2025-09-02 - GenAI, my usage so far

Following the trend: ChatGPT moment

Invited to test the beta version in early 2023, I was using ChatGPT from time to time, and I noticed the improvements made over the different versions of the model. It was helpful for synthesising data, editing notes or translating texts. As a coding assistant, I was also using it to get feedback on system design, to help configure APIs (ie: Sendgrid API on Rails) or to debug errors, but it was not always useful as it could get confused easily. Maybe, I was not clear enough on prompt/context side at that time.

Directly on coding side, I was using Copilot on VS Code. Not really impressed initially, I saw that it got tuned nicely and was helping for basic code autocompletion, especially on unit test generation side. By giving a lot of context, it started to guess many things just with the name of the test file, getting directly the class to test and suggesting a lot of coverage, sometimes too much.

I dug more into Prompt Engineering thanks a Google white book from Lee Boonstra and some articles (like this one from LangChain blog).

Local playground: Ollama and OpenWebUI

Earlier this year in January 2025, I have also started checking on open source projects related to GenAI and I was playing locally with Ollama and OpenWebUI, thanks to the suggestion of a friend.

I was quite surprised to run deepseek-7b model on my Raspberry Pi 5 8GB. Well, it was taking ages to get an answer, sometimes crashing, but with a bit more memory, it looked possible, and it was something real. Especially that I was quite interested in testing small models listed on HuggingFace.

For the setup, I followed instructions from different sources, it was quite straightforward to deploy it using Docker Compose.

On a side, I started using OpenRouter in order to switch models easily on my OpenWebUI instance. The list of supported models was huge, I was still using mostly ChatGPT, but I have been testing also Deepseek and Mistral AI models.

Summer digging: The big picture and shift to Sonnet 4.5

This summer, I took time to explore deeper the AI theme, on a historical and technical side, and I am suggesting 2 good resources for this:

A wikipedia page with the big picture on History of artificial intelligence
A Youtube channel explaining a lot of concept behind GenAI : 3blue1brown This helped me to deepthen my basic knowledge about AI, as the last time I studied it was on a Neural Network class in 2006. Many new things have emerged since, and it is funny to see how Generative AI is taking most of the attention lately, while there are many other AI concepts being developed aside (Reinforcement Learning, Predictive AI, …).

Exploring again the open source tools, I started testing Continue Dev because I was able to plug it easily to my OpenRouter account, and using it with its VS Code plugin and its CLI. Using a CLI for GenAI coding was new to me and I liked the idea behind, making it clearer to define context or to connect with external tools via MCP (Used Context7 for API documentation sync and Playwright for browser access for e2e tests). This is still quite new, so I don’t have clear opinions about it, but it looks promising.

On the model side, I have mostly switched to the Sonnet 4.5 model, as I find it more relevant and structured, especially when you need to synthesise information. I am using it on OpenWebUI and with my ContinueDev plugin in VS Code.

Next step

The GenAI ecosystem is on an ADHD-like phase, it is going on many different directions, especially on marketing side, generating a lot of nonsense products and user confusion.

But if you look more carefully at the technological side, you can see that the steps are quite logical building up the ecosystem:

Improving the models
- Increasing the Context window
- Reducing hallucinations
Improving the contexts/prompts engineering
- Including pre-processed data (RAG)
- Linking external data (MCP)
Improving the automation
- Using n8n for agentic approach
- Defining auto-evaluation of requests and failures

On coding side, I would like to try 2 solutions that many people are talking about lately:

Cursor
Claude Code To be continued…