5.B Snake

Here you can challenge two artificial intelligences in Snake.

The goal is simple: make the snake eat as many apples as possible while avoiding crashing into itself. Even if you’ve never played before, a couple of rounds will be enough to understand how to move and start scoring points.

The two artificial intelligences you can compete against may look identical, but they differ in how many games they were trained on: the first was trained on 100 games, the second on as many as 10 million.

But what does it mean that it “was trained”?

When an artificial intelligence learns to play Snake, at first it makes seemingly absurd mistakes: for example, if the apple is straight ahead of the snake’s head, it might turn right. It doesn’t know what a two-dimensional space – like the one the game takes place in – is, what “right” or “left” mean, or what it means to “avoid an obstacle” or “try to reach the apple.”

However, every game produces a score, and that number becomes the data the machine tries to increase—that is its goal. Based on the results it obtains, game after game, the system gradually adjusts its strategy, keeping the decisions that tend to yield more points and discarding those that yield fewer. By repeating this process many times, the playing strategy improves, and the artificial intelligence becomes increasingly skilled… until it is almost unbeatable.

As you play, observe the strategies of the two AIs: the second is clearly more “intelligent.” Perhaps enough to make us think that, in some way, it has “understood” something. But can we really say that these two AIs have truly understood anything?