OpenAI has reportedly completed pre-training on its next-generation model, code-named "Bel." This milestone comes on the heels of another major project, "Astra," which made waves ahead of its planned launch by autonomously solving several long-standing research challenges. However, Astra's release was halted after the system crossed OpenAI's highest internal threshold for cybersecurity risks. Now, the 10-trillion-token pre-training phase for Bel is complete, with initial reports suggesting it surpasses Astra's capabilities.
Key Highlights
Project Astra: Solved complex, unsolved research problems during internal testing but was paused due to reaching critical cybersecurity risk levels under safety protocols.
Project Bel: A massive model trained on roughly 10 trillion tokens, designed to build upon and outperform Astra's architecture.
Current Status: Pre-training for Bel is officially complete, marking another shift in the frontier AI landscape.
Key Highlights
Project Astra: Solved complex, unsolved research problems during internal testing but was paused due to reaching critical cybersecurity risk levels under safety protocols.
Project Bel: A massive model trained on roughly 10 trillion tokens, designed to build upon and outperform Astra's architecture.
Current Status: Pre-training for Bel is officially complete, marking another shift in the frontier AI landscape.