Original video in Chinese.
Key Takeaway
- AI has entered a “period of calm,” and the main battlefield has shifted: this is not technical stagnation, but a shift in focus from “how strong AI is” (models, products, protocols, Skills) to “how AI enters the environment” (where it works, what it can see, what it can operate on, what permissions it has, how it recovers from failure, and how the results move into the next step of the workflow).
- Big companies are competing for “environmental control”: Tencent’s Workbuddy connects documents, meetings, knowledge bases, and more, linking the three major closed loops of data, collaboration, and capability in an attempt to control the work environment; ByteDance’s organizational restructuring is also meant to deliver a complete solution for “how AI exists in the work environment,” rather than a single-point product.
- Individuals also need to let AI enter their own work environment: use 30 days to make one repetitive task work end to end—pick the task in week one, fill in the context and project files in week two, run it continuously and systematically fix errors in week three, and add checks and feedback in week four. Once that’s done, you’ll have a repeatable, improvable AI work system that truly belongs to you.
Over the past few months, I’ve increasingly felt like AI has “nothing new” anymore.
The models are still being updated, and products are still being released, but those moments of “holy shit, so it can do that too?” are happening less and less.
I’m not saying AI development has stalled. What I mean is that it has probably entered a period of calm. And a period of calm doesn’t mean the technology has stopped moving; it means the main battlefield has changed.
Over the past few years, AI has amazed us many times.
At first, it was the model capabilities themselves.
I think the most representative model was Opus 4.6. It made many people feel for the first time that an Agent wasn’t just something that could answer questions, but something that could genuinely do things for a long time.
Then came the product form.
First Claude Code, then Codex. They made AI no longer just a chat box, but a collaborator that could help you write code, modify files, and run tasks.
After that came the protocol layer.
With the emergence of protocols like MCP, a very important change happened: AI finally had the chance to connect to external systems, read tools, and invoke capabilities. In other words, the wall between AI and the real work environment began to thin.
Then came Skills. You were no longer just asking one question and getting one answer; you could package an entire workflow method and let AI do the work according to your rules, your format, and your process.
Each of these layers was astonishing. But now, that sense of astonishment has weakened. I think the main reason is that the focus has shifted from “how strong AI is” to “how AI enters the environment.”
In other words, before, everyone was competing over: Can AI write code, analyze files, generate images, and perform complex reasoning?
Now it’s become: Where does AI work? What can it see? What can it operate on? What permissions does it have? How does it recover after a failed execution? How do the results enter the next step in the workflow?
This change is already very obvious at big companies. Tencent’s Workbuddy is the most typical example.
In my view, the most impressive thing about WorkBuddy is that it turns Tencent’s originally scattered office environment into an environment that Agents can enter and work in. What it is connecting is not a few isolated functions, but a relatively complete workspace.
For example, Tencent Docs provides a space for collaborative editing and delivering outcomes; Tencent Meeting preserves the context of team discussions; Tencent Drive and ima provide organizational knowledge, and so on.
Workbuddy is completing three closed loops: the data loop, the collaboration loop, and the capability loop. Once these are complete, it gains “environmental control.” And whoever controls the work environment is more likely to control the entry point of AI value in the next stage.
That’s also why ByteDance needs to adjust its organizational structure. Because what you need to deliver is a solution for “how AI exists in the work environment.” This is not something a single-point product can handle.
Alright, enough about big companies. Let’s come back to individuals.
Individuals are actually the same.
A big company’s goal is to let its AI enter the work environment of millions of people. For individual people like us, the goal is also to let AI enter our own work environment.
So what should we do?
My suggestion is: don’t think about doing something huge right away. Instead, pick one thing you repeat often and that wastes a lot of time, and spend 30 days making it work end to end.
In the first week, do only one thing: find a repetitive task worth transforming.
Note: don’t choose a huge goal like “help me run my entire content account.” Choose a specific task, such as turning one topic idea into a video script, or turning a financial report into a company research note, and so on.
So how do you tell whether a task is worth doing?
Look at three conditions:
First, does it happen often?
Second, does it take a lot of time every time?
Third, is it relatively easy to check whether it’s done well?
If the answer to all three is yes, then it’s a good candidate for your first AI workflow.
In the second week, don’t rush to automate. Instead, let AI truly understand the work.
Using content creation as an example, you need to tell AI: what your account positioning is; who the audience is; what content you’ve made before; what style of expression you prefer; what structure your content usually follows; and so on.
Organize all of this into project files. When AI enters this project later, it’s like a new employee walking into the office. On the desk, the account manual, reference materials, work requirements, and past best cases are already laid out. It doesn’t need to ask you again and again what the company does.
In the third week, let AI complete 5 to 10 tasks in a row using the same method.
Every time it makes a mistake, you need to figure out: why did it make this mistake here? What do I need to change so it won’t make the same mistake again next time?
For example, if AI cites unreliable data, don’t just manually replace that data. Instead, add a rule to the workflow: key data must have an original source, along with the date and a link.
That way, each correction improves not just one piece of work, but the entire production process.
In the fourth week, start focusing on “how well AI completed the task.”
For example, in content creation, you can observe: how much time did one draft save; how much still needs to be revised; which headlines have a higher click-through rate; and so on.
Then feed those results back to AI and let it optimize.
After these four weeks, the first week is selecting the task, the second week is filling in the background, the third week is running it repeatedly, and the fourth week is adding checks and feedback.
Once that’s all done, you’ll have a workflow that can be executed repeatedly in the future, continuously improved, and truly belongs to you; you’ll have completed the first real integration between AI and your own work environment; you’ll no longer be someone who uses tools, but someone who starts designing their own AI system with your own hands.
OK, that’s it for this episode. If you want to understand AI, want to reclaim individual sovereignty, and want to find like-minded people, come join our newtype community. See you next time!