A dish of neurons may have taught itself to play Pong (badly)

>> That said, there are a host of caveats. Even the best behaving systems didn't play Pong all that well. The best performing control systems, which were likely to be moving the paddle randomly, consistently outperformed the average trained neural networks, and the worst performances by trained systems were worse than the average control neural networks. So, while the boost in performance was statistically significant, you couldn't necessarily identify a trained, functional system just by watching it play Pong.

The second thing is that there were a lot of individual tests done by the researchers: different measures of performance, different amounts of training, multiple controls, and so on. As such, you'd expect some good results to pop up purely due to chance. So, what you have to look for here is whether the good-looking results consistently point in the same direction.

That consistency appears to be there. The effect was seen in both human and mouse neurons, and several measures of success all moved in parallel: average length of rallies and total number of rallies with at least three bounces off the paddle went up, while the number of "aces" where the paddle never touches the ball went down. Giving no feedback when a game ends with the ball crossing the end line produced performance that was intermediate between trained systems and control systems. Combined, these suggest that there's a real effect there, even if the individual tests produced fairly weak results.

Even humans can't beat pong: https://larsdu.github.io/PongRust/