I was pessimistic halfway through the year, truly believing that breakthroughs could only be made outside of reinforcement learning to go further. Turns out I was wrong. I haven't had the chance to test those code models in real environments, but after trying a newer version, I realized — we've already crossed the inflection point. The performance leap in a short period is so obvious that it's no longer incremental improvement, but a qualitative shift. I have to admit I was mistaken in my judgment.