Chai AI is a platform that's working to advance conversational generative artificial intelligence by connecting language model creators and chat AI fans. The idea is to crowdsource the AGI breakthrough. The company has 12 engineers working on the project, and they've made some progress so far, serving more than 1 million people daily from offices in Palo Alto, CA.
Chai AI has done some research, including work on reward models, rejection sampling, blending and reinforcement learning with human feedback. That work has produced a proprietary dataset of more than 4 billion user-bot messages and trained language models tuned for entertainment purposes, with a compute cost of more than $16 million.
The company also lets people train and submit their own models, which are then used to chat with millions of people, in the hopes of producing more compelling conversations. Chai AI has scale, with more than 500,000 chat AI creators and 4 million people chatting with them. The ecosystem is bolstered by more than 100 foundational model creators who've deployed more than 10 trillion parameters and served more than 100 billion messages.
Chai AI also has a developer platform called Chaiverse where developers can train and submit their models with the Python programming language and compete for cash prizes. The platform supports different architectures, including LLaMa and Mistral, and has a simple submission process. Models are judged on user enjoyment and safety, and a leaderboard ranks the top-scoring models. The company has paid out $132,000 in prizes so far, with a total $1 million prize pool.
Chai AI offers a free app for users, with premium subscription options to remove ads and get unlimited messages. The company doesn't share user-generated data publicly unless users specifically consent, and it's transparent about that through its EULA.
Although Chai AI is focused on entertainment, it's aware of model safety issues and has a lot of checks before models are deployed. It also encourages people to help build a community of people working on safe and engaging AI.
Published on June 12, 2024
Analyzing Chai AI...