Emergence: When Abilities Appear Out of Nowhere

Sometimes, a capability that does not exist at one model size suddenly appears at the next size up. A 1 billion parameter model cannot do 3-digit multiplication at all. A 70 billion parameter model can, with careful prompting. That kind of jump is called emergence.

Examples people have documented

• Arithmetic with several digits
• Translating between low-resource language pairs
• Chain-of-thought reasoning becoming helpful
• Following multi-step instructions reliably
• Using tools without explicit training

Why it matters

If abilities appear without warning, you cannot always predict what the next model will or will not do. That makes both opportunity forecasting and safety planning very hard. You cannot test for skills that do not exist yet.

Emergent abilities change interfaces

1. 1Once the model can follow instructions, you do not need task-specific fine-tuning
2. 2Once it can use tools, it stops being a chatbot and becomes an agent
3. 3Once it can write code that runs, it starts connecting to the real world
4. 4Each new capability reshapes what products are possible

“Scale is a staircase, not a ramp.”

How to watch for it yourself

• Keep a personal test set of hard tasks
• Rerun them on every new model release
• Notice which tests go from 0 to usable overnight
• Share findings — the community learns together

The big idea: scaling produces jumps, not just gentle improvements. Emergence is one of the reasons AI keeps surprising researchers, pundits, and users alike.