CHAI AI: Post Training to Achieve SOTA MoE

▤ Summary

CHAI AI announced that its latest in-house model improved Day-30 user retention by 4% after training with the GSPO reinforcement learning algorithm. The company has been shifting from dense models to Mixture-of-Experts architectures and testing several open-source model variants on its social AI platform. It is also exploring additional RL methods such as DPO, GRPO, and GSPO to improve model quality and engagement. CHAI says its latest model now serves about 1 billion messages per day, and the company continues to push deeper AI personalization and user-generated AI.

▦ Linked Companies