CHAI

Personal entertainment

Users create and interact with custom AI characters for storytelling and roleplay.

Social engagement

The platform optimizes chat sessions to increase user retention and session length.

Creative writing

Build AI-driven narratives and characters for fun or experimentation.

Community building

Users can share and discover AI characters created by others.

Research and development

CHAI’s infrastructure supports advanced RLHF and model blending for AI research.

Custom AI character creation

Users can build their own ChatAIs with unique personalities and storylines.

In-house LLM deployment

CHAI deploys its own 6B and 13B LLMs, trained specifically for social and engagement purposes.

Model blending

The platform invented model blending, ensembling different LLMs to outperform GPT-3 in user retention.

Reinforcement learning optimization

Uses PPO (Proximal Policy Optimization) and DPO (Direct Preference Optimization) to boost engagement and reduce chat session endings.

Reward model scaling

Deploys reward models trained on up to 100 million user signals to decrease retry rates and increase chat session length.

Efficient inference at scale

Custom CUDA kernels and GPU orchestration support 500K daily active users.

Best-of-8 rejection sampling

Uses upgraded reward models and cluster capacity to improve response quality.

Cross-model upgrades

All production blends upgraded to DPO for consistent performance improvements.