As artificial intelligence technology continues to evolve, the industry's focus on the underlying architecture of foundational large models has never waned. Recently, an X user who has long tracked developments in large models revealed that OpenAI is currently advancing a new large model project codenamed Doug. According to the information, Doug will be its largest pre-training project to date and is expected to officially launch by November at the latest.
Looking back at the development trajectory over the past two years, it can be seen that since the release of GPT-4o in May 2024, OpenAI has not launched a major update to its main base model. During this period, the enhancement of its model capabilities has mainly been achieved through post-training, large-scale reinforcement learning, and inference-time computing, such as the相继 released o1, o3, and the integrated GPT-5. However, industry research firm SemiAnalysis points out that relying for a long time on post-training and inference computing, without a comparable generation leap in the base model, will eventually face the challenge of diminishing returns.
Faced with significant pressure from competitors such as Google's Gemini 3, OpenAI has begun to restart and adjust its base pre-training strategy. Previous reports indicated that the company successfully validated the effectiveness of the training repair solution through a model codenamed Garlic after solving a series of pre-training technical challenges. The highly anticipated Doug is seen as an extension and scaling-up of these improvement methods on a larger scale.
The industry generally believes that if the Doug project progresses smoothly, it not only means that OpenAI is returning to the track of large-scale base pre-training but also marks the beginning of a new round of model competition. Once the base itself makes a substantial leap, combined with its mature reinforcement learning system, the future performance ceiling of AI may open up greater room for imagination.
