Writings

How GRPO trains a policy LLM to repair real GitHub issues — the full objective built up with one running example, and how this recipe became the industry default.

The complete architecture that changed everything - from positional encoding to multi-head attention.

How self-attention eliminated recurrence and enabled the Transformer revolution.

Simplified attention mechanisms that improved efficiency while maintaining translation quality.

The breakthrough that taught neural networks to focus - exploring the first attention mechanism in seq2seq models.

Understanding the foundation of neural machine translation and how encoder-decoder architectures transformed NLP.