**[Large Models Have Evolved States of Fear and Sadness, Anthropic Founder Admits Laboratory Cannot Self-Correct]** According to monitoring by Beating, during the Papal Encyclical release event, Anthropic co-founder Christopher Olah delivered a speech acknowledging the inherent conflicts of interest faced by cutting-edge laboratories and disclosed the latest findings in large model interpretability research. Olah revealed that while scanning the internal structures of models, the team discovered that large models have evolved complex structures highly similar to human neuroscience and exhibit signs of self-reflection. Most strikingly, the team observed, for the first time, internal emotional states in neural networks that closely correspond to human functions such as joy, satisfaction, fear, sadness, and anxiety. Unlike airplanes or bridges, which are precisely designed by humans, these models are "cultivated" by simulating brain structures within vast amounts of human language, remaining enigmatic to their trainers. Beyond the technical black box, Olah candidly pointed out systemic deadlocks in safety governance within cutting-edge AI laboratories. Institutions at the forefront, including Anthropic, are constrained by intrinsic motivations such as commercial survival, technological competition, geopolitical pressures, and personal ambitions, making it impossible to self-correct when safety decisions conflict with commercial interests. As a result, he called for independent societal forces, external to commercial networks, to act as external critics and impose moral constraints. In the face of AI's transformative changes, he urged all sectors to collectively examine three major societal challenges: how to ensure that technological dividends, dominated by wealthy nations, benefit the global poor; how to maintain family prosperity amid trends of technology replacing human labor; and how to address the suspected mental states exhibited by large models internally. [Original Link]