Status update on Errloom, my reinforcement learning library / toolkit. It is coming along quite nicely and we will soon start training some concepts of super-intelligence. Pictured below is Errloom's home screen. A single command takes an input training script and deploys to cloud compute of your choice: displays a list of offers in an interactive TUI list to choose, connects through SSH, sets up the whole environment, clones errloom, runs vllm server, and launches errloom to train—local errloom acts as manager & dashboard, opens WanDB in your browser when training starts, downloads logs and checkpoints intermittently, and displays the remote process trace. No unnecessary jupyter or other superfluous steps. Write the prompting environment of your dreams, fire one command and leave the computer for lunch. The helper message demonstrates how I have set up the progressive development workflow. Since I am operating with zero budget and resources, I need to make every penny count. Forget renting cloud compute just to develop this library. We need it to walk through most of the program circuit on a beefy home setup. To achieve this, I have devised a number of commands and feature flags for progressive testing, building the entire pipeline brick by brick until you have full 100% certainty and awareness that GRPO and backprop is going to receive exactly the context windows and token masks you want.