In the past week, three of the technical leads who actually built Grok at xAI announced their departures. Unlike earlier exits of co-founders and executives, these three were the hands-on heads of post-training, pre-training, and voice — the core product lines. Their collective exit directly threatens the next generation of Grok.
Tianyi Zhang confirmed his departure on X today. He led the search and factuality post-training team, previously spending five years as a machine learning engineering manager at Apple. His team slashed Grok's real-time fact-error rate by 70-80%, helping Grok 4.20 hit the top of the Search Arena blind benchmark on launch. That search capability is already integrated into Tesla's in-car system.
Juntang Zhuang announced on May 9th that he left earlier this year. He was deeply involved in pre-training since Grok 2, eventually becoming head of the pre-training team, where he led the training infrastructure and core recipe for the entire Grok 4 series. Before xAI, he spent two years at OpenAI, contributing to GPT-4o and inventing the long-context algorithm for GPT-4 Turbo 128k.
David Haxton confirmed his departure on May 10th. He built Grok's voice agent, real-time thinking voice, and voice cloning from scratch. On Sierra AI's τ-Voice benchmark, his Grok Voice team currently ranks first, beating OpenAI and Google.
The timing of these departures is particularly glaring. The publicly available Grok 4.3 is still just a 0.5T parameter intermediate checkpoint. On April 18th, Musk said the true 1T flagship was "about 5 days from initial training completion," and the next day he specified Grok 4.4 = 1T, expected in early May, and Grok 4.5 = 1.5T, targeting late May. He also claimed the model factory was operational, enabling a new base model every two weeks. Now it's mid-May, Grok 4.4 hasn't appeared, and the core people building these models are leaving faster than ever.
xAI is experiencing a systematic talent collapse, from co-founders down to the core modeling layer.