Researcher Zhang Dinghuai Announces Departure from xAI Reinforcement Learning Position
Researcher Zhang Dinghuai stated in a post that he has left xAI for some time. He wrote that the past year has been the most intense and rewarding phase of his career, focusing on the large-scale training of cutting-edge models: stability, large-scale training, and efficiency.
He studied mathematics at Peking University for his undergraduate degree and pursued his PhD under Yoshua Bengio at Mila, during which he worked at Apple Machine Learning Research and Meta FAIR. After graduation, he served as a senior researcher at Microsoft New England Research and then joined xAI.
At xAI, he worked on the stability of reinforcement learning, scaling reinforcement learning, and large-scale training of programming reinforcement learning. After pre-training, reinforcement learning was used to further enhance reasoning, programming, and agent capabilities, allowing thousands to tens of thousands of GPUs to run continuously for days or weeks without crashing or degrading suddenly.
When Grok 4.5 was released, he wrote about his foundational algorithm changes to promote the stability and capability of reinforcement learning, iterating on infrastructure day and night, and executing large-scale training at the cutting edge. The model was positioned for programming and agent tasks, trained using tens of thousands of NVIDIA GB300 GPUs, with reinforcement learning covering hundreds of thousands of multi-step software engineering tasks.
He thanked colleagues in the overnight war room who helped restart training after system crashes, stating that the team moved quickly and dove into problems until the system was operational again. His academic work also includes research on generative models, probabilistic methods, and enabling models to self-adjust their curriculum in reinforcement learning.
In market mechanisms, the cost is borne by laboratories needing to catch up on programming agent capabilities, while the few researchers who can run large-scale reinforcement learning stably are in demand. The event is driven by personnel movement, with funds flowing from the merged platform to competitors that can still attract key personnel for later training; benefiting are the next laboratories that can open positions, while those needing to continue pushing Grok's programming capabilities face scheduling pressures.
Source: Public Information
ABAB AI Insight
Later-stage training has shifted from "just one more round of human feedback" to "keeping thousands of GPUs running continuously for weeks without crashing." Zhang Dinghuai's focus was on stability and scale efficiency in this area, not just another prompt-writing position. Grok 4.5 markets programming and agent capabilities, relying on the type of large-scale reinforcement learning he participated in. With his departure, the schedule remains.
The capital path is that the merged platform exchanges computational power and war room efforts for model versions, while researchers trade a year of high intensity for cutting-edge training on their resumes. Since the merger of SpaceX and xAI, co-founders and key personnel in pre-training and programming have continuously left. Money flows from the merger valuation into clusters, while people move from clusters to the next opportunity. The speed of talent turnover in later-stage training outpaces the release speed of model versions.
The benchmark is the poaching of OpenAI's coding team, the diversion of DeepMind's reinforcement learning group, and chip companies recruiting core compiler engineers to the next wafer fab. The industry is at a stage of homogenization in pre-training and differentiation in later-stage training: everyone has large models, but the competition is whether reinforcement learning can stably scale in real engineering tasks. Those who cannot retain personnel who have run large training will pass the variance of the next version to newcomers.
Structurally, this belongs to technological replacement. The mechanism is: as capability increments shift from data accumulation to large-scale reinforcement learning, the scarce resource changes from data engineers to those who can prevent training from crashing. Systems can fail, but schedules cannot stop, making the war room a mode of production. After people leave, what remains is the infrastructure and the next large training that has not yet been completed.
ABAB News · Cognitive Laws
The model can scale, provided that training does not crash first.
The war room can produce versions, but also resignation letters.
Once the personnel for later-stage training leave, rescheduling is harder than adjusting weights.