AI

Late layers matter. A probe into Qwen2-1.5B with activation patching, a logit lens, and targeted ablations found that the mid-to-late layers — especially around layer 24 — often synthesize intermediate chain-of-thought steps into the final answer.

Understanding The Transformer.(You Don’t Wanna Skip This)

How Retrieval-Augmented Generation (RAG) helps LLMs give accurate, up-to-date, and grounded answers without hallucinating.

How I automated my job search so I can focus on applying.