Emergent capabilities are abilities that appear in AI models at scale that nobody explicitly programmed in. A small language model can barely string a sentence together. Make it ten times larger, and it still can't do arithmetic. Make it a hundred times larger, and suddenly it can solve math problems, write working code, and translate between languages it was never specifically trained on. These abilities don't appear gradually — they're effectively absent below a certain model size and then show up abruptly, often surprising even the researchers who built the model.

The landmark paper from Wei et al. at Google documented dozens of these abilities across foundation models, including multi-step reasoning, word unscrambling, and chain-of-thought problem solving. The pattern was consistent: flat performance for smaller models, then a sharp jump once a scale threshold was crossed. There's genuine academic debate about whether this is a true phase transition in capability or a measurement artifact — Schaeffer et al. argued convincingly that some "emergent" behaviors look less dramatic when you change how you score them. The honest answer is probably both: some capabilities genuinely appear abruptly, and some we were just measuring badly.

For strategic planning, here's what matters. Emergent capabilities mean next year's model might suddenly do things no one predicted. That makes rigid three-year AI roadmaps unreliable. The company that builds flexible architecture — modular systems that can swap in new model capabilities as they appear — will outmaneuver the one that bet everything on today's limitations being permanent. Plan for surprise. Budget for iteration. Treat your AI strategy as a series of options, not a fixed blueprint.

Further reading:

• Emergent Abilities of Large Language Models (Wei et al., 2022) — the foundational paper documenting dozens of abilities that appear abruptly at scale across multiple model families.
• Are Emergent Abilities of Large Language Models a Mirage? (Schaeffer et al., 2023) — the important counterargument that some emergent abilities may be artifacts of metric choice rather than true phase transitions.