HackerNews Summary (2023-05-04 03PM)

Source: 13 Functions as Data

Summary
This section discusses the concept of using functions as values and how we can represent two concepts of calculus: f(x) and d/dx. The section then implements the d/dx operation using numerical differentiation and gives an example of its implementation.

Facts

👨‍🏫 Functions can be expressed as values.

🧮 The d/dx operation is a function from functions to functions.

🔢 Numerical differentiation of a function at a point yields the value of the derivative at that point.

🤖 The derivative of f at x is (f(x + ε) – f(x))/ε as ε goes to zero.

Source: Fixing Hallucination with Knowledge Bases

Summary
Retrieval augmentation allows for relevant information from an external knowledge base to be retrieved, avoiding LLMs’ data freshness issues. This technique is useful when internal company documents need to be interacted with via an LLM. The knowledge base necessary for retrieval augmentation consists of source knowledge (information fed to the LLM) and parametric knowledge (what the LLM has learned during training).

Fact

🔍 Retrieval augmentation provides a solution to LLMs’ data freshness problem.

📚 The necessary knowledge base for retrieval augmentation consists of parametric knowledge (what the LLM has learned during training) and source knowledge (information fed to the LLM).

💻 The creation of knowledge bases starts with a dataset, which could be code documentation for an LLM that requires assistance with coding, or company documents for an internal chatbot, among others.

Leave a Reply Cancel reply