Who will release the next Large Language Model that has an LMSYS Arena Elo of at least 1334; 75 points better than the current leader?
As of the 22nd of April, 2024 there are 4 models with an Arena Elo between 1249 and 1259 according to the LMSYS leaderboard: 3 versions of GPT-4 and 1 version of Claude 3 Opus. The highest rated GPT-3.5-Turbo version has an Elo of 1119, 46 points behind the lowest GPT-4 version (0613 for both), while the 0314 versions of these models have an Elo gap of 82 points. Thus, a 75 point gap would represent a breakthrough and a new generation of LLMs.
Elo will be evaluated 1 week after the model enters the leaderboard. If 1334 is within the top contender's confidence interval, l'll wait 1 additional week and resolve based on the Elo then. If multiple models meet the criteria, the earliest release date is the winner.
@Sss19971997 That's right. If two models meeting the criteria are released in close succession, the one that was released first will be the winner (not the one that has the higher Elo).