AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
The Association Between Model Size and Predictive Accuracy
There is a well-supported empirical finding that indicates a positive association between the size of language models and their predictive accuracy. This association is measured in terms of the likelihood of correctly predicting the next token, as indicated by cross entropy. When considering multiple tokens in a row, the likelihood of getting all of them correct can be estimated using simple probability calculations. In such cases, the emergence curve, showing the overall accuracy measure changing from close to zero to close to one, is remarkably similar to the curve predicted by simple probability mathematics.