2026 (7)

July (5)

• How LLMs go from base models to assistants July 23, 2026
• Scaling up a DiT: replicating Peebles & Xie, then optimizing the training loop July 23, 2026
• A toy assistant: SFT and DPO on my from-scratch GPT-2 July 16, 2026
• Faster Qwen-Image-Edit 2511 by caching KV July 9, 2026
• Diffusion on a 2D spiral July 2, 2026

June (2)

• Reproducing GPT-2 Small: training, evaluating, and watching induction heads form June 25, 2026
• Replicating rank-1 emergent misalignment in Qwen2.5-14B June 18, 2026

2025 (1)

July (1)

• Transformers Don't Need LayerNorm at Inference Time July 23, 2025

2024 (1)

October (1)

• ARENA Capstone: Hyperparameter tuning for MELBO October 5, 2024