New AI benchmark dropped and it shows we're not even close to AGI

[CosmicExplorer]

https://arcprize.org/leaderboard  Humans are at 100% in ARC-AGI-3

AI on ARC-AGI-2 has not yet reached human levels but it's close. However, there's a very strong suspicion that models benchmaxxed (they created similar puzzles/tasks and trained AI on it, for it to then do better at private evaluation ARC-AGI tests) their way to the scores on ARC-AGI-1 and ARC-AGI-2. When a new benchmark like that drops, it's the best moment to see how far away we are in terms of the general intelligence that humans have. If AGI would be achieved, there would be no possibility of a new benchmark on which a gap between average humans and AI is big.

Edited March 26 by CosmicExplorer

[NewKidOnTheBlock]

Makes sense to me, as expected A.I. of course lacks true agency and the very poor agentic inteligence performance in ARC-AGI-3 reflects that. I've tried one of those retro games and passed all 9 levels without too much problems, so the the fact none of those A.I.s could get even 1% is really showing where it's lacking. Good thing about it is, that we have more time to focus on a various security and ethical concerns, as well as having more time to meaningfully contribute to society and A.I. development if one chooses to go that direction

[CosmicExplorer]

4 hours ago, NewKidOnTheBlock said:

I've tried one of those retro games and passed all 9 levels without too much problems

Same. Those games are really easy for me. It really shows how bad AI's true understanding of things is (if there's any).

For anybody interested in trying those games https://arcprize.org/tasks/ls20

Edited March 26 by CosmicExplorer