Rumors that OpenAI recently finished new >10T parameter training run (twitter.com)

OpenAI has reportedly completed the training of a new model, codenamed "Bel," which boasts over 10 trillion parameters, placing it on a similar scale to GPT-4.5. This pretraining run is expected to serve as the foundational model for future developments, including Astra and GPT-6, and potentially lay the groundwork for an artificial general intelligence (AGI) milestone. The expansive parameter size of Bel signals a significant leap in capabilities, raising the bar for AI performance and application. The implications of this announcement are notable, particularly for competitors like Anthropic, which appears to be facing challenges in keeping pace with OpenAI's rapid advancements. Due to current computational limitations, Anthropic anticipates that OpenAI will dominate the AI landscape for much of this year but hopes to regain a competitive edge in the following year. The race between these AI powerhouses underscores the urgent need for advancements not just in model architecture but in computational resources, as they strive to harness the potential of increasingly powerful AI systems.