# tiny-prl > A documentary of RL training on my 8 GiB laptop, from memory guards to model-specific kernels and the learning experiments that did not solve the task. ## Docs - [tiny-prl](/tiny-prl-docs/index): Making reinforcement learning run on my 8 GiB laptop, one measured change at a time. - [The laptop budget](/tiny-prl-docs/guide/laptop): Why fitting the trainer alone was never the full memory problem. - [One RL update](/tiny-prl-docs/guide/one-update): Follow a prompt through Verifiers, vLLM, the orchestrator and the trainer. - [Moving the policy](/tiny-prl-docs/guide/weights): What a filesystem broadcast and a CUDA IPC refresh actually transfer. - [The kernel path](/tiny-prl-docs/guide/kernels): Why the decode kernels are model-specific, and why exactness came before adoption. - [Credit assignment](/tiny-prl-docs/guide/advantages): The FP32 tie that looked like a learning signal, and the editable function used today. - [The hillclimb](/tiny-prl-docs/experiments/hillclimb): A chronology of constrained execution, kernel work and the attempts that did not survive measurement. - [The wall-time experiment](/tiny-prl-docs/experiments/wall-time): What the 11.74% result measured, what it held constant and what it cannot establish. - [The learning experiments](/tiny-prl-docs/experiments/learning): Zero wins, lower SFT loss and why a dense reward still did not establish task success. - [The installed command](/tiny-prl-docs/guide/command): Keep a small runtime, edit algo.py and remove local model files only after a verified upload. - [Code map](/tiny-prl-docs/reference/code-map): The files to read when changing startup, credit assignment, serving or publication. - [Evidence and open work](/tiny-prl-docs/reference/evidence): What can be checked publicly, what was archived and what still needs a result.