ChatGPT Writes AI: Building an MNIST Classifier with Keras Using ChatGPT

Since the rise of conversational AI models like ChatGPT, developers have been exploring its capabilities beyond simple queries. The author decided to let ChatGPT build a neural network with Keras to solve the classic MNIST digit‑recognition problem.

ChatGPT initially produced a complete, ready‑to‑run script using dense layers and categorical cross‑entropy loss. The author then asked it to replace the loss with sparse categorical cross‑entropy, which ChatGPT did without breaking the code.

To experiment with convolutional architectures, the author requested a Conv2D layer and a MaxPooling layer; ChatGPT updated the model accordingly. Subsequent prompts guided it to write evaluation code that uses the test set, reports accuracy, and later switches the validation split to 20% of the training data.

The author also asked for visualizations of training and testing loss curves, for code that displays sample images with their labels, and for a summary of the model architecture. ChatGPT provided the requested code and explanations, and even generated a diagram of the model structure.

Finally, the author instructed ChatGPT to save the trained model to disk, create a predictor class for inference, and write a script that predicts labels for ten random test images. ChatGPT also produced a confusion‑matrix visualization for the model’s performance.

All the generated code and visual outputs are shared via a public Colab notebook, allowing readers to reproduce the entire experiment.