It is currently the evening of January 31, 2026.
Recently, I've been diving deep into OpenClaw, a highly popular AI tool that can control an entire machine with full permissions. I deployed it on an Alibaba Cloud ECS server, equipped with the MiniMax M2.1 model, and connected it to my Feishu bot. Now, I can chat with OpenClaw through Feishu, and it helps me execute various tasks.
However, OpenClaw is currently more suitable for installation on a local, empty machine rather than a cloud server. The main reason is that cloud servers are relatively expensive. Without installing various tools, such as a browser, OpenClaw's functionality is quite limited. This also highlights that the world's APIs are not yet fully open, often requiring the use of browsers or various apps to complete tasks.
Another issue is that OpenClaw, even with the MiniMax M2.1 model, still seems somewhat unintelligent. The author mentioned that Peter Steinberger said MiniMax M2.1 is the best open-source model, but also noted that Opus is far ahead. In my opinion, MiniMax M2.1 is still relatively weak. I wonder how Opus + OpenClaw would perform.
My GitHub Copilot subscription allows limited access to the Opus model, but it cannot be used with OpenClaw—only with OpenCode. Since the quota resets tomorrow, February 1st, I decided to use up my quota today.
So, I asked OpenCode + Opus to help me write the experimental code for the Capital Protracted War, open-sourced here. I must say, I tried this once before with MiniMax M2.1, and the result was terrible. Opus performed significantly better, truly living up to its reputation as the SOTA model for coding.
The experimental report will be maintained in the open-source project. However, I can share the current experimental conclusion:
Under the high-volatility market model of GBM, a basic mean reversion strategy combined with anti-Martingale money management can still achieve exponential capital growth, even with transaction costs.
Surprisingly, trend-following strategies under anti-Martingale money management cannot achieve exponential growth, whereas mean reversion strategies can. This highlights the advantage of high-win-rate strategies. However, I remain skeptical and need to verify this further. My intuition tells me that even with a lower win rate, trend-following strategies should also be capable of exponential growth.
Further validation is still needed. Please follow the Capital Protracted War experimental open-source project!
