AI policy

One year ago, back when I released the first version of this blog, I used to have strong opinions against AI. My perspective on AI has changed quite a lot since then.

I used to advocate for manual labour from a romantic point of view. I liked to consider myself an artisan and didn't want an LLM to touch anything related to my code.

Dunno if it's because I grew older, lazier or just more tired, but I now regularly use AI. NOT LLMs, but rather SLMs. I like to conduct experiments and advocate for open weights models small enough to run locally to take advantage of a more practical "search engine" without compromising too much the environment. Intellectual property theft is another issue, but I still wonder: if my open-source code is used to make open-source language models anyone can run locally, I don’t care that much, I guess?

I still stand against companies using human data without proper attribution, and since we cannot decide (yet) which language model our code will feed, I still prefer to publish my code in more protected environments, like Codeberg. See my statement on GitHub.

How do I use AI?

As stated previously, I'm all in for SLMs. The SLMs I use are typically in the range of the 4 to 8 billion parameters, which can run perfectly at ~100 - ~70 tokens per second on my machine. Occasionally, if I’m feeling extra lazy or extra blocked, I’d go for a bigger one using a Mixture of Experts architecture.

Currently, I'm using:

Ok, but HOW do I actually use AI?

I think that calls for a dedicated post. It's not a matter of how I use AI, but more about how to work with Small Language Models. As stated previously, I've been conducting some research. As soon as I have some publishable material, I'll link it here.

For the time being, it's useful to note that it looks nothing like working with proprietary large language models. They are mediocre to say the least when it comes to agentic tasks. But they are useful! As long as you elaborate the context yourself and are very, very specific about what the expected outcome should be, they can really help to explore and understand codebases, to avoid some boilerplate, and to have a smarter search engine that can tailor the answer to your specific need. More like using an LLM 1 or 2 years ago (according to the research I've carried out. It wasn't until recently that I started using AI lmao).

I don't think I'm giving up the intelectual struggle. It's just more... convenient I'd say. Nothing to do with productivity. Productivity should stay low and comfy. As a primarily frontend developer, I’d much rather have an AI give me an SQL query than struggle to build SQL queries myself, if that makes sense.

Also, since English is not my native language, I sometimes ask it to review typos and usage of prepositions. ONLY REVIEW. I write my content by myself. AI might have replaced "migth" with "might" in this article, but it never goes beyond that.

OK BUT WHAT ABOUT CAPITALISM

At this point in the dystopia, I think we can all agree that AI hype only benefits large AI companies. I don’t think AI is replacing us anytime soon, but these companies are selling the idea to non-AI-related companies that they can fire their whole staff. They’re making work precarious and cheaper, while causing an overall anxiety and uncertainty. I don’t think anyone who has been forced to use AI in their daily work can say they are happier now. I firmly believe that AI, in the scale that these companies are trying to lure us into, makes us more miserable. I think this article describes it more beautifully than I ever could.

But, for the first time in history, I think we do have a chance to seize the means of production. If we learn to work and to accept the flaws and quirks of SLMs, we could show these corporations that we don’t need their proprietary-soul-eating-earth-burning monsters. That we are okay with open-source small models that can make our work more comfortable instead of threatening to steal it.

Might change my mind in the future but, for the time being, let's leave it like that. I'm open to feedback and new ideas.