wtf is mingpt?

This project, emilwallner/mingpt, is an educational experiment in building a small AI model from scratch. Rather than using a massive, off-the-shelf AI system, it breaks down the underlying mechanics so you can see how these models actually learn. The specific focus here is teaching an AI to solve math problems and, notably, to recognize when it has made a mistake. At a high level, the project trains a mini version of the technology behind modern text AI models. It starts by generating a dataset of math problems, then teaches the model to predict the answers. The interesting twist is a feature the creator calls "self-aware tokens." After the model attempts to solve a math problem, a second process checks if the answer is right or wrong, and the model is then trained to predict its own accuracy. It essentially learns to grade itself, using its past mistakes to inform future attempts. The work is aimed at AI learners and researchers who want to understand the building blocks of artificial intelligence without getting lost in complex, production-level code. For example, if you are a founder or a beginner trying to grasp how an AI can be trained to self-correct, this project provides a tangible, stripped-down example. Instead of treating the AI as a black box, it shows the step-by-step process of data preparation, training, and evaluation. The project is built entirely in Jupyter Notebooks, which are interactive documents popular in data science for mixing code, visualizations, and explanations. This makes it easy to read and follow along. The README consists mostly of a to-do list, indicating the project is still a work in progress. The creator is actively experimenting with ways to measure the model's learning limits and adjust how much computing power it uses at different stages of training.