← AI Terminology

Emergent Abilities

Emergent abilities in LLMs are capabilities that appear suddenly and unpredictably at certain model scales — absent in smaller models but appearing sharply once a threshold parameter count or training compute is crossed.

The term was popularised by Wei et al. (Google, 2022) studying GPT-3-class models.
Why It Matters in AI
Emergent abilities challenge the assumption that scaling produces smooth, predictable improvements. If capabilities can appear discontinuously at scale thresholds, then small-scale evaluations cannot predict what a larger model will be able to do — raising both exciting possibilities (unexpected capabilities) and safety concerns (unexpected dangerous capabilities appearing at scale). They also motivate continued investment in larger frontier models.
Key Points
Aspect Description
Examples Multi-step arithmetic, chain-of-thought reasoning, BIG-Bench tasks, few-shot learning
Contested Schaeffer et al. (2023): emergence may be an artefact of discontinuous metrics, not the model
Definition Task performance near random at small scale, then jumps sharply above random at larger scale
Monitoring Responsible scaling policies include evaluations for dangerous emergent capability thresholds
Scale thresholds Typically appear between 10B–100B parameters — varies by task
Safety implication Dangerous capabilities (bioweapon synthesis, cyberattack) may also emerge unpredictably
Simple Analogy
Ice melting: water's temperature rises gradually (smooth scaling) until it hits 0°C and suddenly becomes liquid — a qualitative phase transition. Emergent abilities are the phase transitions of AI: you scale parameters gradually, then at a threshold, a new capability suddenly appears that wasn't there at all at smaller scale.
Common Usage Examples
  • Chain-of-thought reasoning: near-random at 7B parameters, sharp improvement at 100B+
  • BIG-Bench: 200+ diverse tasks; many show emergent behaviour on GPT-3 scale models
  • Arithmetic multi-step: GPT-3 struggles with 3-digit multiplication; GPT-4 handles it reliably
  • Responsible scaling: Anthropic's ASL evaluations check for emergent dangerous capabilities before each new model class
  • Debate: are claimed emergent abilities real or metric artefacts? (Schaeffer et al., NeurIPS 2023)
Summary
In short: Emergent abilities are capabilities that appear suddenly at scale thresholds — not gradual improvements but qualitative jumps that are hard to predict from smaller-model experiments.