product lines based on Seedance tokens and believe ByteDance will win the video generation race with its Hongguo short dramas + Seedance closed loop.

@joshesye: https://x.com/joshesye/status/2052599953193566214

X AI KOLs Timeline

A detailed tutorial introducing four methods for maintaining character consistency and plot coherence when creating AI short dramas using Seedance 2.0 and GPT-image2, including extending reference videos, using keyframes as the first frame, compositing multiple video segments, and converting storyboards to video.