Analysis
Alibaba's decision to bring on Zhou Hao, previously a key figure in Google's Gemini LLM development, signals a major push to improve the accuracy and reliability of its 'Qianwen' LLM. Hao's expertise in factuality and reducing AI 'hallucinations' promises exciting advancements in the model's performance and user experience.
Key Takeaways
- •Zhou Hao's focus on factuality aims to prevent AI 'hallucinations', ensuring more reliable outputs from the LLM.
- •He was a key contributor to Gemini's advancements, including those related to the accuracy of information.
- •Hao's background includes experience in computer vision, especially video understanding, and large model training at DeepMind.
Reference / Citation
View Original"周浩和团队在Gemini上做的事实性工作，是从模型训练和强化学习的源头入手。通过设计专门的奖励函数和训练策略，让模型在生成每一个token的时候，就内化了“事实准确性”这个约束。"