Exploring deep tech, AI systems, and software architecture.
A code-backed investigation into Mechanistic Interpretability. Why high cosine similarity in Sparse Autoencoders doesn't guarantee causal impact in LLMs.
A reproducible, code-backed investigation into the limits of Mechanistic Interpretability. Why recovering features with Sparse Autoencoders (SAEs) doesn't guarantee causal impact in LLMs.
Bridging condensed-matter physics and machine learning. A deep dive into how Tensor Networks and Matrix Product States (MPS) are used to compress production LLM weight matrices.
A strategic and technical deep dive into digital sovereignty, overcoming international API barriers, and why building a native, localized AI infrastructure is critical for the Algerian market.