**[AI Begins Accelerating Itself: Grok 4.6 Submits 3 Optimizations Now Integrated into Grok Chat Production Environment]** According to monitoring by *动察 Beating*, Grok 4.6 has started optimizing its own inference system. SpaceXAI disclosed in the model card that an earlier version of Grok 4.6 was autonomously accelerating Grok Chat. The model can test its own solutions, but before submitting code, it must prove that the entire system genuinely becomes faster. Within 5 hours, Grok 4.6 attempted 297 optimization solutions in one go, involving MoE (Mixture of Experts), attention computation, low-level operator scheduling, and communication. It also eliminated a batch of solutions that only showed speed improvements in small tests but yielded no benefits in the complete system. Ultimately, the model submitted 7 pull requests (PRs), 3 of which have already been integrated into the Grok Chat production environment. These updates increased throughput during the generation phase by 1.5% and improved input processing by 3.1%. SpaceXAI’s goal is not limited to optimizing inference systems. The model card explicitly states that they are testing Grok’s ability to autonomously take on part of the research and development tasks for training and evaluating new versions of itself. These tasks already include detecting reward hacking during training, generating and auditing training data, troubleshooting large-scale training systems, and designing evaluations for new capabilities. This falls under the category of AI self-improvement, though it has not yet reached recursive self-evolution. [Original Link]