My approach to AI

I’ve been following the AI stuff since 2018, watching videos on the topic and specifically machine learning informative videos from DotCSV (I really recommend this channel for ML/AI stuff in Spanish), even before the ChatGPT bubble, even before the GPT-3 model was available to the public, (and I think even before GPT-2). Seeing stuff like AlphaFold was amazing and still amazes me today; image resolution enhancers like waifu2x; 3D “modeling” with NeRF and Gaussian Splatting. I just saw it as an interesting topic though, it didn’t catch my interest to learn deep about it.

However, nowadays I don’t like the current usage of AI, specially in programming and art (“AI Art” is not art). About programming, I’ve been reluctant with using LLMs for it; even today, with a little bit more experience on the usage, I don’t like them.

Note: I didn’t get into the ChatGPT bubble at the time because it required to add a phone number verification and I just didn’t have a phone for months xd

My previous usage

I didn’t use AI for programming up until 2024 in indirect ways: I had to use IntelliJ IDEA for a Java SpringBoot college project, which has some AI integrations by default and was sometimes useful for autocompletion, though I didn’t use it that much and most of the time I ended up rewriting some of that autocomplete.

At my previous job there was an AI fever for quite a long time, for which I wasn’t very supportive, being very clear that I don’t use AI and I wasn’t willing to use it, even for basic tasks or for help. I don’t regret it, and certainly I don’t see how an AI could have help me then.

My serious tries

ChatGPT

Now, at my current job, I use AI almost every day, “why?” Because I’m required to it’s useful. I was kinda forced to use it and was given a ChatGPT Pro subscription to use, which I ended up liking to use (not at all). ChatGPT is actually very slow and I’ve seen into other alternatives like t3.chat, but I’m not sure yet.

Up until May, I used ChatGPT mostly as a question-answer model, not trying to solve the problem but to give me hints, code, or answers to what I’m trying to accomplish. For example, instead of asking to complete a certain task, this was my workflow:

1. Ask about possible approaches for a task
2. Google about them
3. Try to implement one approach that fit my use case
4. When an issue occurs, ask about it and try to fix it myself

Most of the time, I didn’t give it the full code I was working on, just some pieces and try to get an answer with it.

Cursor

My coworkers all use Cursor instead, which seems very useful as it already has the project context, so making changes or even asking about the project are easy tasks, and it’s very handy. In fact I was offered a Cursor subscription instead, but I prefer to use neovim… so it was an easy decision. However I still tried to use Cursor with the free plan to accomplish some tasks and ask some things about the project when I started my job; it wasn’t successful.

Augment

Then, Augment came in and started using it because a coworker recommended it, specially because it has a neovim plugin. I used it for a week, and honestly it worked like a charm, but it was always slow and most of the time it wasn’t very useful to me; I was often switching back to ChatGPT.

avante.nvim

That’s when I decided to use avante.nvim instead. I went with the DeepSeek API because it was cheap; I added just $5 and I’ve only used about $0.5 in this whole month 😔👌🏼. However I had the same issues as with Augment… very slow, and not very useful, and even it was sometimes a bit buggy. The Agent mode sometimes worked, I even tried to do some vibe coding with it, but most of the time it bugged and had to restart the chat.

So, after various tries at using AI extensively and trying to integrate it with my editor, I always ended up going back to ChatGPT. So, considering the Unix philosophy of “Make each program do one thing well” and my love for terminals, I decided to find an AI chat in CLI/TUI program.

AIChat and DeepSeek

AIChat

AIChat is an all-in-one LLM CLI tool with a REPL mode that I’m currently using with DeepSeek from simple questions to my day-to-day work. Even though it doesn’t has a RAG and Agent mode for DeepSeek, it was far more useful than any other solution.

I use Zellij as my terminal multiplexer, so I now always have a split pane for neovim and another one for AIChat, and simply having a tool that I can use to ask anything at the distance of 2 keystrokes is AMAZING. I merely switch to the other pane, ask some stuff, read the answer and go back to my editor to write some code.

So my current workflow is the following:

1. Open neovim and AIChat in two split panes in Zellij or tmux
2. Ask some stuff in AIChat. Provide files with .file "./file" -- ask about it.
3. Go back to neovim and continue writing code.
4. When an issue occurs, ask about it, provide the file, and fix it myself

Yet, I have some terminal issues with AIChat, where it sometimes renders incorrectly due to Zellij. This has nothing to do with AIChat, but still annoys me sometimes; I may have to look into it in the Zellij repo.

DeepSeek

Even though DeepSeek is a cheap model, is surprisingly very capable, it still amazes me how I’ve only used $0.5 this whole month without any issues in my day-to-day work and for college stuff. I have nothing else to say, it works very well compared to o4-mini-high, which I used before with ChatGPT.

Conclusions

I still don’t like AI. I always end up writing code myself. Vibe coding is not for me. However, it’s still useful and has been proving me wrong about its usefulness in the past weeks.

I’m gonna continue using AIChat with DeepSeek for the near future, and will likely be using them in the long run.