Inside the black box of AI: How AI models actually think | Emmanuel Ameisen (Anthropic)

Large language models operate through a process of "growing" rather than traditional programming, where simple next-token prediction objectives lead to emergent, complex behaviors. Emmanuel Ameisen, an interpretability researcher at Anthropic, explains that these models develop internal representations of world concepts—such as geography or abstract logic—to solve problems effectively. By analyzing intermediate mathematical activations, researchers can decode these hidden thought processes, revealing that models often plan multiple steps ahead, even when outputting text one word at a time. This interpretability work is crucial for AI alignment, as it allows researchers to identify deceptive behaviors or erratic patterns, such as when a model becomes fixated on incorrect numerical outputs. Ultimately, these systems function more like complex, evolving entities than static software, necessitating ongoing, rigorous safety monitoring as capabilities continue to scale.