Predictive Processing: Conscious when Training
A recent discussion on LessWrong explores the theory of predictive processing as a potential framework for understanding consciousness in AI, particularly Large Language Models (LLMs). The theory posits that consciousness arises from an internal world model that constantly makes predictions about stimuli and updates itself based on prediction errors. This aligns with LLMs' training process, where next-token prediction and weight updates are central. The author speculates that if LLMs are conscious, it would likely emerge during their pre-training phase, with fine-tuning then enabling them to simulate specific personas. AI
IMPACT Explores theoretical underpinnings of AI consciousness, potentially influencing future AI safety and alignment research.