AI Starts Accelerating Itself: Grok 4.6 Submits 3 Optimizations and Enters Grok Chat Production Environment

According to Perceive Beating monitoring, Grok 4.6 has started participating in optimizing its own inference system. SpaceXAI revealed in the model card that they had let an earlier version of Grok 4.6 autonomously speed up Grok Chat. The model can propose solutions on its own, but must prove before submitting the code that the whole system has indeed become faster.

Within 5 hours, Grok 4.6 attempted 297 optimization proposals in one go, involving MoE, attention computation, low-level operator scheduling, and communication. It also eliminated a batch of solutions that only sped up in small tests but did not benefit when placed in the full system. In the end, the model submitted 7 PRs, with 3 already entering the Grok Chat production environment. This code has increased throughput in the generation stage by 1.5% and in the input processing stage by 3.1%.

SpaceXAI's goal is not only to optimize the inference system. The model card explicitly states that they are testing having Grok automatically take on some of the R&D work of training and evaluating new versions of itself. Tasks have already included detecting reward cheating during training, generating and reviewing training data, troubleshooting large-scale training systems, and designing evaluations for new capabilities.

This is already part of AI self-improvement, but it has not yet reached recursive self-evolution.