What is an 'internal OpenAI model' and why is DeepSeek catching up to OpenAI?

The article examines the nature of 'internal' AI models and their development lifecycle. The author debunks conspiracy theories about hidden AGI, explaining that frontier labs create an ecosystem of models ranging from research checkpoints to distilled production versions. The focus is on knowledge distillation, where powerful teacher models are used to train smaller, more efficient student models. This allows companies like OpenAI to optimize the balance between computational power, cost, and latency for end users. The author emphasizes that research and production models serve different purposes: the former maximizes capabilities, while the latter balances intelligence, latency, and GPU costs. This approach to optimization and scaling explains why competitors, including DeepSeek, are able to rapidly close the technological gap by effectively utilizing synthetic data training and distillation methods.