Sakana AI Labs' LLM Squared Unveils DiscoPOP to Beat DPO

Jun 13, 2024, 02:08 AM

Sakana AI Labs, an AI research company based in Tokyo, Japan, founded by former Google researchers, has introduced a new state-of-the-art (SOTA) preference optimization algorithm called DiscoPOP. This algorithm, discovered and written by a large language model (LLM), aims to automate AI research and discovery. The initiative, referred to as LLM Squared, involves using LLMs to generate hypotheses, write code, and create better ways to train LLMs. DiscoPOP specifically aims to beat the DPO algorithm. The goal is to advance AI by leveraging AI itself, moving one step closer to artificial general intelligence (AGI).