
Designing Paper to Code: papers you can read, see as one picture, and run in a few minutes.
Is your image model actually getting better? An evaluation harness for text-to-image diffusion.
Tools I've built this year, mostly for making, testing and understanding generative models.
The code behind my sleep post: six ways for GPT-2 to learn new facts without forgetting old ones, with notebooks and 3-seed results.
An evaluation harness for text-to-image diffusion models: one place to measure whether a model really improved.
torchinfo with interactive model reports: graph, flow, treemap and table views of any PyTorch model.
Tests for generated images: turn a request into yes/no checks, verify each image, retry until it passes.
Model-agnostic 3D medical image-to-image translation, such as MRI to CT, with PyTorch and MONAI.
Resolves GPU-aware Python stacks that actually run on your machine: driver, CUDA and torch versions that fit together.
Agent rollback: every action is recorded with its undo recipe before it runs, so nothing irreversible happens.
Small experiments that test an idea from a paper, written up so anyone can follow and rerun them.
What the brain's memory replay teaches us about catastrophic forgetting, tested on GPT-2 with three seeds.
The script, Colab notebooks and every result from the post, on GitHub.
A pixel-art introduction: meet the player and see the quest log.