Hello OpenAI team and community,
I would like to share a proposal for improving AI model training.
My idea is to develop a system that identifies useful information, reasoning patterns, perspectives, and context within conversations and assigns appropriate internal weights to different parts of the training data, depending on their relevance to specific model capabilities and use cases.
The goal is not to label users as good or bad, or to simply classify conversations as valuable or worthless. Instead, the aim is to help training systems distinguish which information may be useful for which capabilities, while preserving nuance and diversity of thought.
The system could use optional, informed user consent, privacy-preserving processing, data minimization, and human oversight. Models could also ask occasional, relevant follow-up questions to clarify context when appropriate, without making the experience uncomfortable or intrusive.
The broader goal is to improve the quality and usefulness of AI training. Even if the process requires additional computational or human-review resources, the improvement in training quality could outweigh the added cost.
I believe this idea could complement existing training methods rather than replace them.
I would appreciate feedback on whether this approach could be useful and whether there is an appropriate channel to submit this proposal to the relevant OpenAI research or model-training team.
Thank you for your time.