Every major ai model failed this test

Added: 2026-05-18

[00:00:00]Research shows that every major AI model was just given a test that 135 regular humans solve with no prior knowledge.

[00:00:06]And it's worth paying attention to.

[00:00:08]Reason number one, the test is designed so that you cannot cheat. It basically drops an AI into a game environment that has never existed before. No instructions, no goals, no languages.

[00:00:17]And the AI has to figure out what winning even looks like. Number two, researchers tried building a custom setup engineered for individual puzzles.

[00:00:24]scores jumped as high as 97% on those puzzles. But given the same puzzle setup, it wasn't built for, that score crashes to zero. The intelligence was never in the model. It was in the engineers who studied the puzzle first.

[00:00:37]Number three, without that kind of custom engineering, the best any specialized system has ever scored is 12 and a half%. There's a $700,000 price pool for hitting 100% and it has not been claimed since March. It's like hiring somebody who aces every practice test but freezes the moment the exam looks different. That's every enterprise AI deployment right now. And engineers, they study the workflow, they build the logic, they encode the answers, and then AI runs it. So companies right now think that they're buying intelligence, but they're actually just buying engineering.