Emergence in Neural Networks › Emergence in Neural Networks

Throughout our journey from simple n-gram models to transformer architectures, we've observed a consistent pattern: more data and more parameters generally yield better performance. The scaling laws we explored in Part XXIII formalize this relationship, showing predictable improvements in loss as we increase model size and the resources used to train it. But scaling laws tell only part of the story. Some capabilities do not improve gradually. Instead, they appear suddenly and unexpectedly once a model crosses a certain threshold. This phenomenon, called emergence, challenges our understanding of how neural networks learn and what they can reach.

Emergence matters because it suggests that tomorrow's models might develop capabilities we cannot predict from today's performance curves. If a model shows no ability to perform multi-step arithmetic at 10 billion parameters but suddenly succeeds at 100 billion, what other abilities might be waiting beyond our current scale? This question matters for AI safety, capability forecasting, and understanding intelligence.

This chapter explores the concept of emergence in neural networks. We'll define what emergence means in this context, examine the evidence for phase transitions in capability acquisition, survey examples of emergent behaviors, discuss proposed mechanisms, and engage with the ongoing scientific debate about whether emergence is a real phenomenon or an artifact of how we measure model capabilities.