Emergent Abilities
Capabilities that appear in large AI models but are absent in smaller ones — behaviour that seems to switch on abruptly as models scale, rather than improving smoothly.
Emergent abilities describe skills that show up only past a certain model size: multi-step arithmetic, instruction following, in-context learning from examples, and chain-of-thought reasoning are often cited. In smaller models the benchmark score sits near zero; in larger models it jumps, sometimes within a narrow range of scale.
The phenomenon is debated. Some researchers argue emergence is partly an artefact of how we measure: benchmarks that award credit only for fully correct answers turn smooth, incremental improvement into an apparent cliff. Score the same outputs with partial credit and the curve looks gradual.
Why it matters:
- Forecasting: if abilities appear abruptly, you cannot extrapolate safety and capability from small runs.
- Safety: a new capability may arrive without warning in a larger model.
- Evaluation: tests must be sensitive enough to detect partial progress, not just pass/fail.
The practical takeaway is that scale changes behaviour in ways that are hard to predict, which is why labs evaluate models at every size and why benchmark design is treated as a scientific problem, not a marketing exercise.