Frontier models

Frontier models is a generic term to refer to the biggest and best LLMs in existence. They tend to cost tens or hundreds of millions of dollars to train¹ and sit at the top of the LMSYS Chatbot Arena Leaderboard.

The way frontier models are developed starts with applying scaling laws to predict the performance of a new model. Then the model is designed, the training dataset is created, and LLM training at scale is performed.

However, the precise capabilities of the resulting frontier model are not know a priori; they’re training “a mystery box.”² This is a bit scary, because it reflects a reality that we don’t know what the next big model will be capable of until it is trained and carefully examined.

Footnotes

1. GPT-4 is estimated to have cost $78MandGeminiUltra$191M. See AI Index Report 2024 – Artificial Intelligence Index (stanford.edu) ↩

2. Eight things to know about large language models ↩