Grok 4.6 deploys three optimizations in production, enhancing chat performance.

According to monitoring by Beating, Grok 4.6 has begun participating in optimizing its own inference system. SpaceXAI disclosed in the model card that they enabled an earlier version of Grok 4.6 to autonomously accelerate Grok Chat. The model can test optimization strategies on its own, but must prove that the entire system actually becomes faster before submitting any code. Within five hours, Grok 4.6 tested 297 optimization approaches involving MoE, attention computation, low-level operator scheduling, and communication. It also eliminated solutions that showed speed improvements only in small-scale tests but yielded no benefit in the full system. Ultimately, the model submitted seven pull requests, three of which have already been deployed to Grok Chat’s production environment. These code changes improved generation throughput by 1.5% and input processing by 3.1%. SpaceXAI’s goals extend beyond optimizing the inference system. The model card explicitly states they are testing Grok’s ability to autonomously take on parts of the research and development workload for training and evaluating its own future versions. Tasks already include detecting reward hacking during training, generating and reviewing training data, debugging large-scale training systems, and designing evaluations for new capabilities. This constitutes AI self-improvement, but has not yet reached recursive self-enhancement.