AI - Gold (2) The Golden V2 version was launched after one week of running the V1 version (which was the most glorious first week of V1) At that time, I thought to myself that even a handful of V1 could do this. So, if we thoroughly build a wave of V2, integrate more data, and re optimize the agent from top to bottom, can we achieve a better result? So that weekend, I rubbed this wave of V2 V2 actually mainly incorporates data from COMEX's gold futures GC tick . So compared to V1, it mainly adds more agents in terms of gold order flow But V2 is not a completely improved version of V1 Should V2 draw inspiration from V1's approach. At that time, I thought V1 had to continue running, so V2 didn't need to run the good ones in V1 again So V2 still follows the idea of doing things again in V2, based on what V1 hasn't done well . Finally, V1 and V2 run in parallel, and during the operation, various models are tested, including opus 4.6， gpt 5.4, kimi 2.5, kimi 2.6, qwen 3.6 plus, deepseek v4 pro, glm5.1, mimo v2 pro, minimax m2.7 wait. . It can be said that we have tried all the mainstream options So during the craziest period, it was equivalent to three sessions per day (Asian, London, New York) * each session V1+V2 added up to nearly 20 agents with different ideas * each V1 V2 of these 20 agents used many models to produce results simultaneously with four .. Each run of V1 and V2 using a large model requires approximately 300k to 500k tokens In short, if the data is full and the budget for the token is full, let's see which combination of model and agent can produce results .. (The upper limit for a single agent is 65536, because if it is found that a single agent has 30K, things like Kimi 2.5 often use up 30k during thinning, resulting in no output results in the end.) In the end, it was discovered Domestic models can indeed produce results The foreign model is undoubtedly still opus After four weeks, it was found that (Figure 1-4) In the first two weeks, V2 gave me good confidence because it outperformed V1 in terms of win rate and PF But last week and this week were a hellish market for V2 .. Overall, V2 is basically in a 4-week no profit state . Prepare to run for another 2 weeks for a trial ..