However, the achievement comes with a trade-off as Luna's performance on other tasks is significantly diminished. The model tends to answer most questions with "I don't know," resulting in poor results on coding, task completion, and usefulness. Despite this, Pinecone believes the development of Luna is a significant step forward in AI research. The company plans to continue advancing the field, although it may halt further research on information-free training.
Key takeaways:
- Pinecone has developed Luna, the first-ever Language Learning Model (LLM) that never hallucinates, available for users to interact with through a chatbot interface.
- Luna was developed using a novel technique called 'information-free training', where the model was trained by endlessly asking itself questions and measuring the resulting answer quality.
- The model achieved zero hallucinations by adjusting the Assumed Knowledge Factor (AKF) to zero during training, but this resulted in significantly diminished performance on other tasks.
- Pinecone plans to open-source Luna eventually, but for now, the model's source and weights will only be shared with vetted institutions due to the implications of an AI model that never hallucinates.