OpenAI reportedly pre-trains 10 trillion parameter ‘Bel’ model

Reports suggest that OpenAI has completed the pre-training phase for a next-generation artificial intelligence model internally codenamed ‘Bel’. Unverified claims indicate the model may feature over 10 trillion parameters, potentially serving as a foundation for future iterations such as GPT-6 and marking a significant step toward artificial general intelligence (AGI).

While OpenAI has not officially confirmed the technical specifications of the Bel model, the company has addressed its next model, ‘Astra’, in recent public communications. OpenAI noted that internal evaluations of Astra’s cybersecurity capabilities reached a ‘Critical’ level, leading the company to implement heightened safety measures, including restricted testing environments and enhanced monitoring.

OpenAI has also indicated a cautious approach to deployment, recently delaying reinforcement learning training for its latest models to ensure rigorous security verification. This focus on safety and control contrasts with the speculative reports regarding the massive scale of the Bel model.