jev by TypeSafe: The AI That Beats 2048 Without Thinking Hard

TypeSafe's jev is a "System One" model, designed for fast, almost instinctive decisions. A developer put it to the test with 2048, an seemingly innocent little game. The result? It played surprisingly well, without truly "understanding" the game.

jev is a model from TypeSafe, a company perhaps not everyone knows. This model falls under the "System One" category, meaning it thinks fast, almost instinctively. An engineer, cablehead, tested it with the famous game 2048, publishing the results on GitHub on June 11, 2024.

The trick? jev doesn't "reason" like us. Instead of planning complex strategies, it simply generates a sequence of probable moves. Then it executes them, relying on this generative "intuition." It's a bit like an experienced player who feels the right move, without explicit calculations.

The results were remarkable. jev managed to regularly reach the 1024 tile and, in some cases, even the legendary 2048. Not bad for a model that has no clue what a "strategy" or a "score" is, right?

The test's author, cablehead, noted that jev reaches the 1024 tile quite frequently. In fact, it even reached the 2048 tile, the game's maximum, on several occasions. jev regularly reached the 1024 tile and the 2048 tile in several instances during cablehead's tests. This suggests that synthetic intuition isn't so bad after all.

This TypeSafe experiment shows us something interesting. You don't always need an AI that "reasons" deeply to get good results. Sometimes, a fast, generative model that "hallucinates" the next move can be enough. It's a different, leaner approach for certain types of problems.