Reinvent Projects
This page collects my newer deep learning reproductions, research blogs, and open-source style projects.
They are separate from the older school project archive.
Ordered by time from most recent to older. Most of them are planned as Colab notebook blogs that users can copy and run by themselves.
vision-banana-norm-expert
Type: Colab ipynb blog
Start: June 13, 2026
End: Ongoing
Status: Ongoing project
Open source image generative model work to reproduce what Nano Banana can do.
- Users should be able to copy the notebook into Colab and train it by themselves.
- Current progress: trained with the world-renderer dataset for a sanity test.
- Next step: move to a larger dataset.
ppo-lunar-landing
Type: Colab ipynb blog
Start: June 6, 2026
End: Ongoing
Status: Just started
A PPO reconstruction project focused on one practical question: what exactly in PPO makes RL training stable?
- Users should be able to copy the notebook into Colab and train it by themselves.
- Goal: implement every important PPO detail explicitly.
- Current progress: still at an early stage and not much has been finished yet.
ddpm-train-your-MNIST-generator
Type: Colab ipynb blog
Start: May 30, 2026
End: Ongoing
Status: Training finished, blog in progress
A DDPM notebook project about why denoising does not become intuitive until you really inspect how the training process works.
- Users should be able to copy the notebook into Colab and train it by themselves.
- Main idea: do not assume the model learns denoising in an obvious way.
- Current progress: training is done and the instructive write-up is being written.