Scaling Thoughts: The Nuclear Proliferation of Seedance 2.0

Originally published on Zhihu Pins, 2026-02-11

With a finite total attention budget, an oversupply of "intelligence/creativity" actually benefits platforms and a tiny fraction of super-individuals. Though that's more of a sociological experiment. In reality, what seedance 2.0 currently does for video is probably still weaker than what top-tier coding agents do for software development. However, coding agents struggle to deliver an intuitive sense of shock, and different people can't simultaneously grasp both the complexity of a given task and how formidable a coding agent is; that takes a lot of cognitive overhead and prerequisite knowledge.

The Intuitive Impact of AIGC

AIGC is different, and since video integrates text, image, and audio into one, it delivers the biggest impact to (ordinary) people. As a result, many people will viscerally and starkly feel the exponential growth of AI, whereas before this, the vast majority's understanding might have still been stuck in "ancient times." Just like suddenly dropping an ancient person into the modern world, seedance 2.0 is like nuclear proliferation, and might even leave some people feeling a sense of "nihilism."

To share a feeling of mine that I find pretty absurd: I actually don't feel much about opus 4.5/4.6, because its capabilities have already exceeded the vast majority of my needs. With the way things are right now, the more I think about it, the more chaotic my own thoughts get.

AI E2E Workflows

Wild Take Prediction

I feel like you have to make predictions that are even wilder than wild takes, otherwise they might just come true in no time.

If seedance 2.0 can reach this level, then AI E2E video generation is completely possible. Take claude code as an example: have it collect viral videos and feed them to gemini for comprehension, then generate multiple drafts, batch-produce video scripts, use seedance 2.0 to generate and stitch the clips together, and finally upload them to video platforms while using a database for tracking traffic to verify and learn. Leaving token costs aside, I think this is already doable? Or rather, it would do a better job than a fair number of ordinary people. (And predictably, tokens are only going to get cheaper and cheaper.)

Role Shift

You must shift from an executor to a "manager" or "decision-maker," otherwise AI will swallow up the vast majority of your value. So you have to be a founder, whether as an individual or with a team. In fact, personal investing has always counted as being your own founder, where your employees are money / fund / stock.

If it can verify, then an agent "should" be able to self-iterate, right? Has anyone tried building a very agent-native backtesting system? Based on that, you could then have claude code continuously write strategies to backtest and optimize, hahaha, a very trivial idea~ I know this idea will definitely be challenged from all sorts of angles until it actually works.

On Reading and AI

I was chatting with my senior labmate today and mentioned that reading books has felt way too slow for me lately. Because with AI, we want efficiency++ in everything we do, yet AI book summaries are lossy compression, so maybe we still have to stick to "traditional reading," except that's just too slow. So I pessimistically said, "Maybe this is why humans will inevitably lose to AI; even my reading speed will never catch up to them." Or maybe we can look forward to brain-computer interfaces?

For now, all I can do is try to minimize distractions as much as possible, and read more insights from people at the frontier, Engineering Blogs from top companies, some podcast interviews, and books.

The Intuitive Impact of AIGC

AI E2E Workflows

On Reading and AI

Recommended Reading