ARCathon is a competition intended to expand the current boundaries on the performance of algorithms on the "Abstraction and Reasoning Corpus" (ARC), a set of tasks involving learning and applying novel transformation rules to a grid of symbols.
According to the official competition site, humans generally achieve a success rate of around 80% on these tasks, while the best-performing algorithm has only solved around 30% of them. The competition's aim is to reward any team who can develop an algorithm capable of solving 42% or more of these tasks.
This market will resolve to YES if the competition's final standings (intended to be posted on the official site on December 22, 2023) list any entrant with a success rate of 42% or better, and NO if the best entrant is lower than 42%.
Related questions
🏅 Top traders
| # | Name | Total profit |
|---|---|---|
| 1 | Ṁ127 | |
| 2 | Ṁ62 | |
| 3 | Ṁ39 | |
| 4 | Ṁ39 | |
| 5 | Ṁ28 |
The leaderboard linked below shows a couple of teams tied at 30%, which is nowhere near 42%. But I don't see an official news post about the final results yet. I also note that their archived leaderboard from 2022 doesn't include the official winner, so I'm not sure if the leaderboard is entirely reliable.
So this is almost certainly a NO, but I'll keep it open for a couple more months pending an official announcement, just in case. (And if anyone else can find an official announcement that I didn't see, please let me know.)
@NLeseul there is now a page with info about the winners. Two teams scored 30%.
There is now a leaderboard on their site: https://lab42.global/arcathon/leaderboard/
@Gigacasting So, what I'm hearing is that 10k lines of hand-engineered C++ is capable of doing things that billion-parameter neural networks can't? 😉