Original video in Chinese.
Key Takeaway
- System-level shift: moving from the old add-on RAG system, which had major limitations, to a new system centered on a Coding Assistant (such as Cursor/Claude Code). Retrieval and editing capabilities have generalized into daily work; domain experience has moved from Workflow to Skills.
- Assessment of the giants: Google has the strongest full stack (leading multimodality, a complete ecosystem, Gemini + Nano Banana visual productivity), and I will remain in the Google camp for the long term; Anthropic is steadily advancing with a focus on Agent/Skills; OpenAI is fighting on multiple fronts, losing focus strategically and struggling to support its valuation.
- Personal gains: my thinking has matured and converged; I am less inclined to tinker with open source or deployment, and more focused on refining my personal AI OS; data assets are non-transferable, so once I choose a platform, I will keep building on it for the long term.
It’s time for another year-end review. I’m planning two videos, one looking back at the industry and one at my personal journey. In this video, I’ll first talk about what I gained from using AI over the past year.
One of the benefits of being a tech blogger is that whenever I have an idea or try something out, I record it in video form. I compared the videos I posted in 2024 with those from 2025 and found two very obvious changes.
First, at the system level. I abandoned the old add-on RAG system and shifted to a new system centered on a Coding Assistant.
Those of you who know me well know that I originally made a lot of videos about RAG. That’s because I firmly believed:
For AI to become useful, two things must be added to it — domain knowledge and domain experience.
In AI terms, domain experience used to be Workflow, and now it is Skills. After Anthropic launched Skills, I made many recommendation videos about it for exactly this reason.
In AI terms, domain knowledge is RAG. Through RAG technology, when AI generates responses, it relies not only on the general knowledge it was trained on, but also on domain knowledge provided by the user.
In 2024, I explored many tools around RAG, and I also discovered RAG’s major limitations — I even made a special video about this before.
The turning point came this year. After Vibe Coding took off, I started working on products. When I got in touch with Cursor, I suddenly realized that its retrieval and editing capabilities are not only suitable for programming, but also perfectly fit everyday work scenarios!
Think about it: model context windows are longer now, and native tools like Cursor have internalized retrieval well enough that users no longer need to worry about intermediate steps like vector databases or chunking.
So I connected Obsidian and Cursor. Before this, I had used many AI plugins, but none of them quite did it for me. Now that I have Cursor, I found I can open my note repository with Cursor. In this setup, Cursor handles editing at the front end, while Obsidian steps back to the backend and handles storage.
This idea was greatly strengthened later on.
After MCP came out, we were able to connect more tools. For example, connecting to the MCP for Tavily or Perplexity enhances the system’s external search capabilities.
Claude Code, as the strongest Coding Assistant, has powerful tool calling, and it can also switch the model used as the driving engine, which gave the whole system a qualitative leap.
And later, the launch of Skills allowed us to encapsulate domain experience and give it to AI. In this way, both the domain knowledge and the domain experience I talked about before were now complete!
As it turns out, my approach was completely right. Quite a few people both in China and abroad have also come to the same conclusion. Anthropic renamed the Claude Code SDK to Claude Agent SDK. In their official blog post, they said this thing can be used in a much broader range of scenarios, not just programming.
The AI OS video I made some time ago was a staged summary of this line of thinking. If, like me, you care deeply about maximizing productivity and about building a personal AI system, you should definitely try my method.
Second, at the tool level. Fully embracing the Google ecosystem.
I started using Gemini 2.0 as my daily main model from that period onward. Because it is really so comprehensive. The more I use it, the more I like it. And with model support, the products within Google’s ecosystem have also seen a qualitative leap.
For example, the newly launched NotebookLM is currently recognized as the strongest AI learning tool. Older products like Gmail and Docs have also integrated Gemini and been connected horizontally, which has greatly improved productivity and the overall experience.
The AI OS I introduced before relies on the Claude Code framework and is only suitable for use on desktop. Google’s ecosystem products, on the other hand, cover everything from desktop to mobile, making them a great fit for my lightweight needs.
Among the current AI top three — Google, Anthropic, and OpenAI — on the consumer side, I still firmly favor Google. I have said more than once that once native multimodality really takes off, it will bring enormous improvements to both models and products.
That’s why I said in the Knowledge Planet that next I would gradually store a copy of the materials I’ve accumulated in Google Drive. Both Gemini and NotebookLM can import materials from Google Drive. Or you don’t even need to import them — you can just open Gemini in Google Docs and have conversations based on the document content.
On the enterprise side, I’m optimistic about Anthropic. Like I said earlier, the Claude Code stack is already a general-purpose Agent framework. This isn’t just Anthropic boasting about itself — it’s what they saw a large number of users actually doing. For companies, if they want to develop Agents, rather than building a framework themselves, it is more cost-effective to use Anthropic’s SDK.
As for OpenAI, I think they are now fighting on too many fronts and are a bit overextended.
OpenAI is now competing with Google on multimodality and ecosystem, competing with Anthropic on coding ability and Agent capability, and also building massive infrastructure. They are fighting on all sides, trying to support their valuation. But judging from the situation over the past few months, strategic drift has already made OpenAI somewhat strained.
As a user, even setting aside my personal preference for products, I don’t dare choose OpenAI. As I said in the Knowledge Planet, choosing a platform, putting your data assets on it, and using it long term so the model gets to know you better — that’s something you cannot casually migrate.
So at the tool level, I will probably stay in the Google camp for good.
These two levels of things are my biggest gains in 2025. If you’ve been watching my videos all along, you’ll know that all my content this year has revolved around these two levels.
Compared with 2024, I’m not using open-source models as frequently, and I’m not as enthusiastic about deploying all kinds of open-source projects. This shows that my whole line of thinking has started to mature and converge. I’m putting more energy back into refining my own system, rather than running around trying new tools.
If I have more thoughts later, I’ll share them within the community. The environment is getting worse and worse now, and anything you say gets attacked.
OK, that’s it for this episode. If you want to understand AI, want to become a super individual, and want to find like-minded people, come join our newtype community. See you in the next episode!