Not One Revolution, But Two
Talking with a few friends recently, plus spending over ten hours a day immersed in coding agents, one feeling keeps getting clearer: this wave of AI is not one revolution, but two.
The Programming Thread Has Been Discussed Enough
The first thread is programming and productivity. Coding agents, AI office assistants—the entire workflow of white-collar workers in cognitive processing and software creation is being reshaped. There's been plenty of discussion on this, and I've written about it before, so I won't dwell on it.
A few numbers to convey the scale: GitHub's statistics say AI has written 26.9% of production code, developers using AI save an average of 3.6 hours per week, and merged PRs are up 60%. There are also dissenting voices—METR's randomized controlled trial found that core contributors to 16 large open-source projects actually slowed down by 19% after using AI tools, though they felt 20% faster. Faced with complex, large codebases, AI hasn't reached the point where you can use it mindlessly.
My own experience is that when you know what you want, coding agents can indeed boost efficiency by an order of magnitude. The key phrase is knowing what you want.
The Other Thread, Probably Underestimated
The second thread is content creation. Text-to-image, text-to-video, image editing—what's happening in this direction is no less intense than on the programming side.
ByteDance's Seedance 2.0 is a watershed moment.
Released in February, API opened at the end of March. Artificial Analysis's benchmark gave it an Elo of 1269, surpassing Google Veo 3, Sora 2, and Runway Gen-4.5. Not by a little, but by a wide margin. It can generate up to 20 seconds of 1080p video in one go, with music, dialogue, and sound effects synchronized, no post-production dubbing needed. Camera movement, lighting, and character motion can all be precisely controlled.
A friend of mine working on AI short dramas says there are plenty of teams in the industry sitting on tens of millions in cash waiting to use it. The day the API opened, my social media feed exploded with founders in related spaces—shorts made with it were everywhere, and the quality was on a completely different level from before.
The cost shift is staggering. Previously, a 25-minute episode of Japanese animation cost roughly RMB 1 to 3.2 million. Attack on Titan ran about RMB 1.1 million per episode, and Jujutsu Kaisen was similar. Now a 3-minute AI short costs RMB 400 to 1,200. The per-minute cost has dropped to a few percent of traditional methods. In blind tests, 73% of viewers can't tell the difference.
Another friend runs a small live-commerce team. They told me content production costs have dropped to one-tenth of before, and speed is up 10x. A content creator used to cost over RMB 10,000 a month; now the same output costs an order of magnitude less. Demand completely outstrips supply.
Imagine someone spending a few thousand RMB and one week to make a 20-minute anime short. Impossible before. Now the outline is already clear—text-to-video has reached its "ChatGPT moment." Once the path is found, cost reduction is only a matter of time. That's how AI works: someone blazes a trail, and cheaper solutions inevitably follow, because at its core it's software, endlessly iterable through data and training.
The Two Paths Are Diverging
I recommended coding agents to a friend. He said he hasn't had the energy to dig into them lately—not because he isn't smart, but because all his time and passion are going into content creation. Every day he's figuring out how to build pipelines with the latest tools, how to express ideas at low cost. I thought about it: why should he study coding agents? That's not his direction.
The reverse is also true. I spend over ten hours a day in coding agents; if you asked me to research content-creation pipelines, I wouldn't reach that level anytime soon.
In my last post I wrote about the split between Builders and Promoters. This time the feeling is more concrete: these two paths differ not just in role, but in tools, skills, and the passion required. There is overlap, but it's shrinking, while the divergence is widening.
This means it's not just product builders undergoing massive change. People passionate about expression—content creators—are about to receive tools of the same magnitude. If you have a strong urge to express something about a topic, for the cost of a few dozen RMB and half a day, you can make a one-minute short video. Smartphones and TikTok already lowered the barrier to filming once; next, that barrier will drop by another order of magnitude.
Agent to Agent: Meetings Are Dead
Both paths share a common downstream effect: they are changing how people interact with one another.
A friend asked me: with AI this powerful, would it be safer to return to face-to-face fields—sales, talking to investors, managing supply chains?
I don't think so.
I now see meetings as an extremely inefficient format. I record a voice memo, AI transcribes it and polishes it into a document with more information than a two-hour meeting. The other side uses an agent to digest the document and extracts the key points in minutes. What I do in five minutes might take two hours of meetings just to get the full picture.
The efficiency gap is 100x. That's not rhetoric.
The infrastructure is already taking shape quickly. Google released the Agent-to-Agent protocol (A2A) last year, backed by over 50 enterprises including Salesforce, SAP, and PayPal. Anthropic's MCP lets agents plug into various tools and data sources. Earlier this year, the Linux Foundation launched the Agentic AI Foundation, joined by OpenAI, Anthropic, Google, and Microsoft, with over 100 enterprises following by February. Gartner predicts that 40% of enterprise applications will embed task-specific AI agents by the end of this year, up from less than 5% last year.
Your agent and the other party's agent analyzing, transmitting information, and coordinating in the background will be far more professional than two people chatting face-to-face. Agents don't need small talk; they don't need two weeks to build trust. They synthesize judgments from all verifiable information, which is more reliable than hearing someone say two sentences and deciding "this person seems fine."
What E-Commerce Eliminated
This reminds me of e-commerce.
Before e-commerce, all transactions were face-to-face. To buy something, you had to meet the seller, build a relationship, and size each other up. The seller's social skills and charisma were critical to closing the deal.
E-commerce arrived. Products are roughly the same, compare prices, read reviews, place an order in ten minutes. You don't know the seller, you've never met him, and he doesn't even know you bought it. The B2B space is the same—buyers have completed 57% to 70% of their purchasing research before contacting sales, and 67% of the buying journey happens online.
E-commerce didn't eliminate retail. Physical retail still accounts for 81.6% in the U.S. But it restructured the relational logic of transactions—face-to-face social skills went from "essential" to "nice to have."
Agent-to-agent will bring a similar but far stronger wave. When the efficiency gap is 100x, people will choose the more efficient route in many scenarios. It's not that humans don't matter; it's that business matters now have a better channel.
Play Ball Without Talking Business
There's a side effect I think is quite positive: human relationships will become purer.
Before, you had to talk business on the court and discuss deals over dinner. Face-to-face was the most efficient business communication channel, so socializing and commerce were always intertwined. You didn't go play ball purely because you loved the game; you went because the court was where you could meet the people you wanted to know.
What if agents handle the business part? Then playing ball is just playing ball. Socializing is just socializing. You no longer have to maintain unwanted relationships in settings you don't enjoy.
Intimacy, entertainment, interests—these needs will be stripped out of business contexts. It's not that they aren't important; it's that they can finally be nothing but themselves.
Go Find It
Two revolutions are running in parallel, each demanding serious time to master.
The pattern I've observed is simple: people with passion put in ten hours a day and master every tool. Those without passion open them occasionally, and the gap quickly becomes an order of magnitude. The leverage of AI is right there; whether you can move it depends on whether you're willing to keep applying force.
Whether building products or creating content—go find the thing you can't stop doing.
Originally published at https://guanjiawei.ai/en/blog/two-ai-revolutions