Musk promised a new version of Grok every two weeks, but the three core people who built the model all left within a week.

According to AIMPACT, on May 14th (UTC+8), Beating reported that three core technical personnel directly responsible for building the Grok model at xAI announced their departure in the past week. Unlike previous departures at the co-founder and executive levels, these three were the actual heads of three core product lines: search post-training, pre-training, and speech. Their collective departure will directly impact the iteration capabilities of Grok's next-generation model. Tianyi Zhang confirmed his departure at xAI today. He was the head of the search and factual post-training team and previously served as a machine learning engineering manager at Apple for five years. He led the team to reduce the factual error rate of Grok's real-time model by 70% to 80%, helping Grok 4.20 to top the Search Arena blind test leaderboard upon its launch. This search capability has also been integrated into Tesla's in-vehicle system. Juntang Zhuang announced on May 9th that he had left earlier this year. He was deeply involved in pre-training from Grok 2 onwards, later becoming the head of the pre-training team, leading the training infrastructure and core recipes for the entire Grok 4 series of models. Before joining xAI, he worked at OpenAI for two years, being a core contributor to GPT-4o and the inventor of the GPT-4 Turbo 128k long context algorithm. David Haxton confirmed his departure on May 10th, having previously built Grok's voice agent, real-time thinking speech, and voice cloning capabilities from scratch. In Sierra AI's τ-Voice voice agent benchmark, his team's Grok Voice currently ranks first, ahead of OpenAI and Google. The timing of this departure is particularly striking. The currently released Grok 4.3 is still only a training intermediate version with 0.5T parameters. On April 18th, Musk stated that the true 1T flagship was "about 5 days away from completing initial training." The following day, he clarified that Grok 4.4 = 1T, expected to be released in early May, and Grok 4.5 = 1.5T, aiming for delivery by the end of May. He also claimed that the model factory was running smoothly, with a new base model being released every two weeks thereafter. Now it's mid-May, and Grok 4.4 has still not appeared, while the core personnel actually building these models are rapidly leaving. xAI is experiencing a systemic talent collapse, from the co-founder to the core model layer. (Source: ME)