Source @synthwavedd posted on X today (August 26), claiming that OpenAI has completed pretraining its next-generation model, internally codenamed Bel. It is the direct successor to the Doug model, with total parameters exceeding 10T.
The report says OpenAI hopes to develop Bel into the foundational model after GPT-6, potentially even making it a model that reaches the threshold of artificial general intelligence (AGI).
Artificial general intelligence, or AGI, is a hypothetical intelligent system capable of learning, understanding, and performing any intellectual task that humans can. Unlike current narrow AI, which can only handle specific tasks, AGI would possess cross-domain generalization and reasoning abilities, along with the capacity to learn autonomously and adapt to unfamiliar environments.
ITHome translates the source's views as follows:
Exploration of these models could put pressure on Anthropic. OpenAI believes that Anthropic has no effective response to Astra, which is about to be publicly released. Based on what I understand, that is indeed the case, mainly because of compute constraints. Anthropic is psychologically prepared for the situation to favor OpenAI for most of the remainder of this year, but it believes it can regain the lead early next year.