This is a re-implementation of the original Karpathy's Atari-Pong script ported to TensorFlow 2.
The approach here is a bit different from Karpathy's; while it still uses Policy Gradients, the AI trains through Gradient Ascent where cost function is the expected reward for running the current policy. Please refer to the spinningup docs from OpenAI for a thorough discussion on details.
Here's an example of the AI playing after training for roughly 15k episodes; green player is the trained AI:
The trained model weights are located in the file model.pckl.