How I use AI

Foreword

AI is a very vast term, and I believe first I need to make a clear distinction in the domains of application.

I use AI in three different domains :

• Photography
• IT
• General research

In these domains, I use them a bit differently.

What I do

It’s pretty simple.

In photography, I use AI as a tool at 100%. AI masking is an absolute game changer, making the tedious task of creating a mask - which I’m not the sharpest at - a breeze. It clearly improves my potential there, but let’s place back the important information, as of writing this, it’s not a job so I don’t have any constraints. If it were a job, I would still use it, and find a way to get the following functionality :

Sketch2img

Why sketch2img ? I spend a lot of time composing stills, when I’m not out and about hunting for wildlife and thinking about composing a photograph placing an animal as its center. The first, I could totally control, the second is harder. But using sketch2img could allow me to project an idea on paper, photograph it, and then “fill in the gap” to verify that an idea is viable. Then, I would make the image.

It’s important to explicit that this is not GenAI here, it’s a different kind of AI.

In IT, I have multiple ways, but to sum them :

• I have a rough idea regarding architecture, and am looking for tools to fill specific gaps
• I need a rubber ducky to test the viability of a design choice
• I need a quick check on some code snippet (DevOps)

I have LLMs that I use offline (LM Studio), I use ChatGPT as a search engine since Google is progressively gutting its search engine from functionalities and dorks. For code, I use Copilot. I intend to host models using Ollama locally, but that’s not done yet. The purpose is to rely on sovereign systems as much as possible there, to build habit.

For general research, it can be extremely varied. But I can use it to summarize documents with a fixed scope, perform translations offline with added notes on some specific vocabulary. But it’s lacking a bit there, as the LLM itself is just a brain, it lacks some RAG. Examples will be for CV building, cover letters, providing critique on writing. An LLM absolutely shines on anything language-related.

Both these two domains, I use GenAI, more specifically LLM. This small blog post wasn’t written with LLM, though.

Direction I want to take ?

I’d like to have some tools that will follow some mixed feeds. Then I’d like to have different processing steps making use of LLM along with RAG, to summarize them in some automated reports that I can read more easily.

It’s become incredibly difficult to follow news outlets, as information speed has only gone up. With the addition of LLMs in the pipeline a few years back, quantity of words also seems to have gone up, increasing the weight of information splicing. Thus, I want to move in this specific direction.

LLM is not the devil, but if a user wants it to become one, that’s what it’ll be. Some kind of Pandora’s Box, if you will.