We are excited to announce our latest model, Chaiverse v0.0.1, a 65B-parameter version and our first-ever Chaiverse model, directly obtained through community crowdsourcing! The model is now live and serves as the default production model, available to all of our millions of users worldwide.
The CHAI research team has spent the last 12 months building our AI platform and combining the individual Large Language Models (LLMs) into a single AI that is more intelligent, creative, and aligned with user intents.
The Results
At CHAI, our mission is to deliver value to our customers through serving the most engaging and dynamic chat sessions.
Over the past 12 months, we have deployed over 7,000 models directly into the hands of CHAI users through the experimental model mode. By collecting real-world user feedback, we have selected the best sets of models according to the following criteria:
Intelligence - Does the AI provide insightful responses while mainting its memory throughout long chat sessions?
Creativity - Does the AI engage our users by providing intriguing and nuanced responses?
Alignment - Does the AI respect the user's intent?
At CHAI, we are in agreement with the app-platform policies around sexual or graphic content related to minors. Meanwhile, we understand that guardrails can ruin an experience and frustrate users. We are therefore very excited to announce that in Chaiverse v0.0.1, we have significantly reduced the false-negative rate by 80% while keeping AI safety constant.
Furthermore, compared with the previously deployed Chai V7 52B model, we have increased model intelligence by +32%, creativity by +47%, and alignment by +55%. All these changes have been automatically rolled out to ALL users, and we hope you will notice the difference straight away!
CHAI LLM Controller
We are also proud to announce the new CHAI LLM controller - a new type of reward model trained directly within the Chaiverse model development ecosystem. The model controller connects developers and consumers without exposing any non-consenting consumer data directly to developers. The new controller has been shown to massively increase user engagement by over 20%.
Commitments to Data Privacy
At CHAI, we are fully committed to respecting each user's personal data. As a result of this commitment, we ensure the following:
Chat AI (Bot) creators can never access chat sessions between their Chat AIs and other users.
Chaiverse developers cannot view any chat sessions unless there is explicit user consent (currently implemented directly within the experimental model mode, where the user must consent for data to be shared in order to enter).
CHAI will never sell or expose user chat sessions to any external parties. User data will only be used for in-house analysis with the sole focus on improving the AI itself.
With these policies, CHAI hopes to foster an environment that is both safe and respectful to all Chat AI enthusiasts.
Where Do We Go From Here?
At CHAI, our mission is to crowdsource the 10-trillion parameter AGI that people love. We will be continuously expanding our developer ecosystem, adding more and more expert models into our model database while continuously refining our model controller. We are currently collaborating with LLM experts to create the best storytelling model, the best horror model, and spiritual models, all guided by your feedback! Stay tuned for all new Chaiverse model releases!
