CrackedAI

CrackedAI offers a collection of machine learning coding problems that users can solve locally on their own hardware. The platform provides 571 problems implemented in both PyTorch and JAX, covering a range of topics from basic neural network primitives like softmax, relu, and batch normalization, to frameworks such as autograd, vmap, einsum, and jit. Users are tasked with implementing advanced techniques from research papers, building models like transformers, and working through end-to-end projects involving MLPs, CNNs, and RNNs.

The tool is designed for those interested in hands-on practice with machine learning concepts and frameworks, supporting both PyTorch and JAX. CrackedAI allows users to browse problems by category, including primitives, frameworks, end-to-end models, and research topics such as RoPE, GQA, LoRA, and MoE. Each problem is intended to be solved locally, ensuring that all code runs on the user's own machine and hardware, with the platform explicitly stating that it does not access or see user code.

The setup process is described as taking approximately 30 seconds.

As a platform for practicing and reproducing machine learning techniques, CrackedAI serves learners and practitioners seeking to deepen their understanding of PyTorch and JAX through practical coding exercises executed locally.