Loading slide
Here is the pattern that unsettled people. A small model cannot do arithmetic at all. Make it bigger, step by step, and it still cannot. Then, past a certain size, it can. Not gradually better. Off, then on.
The same happened with other skills. Small models could not translate a language they were never specifically trained on, then past a size they could. Small models could not follow a multi-step instruction, write a working snippet of code, or lay out their own reasoning, then past a size they could. And nobody had trained them to do any of it on purpose. These abilities arrived as passengers, riding along with sheer scale.
Researchers borrowed a word for this from physics and biology: emergent abilities. The word names something we already met in the going-deeper module. A single water molecule is not wet. Wetness is not hiding inside one molecule, waiting. It only exists once you have enough of them together. The property belongs to the crowd, not to any member of it. Stack enough simple parts and something genuinely new appears that was in none of them.
The uncomfortable twist is that these abilities switched on at sizes nobody could predict in advance. The smooth curve told you a model would get better at what it already did. It said nothing about when a brand-new skill would suddenly click into place. You often could not tell what a model could do until you built it and asked.