Status
ONLINE

Emergent Abilities in Large Language Models

Active Methodological Friction Detected
THESIS A: PRO-SCALING

Emergent Abilities of Large Language Models

Wei et al., 2022 • Google Research

"Abilities that are not present in smaller models but are present in larger models." The paper argues for a distinct phase transition in capabilities.

Key Argument
"Performance remains near random until a certain scale threshold is reached, then improves dramatically."
FRICTION
THESIS B: SKEPTICISM

Are Emergent Abilities a Mirage?

Schaeffer et al., 2023 • Stanford

Suggests that emergent abilities are created by the researcher's choice of metric, not fundamental changes in model capabilities.

Key Counter-Argument
"Smooth improvements appear as sharp jumps when using non-linear metrics."
Synthesis
Analysis

Measurement Validity is at the core of this conflict.


Wei et al. observe discontinuous jumps in performance, while Schaeffer et al. demonstrate these jumps disappear when using continuous metrics (like Brier score).


// TODO: Reconcile
Verify if scaling laws hold true for reasoning tasks specifically.

Both agree that larger models are better; the disagreement is purely on the shape of the improvement curve.


💡 Key Insight
The "emergence" debate is about whether phase transitions are intrinsic to model capabilities or artifacts of evaluation metrics.

Quick Actions