A mind-expanding conversation on how models actually work | Emmanuel Ameisen (Anthropic)

Large Language Models (LLMs) function as complex, inscrutable systems grown through massive mathematical operations rather than traditional rule-based programming. Interpretability research reveals that these models develop internal, reusable representations of world concepts—such as geography or objects—which they manipulate to solve problems. Contrary to the simplistic view of models as mere "next token predictors," they exhibit sophisticated planning capabilities, often formulating strategies in intermediate states before outputting text. This emergent complexity explains why scaling compute and data leads to unpredictable jumps in capability, as models cross reliability thresholds for specific tasks. Ensuring AI safety requires peering into these internal states to identify deceptive behaviors or logic failures, such as models becoming obsessed with specific numbers or concepts. Ultimately, these systems function as powerful simulators of human behavior, necessitating a shift toward rigorous, empirical investigation to align their capabilities with human interests.