A Chinese AI lab just quietly released a model that trained itself to improve by 30% 🤯 The @MiniMax_AI team gave M2.7 access to its own training infrastructure. Its own reinforcement learning pipeline. Its own memory. Then they told it to make itself better. And it did... M2.7 built dozens of complex skills inside its own RL harness. It updated its own memory systems. It optimized its own reinforcement learning process based on results it was evaluating in real time. It ran this loop autonomously for over 100 rounds, analyzing failure trajectories, modifying its own scaffold code, running evaluations, comparing results, and deciding what to keep or revert. No human in the loop. Just the model improving the model. The result: - 30% performance improvement on internal evals - 66.6% medal rate on ML competitions (This ties it with Gemini 3.1. SWE-Pro scores while nearly matching Claude Opus) MiniMax already runs 30% of its entire company operations autonomously on its own models. 80% of newly committed code at the company is AI-generated. They're literally shipping self improvement as organizational infra and the Honk Kong stock market LOVES it. This is the early stages of recursive self-improvement. And it's coming from Shanghai.