CHAI

Consumer AI characters

Users can build their own AI characters and stories for entertainment or social interaction.

Research and development

Developers and researchers can test and deploy advanced chatbot models for conversational AI experiments.

Engagement optimization

Companies can leverage CHAI’s RLHF and PPO models to decrease chat session termination and increase user retention.

Model blending

Teams can ensemble different LLMs trained on different targets to outperform single models like GPT-3 in user retention.

Scalable deployment

Organizations can use CHAI’s custom GPU orchestration and inference systems to support high-demand user bases (e.g., 500K DAU).

Preference-based training

Developers can apply Direct Preference Optimization (DPO) using user preference datasets to boost engagement by 20%.

In-house LLM deployment

CHAI deploys its own 6B and 13B LLMs, adapted for social and engagement purposes, boosting user engagement by +10%.

Reward model training

Users benefit from reward models trained on user signals (e.g., Best-of-4, Reward Model XL with 100 million signals) to decrease retry rates and increase session length.

Proximal Policy Optimization (PPO)

The platform uses PPO reinforcement learning to optimize the base foundation model, reducing the probability of chat sessions ending.

Model blending

CHAI invented model blending—ensembling different LLMs trained on different targets at the conversation level—outperforming GPT-3 in user retention.

Direct Preference Optimization (DPO)

DPO models are deployed across all production blends, boosting engagement by 20% and stacking well with existing reward models.

Custom GPU orchestration

CHAI implements custom CUDA kernels and its own GPU orchestration system to support 500K DAU, moving beyond off-the-shelf load balancing and vLLM.

Scalable GPU reservation

The platform reserves additional GPUs (e.g., 1,000 A100 GPUs) to handle high demand and scaling issues reliably.

Best-of-8 rejection sampling

With increased cluster capacity, CHAI implements Best-of-8 rejection sampling using upgraded reward models for better output quality.