Modern LLMs have tiny GPTs hidden inside them (invertedpassion.substack.com)

Recent explorations reveal that modern language models (LLMs), such as GPT-2 and Qwen, may contain internalized models of previous LLMs, fundamentally enhancing their token prediction capabilities. An informal study examined this hypothesis by prompting GPT-2 to generate text continuations and subsequently evaluating Qwen's responses. The results suggested that when Qwen was tasked with completing GPT-2's partial outputs, its completions bore more resemblance to GPT-2's style than its own independent continuations, indicating an implicit awareness of earlier model outputs. This finding carries significant implications for the AI/ML community. If LLMs can effectively model other LLMs internally, it opens the door to advanced metacognitive abilities, allowing models to simulate outputs before generation. Such capabilities could improve uncertainty estimation and enhance the robustness of text generation. Furthermore, the potential for self-representation among models could lead to richer, more nuanced interactions in future AI applications, prompting further research into the architectures and training methods that could foster these unique internal models.