Elon Musk, CEO of xAI, revealed today on the X platform that Grok 4.8 will be an AI model with 2.5 trillion parameters and will be trained using a completely new C++ software stack. He also provided a clear timeline: Grok 4.8 will complete training this week and then proceed to the reinforcement learning phase.
Looking back at the timeline, there was another incident. Musk previously announced that Grok 4.7 would be released on September 12, but later changed his mind, saying it might take a few more days before it is officially launched. The reason for the delay he mentioned was quite practical - the penalty for response length during the reinforcement learning phase was too heavy, causing the model to give up on difficult tasks too early and fail to learn how to carefully review its own work, which was a critical obstacle in the training process that needed to be resolved.
