This month, I spent almost all my spare time in a loop of "Trial and Error" with OpenClaw.
It started out of simple curiosity. I installed this open-source AI agent platform on a Mac Mini M1 and watched it turn into a "front-end agent" capable of understanding complex commands. It gave me a strange sense of déjà vu. This week, I dug out a Raspberry Pi 4, flashed it with Ubuntu Desktop, and stuffed OpenClaw into it as well. Watching that tiny development board smoothly run an "agent brain" made the image in my mind finally crystallize:
This is exactly like the "Little Formless Skill" (Xiao Wu Xiang Gong) from Jin Yong's wuxia masterpiece, Demi-Gods and Semi-Devils.
Formless and Shapeless, Driving All Methods
The essence of the "Little Formless Skill" is that it has no fixed moves, yet it can power any martial arts style from any sect. Jiumozhi used it to mimic the Seventy-Two Supreme Techniques of Shaolin, making others believe he was a master of every school.
OpenClaw is exactly like that.
It isn't a single "Super AI Model"—it's not GPT-4, Gemini, or Claude. It is a "Mental Cultivation Technique" (an inner power method): it calls and integrates various top-tier models in the back-end while presenting itself as a unified "Agent" in the front-end.
Using OpenClaw on my Mac Mini, it connects to three or four AI services simultaneously: one specialized in reasoning, one in coding, and one in writing. Yet, on the front-end, I feel no transition at all. I only see a "master" who understands complex instructions, automatically breaks down tasks, and completes them beautifully.
Transitioning from a "Q&A machine" to an "Agent master"—this is a qualitative leap.
The Price: The "Little Formless Skill" Consumes Inner Power
But a month of trial and error has taught me one thing: this "divine skill" is extremely demanding on one's "inner power."
Powering Shaolin techniques with the Little Formless Skill requires a massive foundation of inner energy; Jiumozhi eventually suffered from "qi deviation," which was to some extent a result of overtaxing his energy. OpenClaw is the same—it consumes AI Quota heavily.
I've experienced this: refilling my API credits in the afternoon, only to receive an "insufficient balance" warning by evening. Checking the logs, I realized that for just a few rounds of "research a topic, summarize it into a report, and write an email based on that report," the agent had swapped models back and forth dozens of times behind the scenes.
Every sentence, every sub-task, and every retry burns through your quota. It's just like Jiumozhi consuming a portion of his "true qi" every time he executes the "Formless劫 Finger" (Wu Xiang Jie Zhi).
After moving OpenClaw to the Raspberry Pi this week, I started seriously researching "How to conserve inner power (AI Model Quota)":
First is Precise Task Decomposition: Don't let the agent solve problems through "brute force calls." Instead, think clearly about which steps truly require an LLM and which can be handled by rules. Next is Set Clear Boundaries: Tell it, "Think for a maximum of three rounds; if it's not working, give up," to avoid falling into infinite recursion. Along with this is Cache Repeated Results: Don't ask the same question twice (At least I tried). Finally, Choose the Right Model: Use cheap small models (Mini or Flash) for simple tasks, and only deploy GPT Pro for complex work.
These are lessons bought with "blood" (and money). But then again, what martial art doesn't consume energy? I've just learned to be more calculated, so I'm not "releasing a grand ultimate move" for every tiny task.
The "Little Formless Skill" on a Raspberry Pi
The most surprising thing this week was OpenClaw's performance on a Raspberry Pi 4.
The computing power of this small board is less than that of my phone. But because OpenClaw only handles the "scheduling of the skill," while the heavy lifting is done in the cloud, it still plays the role of a "master agent" remarkably well.
Looking at that tiny circuit board, using the "inner power" of OpenClaw to mobilize the world's top AI models to work for me—the scene itself is quite magical.
Conclusion: The Price and Future of Inner Power
A month of trial and error has led me to a realization: platforms like OpenClaw mark a turning point in AI development. In the future "Jianghu" (world), the competition won't be about whose model has more parameters, but about the depth of one's "Inner Power"—who can better integrate the strengths of many schools and who can do more with less "energy."
And I am still learning how to master this inner power before my quota burns dry.
After all, who wouldn't want to possess their own "Little Formless Skill"? Just as long as I don't follow in Jiumozhi's footsteps and end up with "qi deviation."