I am currently exploring underexplored parts of the training stack that enhance the capabilities models acquire during training. I want to understand ML on both a theoretical and practical level, which means reading papers, implementing ideas, and occasionally discovering why something that sounds elegant falls apart in practice. My most recent endeavors include:

• Temporal Regularized Learning: Self Supervised Learning Local In Time And Space by Davide Wiest. [2025] Code Paper
• pptrain — a PyTorch library for pre-pretraining language models on synthetic tasks [2026] Code
• Closed-Form Initialization — exploring whether neural networks, including transformers, can be initialized analytically [2026] Code
• Gradient Routing for Continual Learning — testing whether sparsity can reduce catastrophic forgetting [2026] Code

Outside of work, I go on long walks when I am stuck on a problem or visit the library for inspiration. Philosophically I lean toward anti-essentialism and structuralism, with bundle theory as an interesting framing.

Reach me at: davide.wiest2 [at] gmail [dot] com