Building AI Agent from first principles with a durable understanding that frameworks alone cannot, This post structures a learning path that separates the core theory from the "from scratch" implementation, using resources that are independent of major agent frameworks, such as LlamaIndex.

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 theorem serves as the fundamental existence proof for deep learning, demonstrating that a standard neural network with sufficient neurons has the theoretical capacity to approximate any continuous function. It assures engineers that the architecture itself is capable of modeling any complex real-world pattern, shifting the challenge from "is this representable?" to "how do we find the parameters to represent it?"

AI is not just a consumer of linguistic data; it is a producer of new philosophical questions. It forces us to be more precise about what we mean by "meaning," "understanding," and "thought," turning centuries of abstract debate into a pressing, practical challenge.

My first ASR model fine-tuning hit really hard at an error by gradient checkpoint which draw my attention to this topic. Gradient Checkpointing (also known as activation checkpointing) is a memory-optimization strategy used during the training of deep neural networks. It allows you to train significantly larger models (or use larger batch sizes) on limited GPU memory by trading computation time for memory space.

Why would AI work? Certainly technological advancements and market conditions have their contributions. But it is the philosophical justification presented by one of the greatest philosophers, David Hume, that has made it inevitable which this post shows by examining and linking his most phenomenon work - "An Enquiry Concerning Human Understanding"

If the Universal Approximation Theorem expresses the unprecedented power of AI, the Halting Problem poses the opposite - its computational limit.