Let the users grade it
Last time, we saw AI tests can leak. So let's drop the test and let the users be the judges.
BLIND MATCH
Two anonymous answers
You type a question and get two answers from two unnamed AIs. After reading, you click the better one. Only then are the names shown.
ELO SCORE
Chess-style scores
Each vote is like a chess game: the winner gains points, the loser loses some. Beating a strong opponent earns more. This is called an Elo score, like chess player rankings.
LEADERBOARD
Millions of votes
Add up millions of votes like this, and you get a leaderboard. The best-known site is LMArena, and anyone can try it.
PROS
Close to real life
The good part: the questions are real questions from real users, new every day, so it's very hard to cram for.
CONS
Long and pretty isn't always right
The downside: voters tend to prefer long, nicely formatted, confident answers. A short answer that's correct can sometimes lose.
PART 4
Write your own test
But a general leaderboard still isn't your work. Write your own test. See you in part 4.
This article is based on the video The AI arena from the Mark học AI channel. Watch the video (in Vietnamese) to see the animations.