Since the launch of DeepSeek, a wave more intense than ChatGPT has surged. The general public has been directly impacted by AI. For a time, AI tools seemed invincible. Most people have interacted with AI tools to varying degrees - or been exposed to them. Some use it for fortune-telling, others as a tool for social mobility. This trend has also reached Obsidian. I myself have used at least 4 AI plugins - Smart Composer, Copilot, Star AI, Text Generator. Don’t misunderstand - I’m not here to evaluate plugins or pit them against each other, but to explore the practicality of AI tools.
Based on my personal experience, I believe AI assistants are overhyped. In terms of operation, various AI tools essentially embed AI features - from Copilot’s dialog boxes to Smart Composer’s automatic modifications requiring human approval, to Star AI’s fixed prompts.
None of them significantly improve workflow efficiency: they eliminate the need for extra steps like copy-pasting (though scripts could handle this, I doubt anyone would choose such complicated methods).
On the other hand, features like tag association, property generation, and article expansion save massive amounts of time and energy.
But fundamentally, current AI assistants cannot replace human roles.
Why?
Firstly, the human brain’s ability to retrieve and associate different concepts remains far superior to current AI: AI that takes ages to search databases is impractical for real-world use (unless you have access to a super high-performance AI lol). The slowness isn’t due to insufficient search capabilities - quite the opposite, it’s because it needs to search everything every time.
The solution would be to establish a hierarchical system, similar to the human brain - first retrieving general directions then recalling specific content. In Obsidian terms, this would mean first searching by note titles, bidirectional links and properties, then drilling down.
Additionally, mimicking human cognition, we need a “recent access” database - which in Obsidian might involve checking recent files, or prioritizing notes with the most recent edits or longest editing time.
Another crucial point: AI ultimately isn’t you, and doesn’t possess your knowledge. This became apparent when I created Anki flashcards: flashcard content needs dynamic adjustment. For instance, BIOS is basic knowledge for computer experts but crucial for beginners. However, AI lacks your personal knowledge base - it can’t discern what’s essential vs redundant, even when given learning objectives.
Even worse, AI can be remarkably dumb at times (or perhaps inherently dumb): If you feed it lots of data and principles (like Anki card creation rules), it doesn’t know what to do beyond mechanically generating cards. But with explicit instructions, its performance improves dramatically. My approach is to provide operational procedures and principles, making it process information accordingly - first reviewing principles, categorizing knowledge, then formatting it, etc.
I find content modification by AI currently unreliable. A better approach might be using AI tools to update project properties, delete useless notes, generate dataview code (though components can handle this haha) - helping humans eliminate unnecessary tasks rather than producing hallucination-filled AI articles (Frankly, what’s the difference between using this inside vs outside Obsidian? It completely wastes the platform’s advantages!)
Finally, a small joke: If anyone actually implements my ideas, please credit the source (though you can claim it’s your own idea XD)