This is part of the down-introspect series on ML/AI on Transformer. I was originally studying LLM on Hugging Face (https://huggingface.co/learn/llm-course) and notice that transformer occupied significant portion of their learning materials so I decided to look into it deeply

This is part of the down-introspect series on ML/AI on Word Embedding. In this post, we will go over the concept of embedding, one of the fascinating ideas in machine learning, and the mechanics of generating embeddings with word2vec.

"Structured Outputs" itself has largely become the standard industry-wide term adopted by open-source libraries, academic papers, and competing cloud providers (such as Google Cloud Vertex AI, Anthropic integrations, and AWS Bedrock). It refers broadly to the developer pattern of supplying a schema (e.g., JSON Schema, Pydantic model, or Zod schema) to guarantee an API or local runner returns valid, structured data instead of free-form text.

Specifically created to solve the massive memory bottleneck and inefficiency of serving Large Language Models in production, vLLM at its core resolves the problem of LLM serving and inferencing. Before vLLM, the bottleneck in serving LLMs wasn't just the model weights themselves, but the KV Cache (Key-Value Cache) - the memory required to store the attention context for generating tokens sequentially. This posts helps us systematically learn KV cache.

We are all familiar with RAG system as a customized LLM trained with private data. How about applying a codebase as the private data as a RAG for coding agent? ADR (Architectural Decision Records) does this for us

The ReAct paradigm bridges the gap between digital reasoning and purposeful action by giving agents a digital form of inner speech. By generating internal thought traces before executing tasks, AI can plan, observe, and self-correct in a continuous loop. This approach moves us away from black-box automation and toward building transparent, deliberate agents that can navigate complex problems with genuine logical integrity.

AI agents are finding their inner speech. By "talking to themselves" to plan and self-correct before taking an action, they mirror the way we use internal monologues to navigate complex choices. This shift from reactive automation to deliberate reasoning marks the evolution of AI from a simple tool into a truly metacognitive collaborator.