LLM Basics

LLM fundamentals and their implications for end users of AI-Native Applications

Input and Output

Large Language Models (LLMs) take data (e.g., text, image, audio, etc.) and output text (e.g., SQL queries, code, natural language).

Above is the simplest description of what is happening, but there are details we are missing here:

• How will we know when it outputs an SQL query, or blob of text in English vs code?
• Can it output gibberish?
• Can it just stop giving answers?
• How will I execute the SQL query it gives?

Nothing is Certain

Before we scare you about uncertainty with LLMs, consider this scenario: You ask your friend “What is up?”, are you certain what they will answer? No, but you know what to expect, the kind of answers to expect.

LLMs have been trained to be predictable in intent and serve some common use cases like coding, conversing with humans, helping out humans with good intent.

But that being said, the fact is:

There is no guarantee of what an LLM will output. No matter what applications built on top claim, nothing is 100%

LLMs mimic Humans (Can we trust LLM?)

LLMs have been trained on human-generated content, and they have been built to mimic the human brain in thinking.

Don’t worry, the best researchers don’t understand how LLMs think, they can’t even agree on whether it pretends to think or actually has intelligence. This is a rabbit hole, don’t go into that. We should just interact with them and learn what to expect.

Most people reading this must already have a good idea of what to expect when conversing with ChatGPT. Remember, ChatGPT is an application built with a lot of features in the backend and not just a bare LLM. A lot of engineering has gone into making it remember your preferences and be useful in daily tasks.

But would we want an LLM to suggest SQL queries on complex data (different tables, databases, DBMS)? Not really, unless there is a specially curated application for it (think Incerto).

To gain trust in an application (built on LLMs) you must use it on dummy data and then, decide for yourself

AI-Native Applications

The word gets thrown around a lot, but it means the core features of products are dependent on LLMs (e.g ChatGPT, Claude, Perplexity). As a user, we should have the skepticism of uncertainty of LLMs causing uncertainty in the end product — which is always possible.

AI-Native Applications always have these features:

• They take dynamic input from user chat, system, database, etc.
• They transform the input and route it to the LLM
• They process the output of the LLM and depending on its intelligence take actions (show text, do SQL queries, ask clarifying questions, etc.)

Always be clear of what feature is LLM powered and never blindly trust.

Incerto is also an AI-Native application made for interacting and managing Databases. Interact with it and see if you can trust it or not.