Loading slide

Loading contents...

[████░░░░░░░░░░░░░░][██████░░░░░░░░░░░░░░░░░░░░░░]8 / 37
<back>next

When an ability seems to switch on

Researchers sometimes saw a startling pattern. Several smaller models scored zero on a task. A larger one crossed the line for a correct answer and suddenly earned points. On the chart, the ability looked off and then on.

This was reported for tasks involving arithmetic, translation, instructions, and code. Researchers called such jumps emergent abilities, meaning abilities that were not visible in the smaller-model measurements but appeared in larger ones.

But the chart can fool us. Imagine a maths test scored only as pass or fail. A learner may improve from 1 correct step to 4, then finally reach all 5. The score stays at zero until the last step, even though the underlying progress was gradual. Researchers have shown that some apparent jumps shrink or disappear when the metric gives partial credit.

Other studies still find threshold-like changes under some conditions. So emergence names an observation and an open measurement problem, not a settled law that size magically creates abilities from nowhere.

# citations(2)↓
  1. [1]arxiv.org
  2. [2]arxiv.org