Here I am, two hours before midnight, writing an article about a detailed opinion on how I feel about AI.
Today, in a Matrix Chat room I am occasionally active, there was a discussion about using LLMs, and about my position on it specifically. One of my previous articles, “AI is not only stealing my blog posts, but also my motivation”, was linked at used as a reference point for my statements in said chat room.
This is why I am here writing an article about what I already did with generative AI, what I think it is worth using it for and what I think is just a waste of resources.
TL;DR: Usage of LLMs should be reduced to a minimum, for the sake of the environment, as a protest against the practices of the hyper-scaling AI companies, and to protect you from loosing your ability to think.
If you have read my other articles, you will have realized that I am very against the usage of generative AI. This is not to confuse with normal AI, which I think can be of great use for enhancing the human experience in different aspects of life, but I will go into more detail on that later.
Generative AI
For the sake of simplicity in this article, generative AI, or “genAI”, will reference any system that is made to accurately copy human output stylistically. This includes normal LLMs, but also diffusion models to generate audio and video, and whatever else there may be.
Generative AI, as we all know, has many, many problems. I will list some of them now:
Environmental Impact: the current frontier models are all hosted in huge data centers. Those data centers use gargantuan amounts of electricity and cooling water. The electricity is not fully generated from renewable energy sources, and the cooling water is mostly not cooled down enough to have no significant impact on the water life and environment it is in. Each request to a frontier LLM is contributing to that. This is a clear counter-argument to ever using LLMs.
The fundamental business model: In my eyes, it should not be okay to build a whole industry around mistreating active laws around intellectual property. While I personally do not have a problem with someone citing my blog posts without credit (even though I may reach out and ask politely if they could add a reference to the blog), scraping them for the essence of it being my expression, and sell it to millions of others in form of a contribution to closed-source weights with the pure intention to profit from it feels different to me. If you need some kind of argumentation for that, you won’t be satisfied. This is my subjective perception of this.
The behavior of frontier AI labs: For three years I am hearing that I will be jobless in 6 months. This is intentional to keep the money flow. I don’t think this needs much of proof, and if it does for you, you can look it up on the internet yourself. I am tired of hearing this.
Protecting your own ability to think: The usage of generative AI has provable impact on your ability to think for yourself. Recently, more and more studies are appearing that show a gigantic reduction in brain activity during otherwise heavy-thinking tasks like writing an article or doing some research (I can feel my brain thinking very much writing this, believe me). And I see it myself everywhere. People who cannot live without their agentic assistants anymore. “Hey ChatGPT, what is 5+5”, not knowing that there is a calculator that is much more resource efficient than kicking of a multi-trillion matrix multiplication just to produce a probability that it may be 10.
The first three points are exclusively telling you to stop using LLMs. The last one though hides an exception on where I personally think it is okay to use LLMs: when you already tried your best, and in a last-resort effort you are firing up an LLM to see if it can solve what you couldn’t. Sometimes it can, sometimes it can’t.
While I wanted to write about that in a separate blog post (and probably still will), I had the “possibility” (keeping up the positive language) to use frontier models during a software development internship. This was the first time for me to actually have access to coding agents and the like. I admit, I have used them more than I wanted, and they definitely helped me to navigate code bases I never thought were real. This saved me a lot of time and probably nerves. This is a luxury though, and it may be a bit hypocritical regarding my previous statements, and I cannot defend that. I am only human, I am not perfect. If what the AI delivers is “good enough”, I couldn’t really resist pressing that button that will do all my work for me. I had time to focus on the really cool problems AI couldn’t solve then.
There is a category of models that is special: Open Source models. The profit-intent argument is falling apart. The environmental impact now depends on how you infer with it. Is it locally on your PC that is powered by some solar cells on your roof? Then it is green. If you are renting GPUs in a data center, nothing really changed. You still won’t think on your own though.
Non-generative AI
There is also another branch of AI that I do not have a problem with at all. I will keep this paragraph short though.
I love playing around with different neural networks, like currently where I am trying to train an OCR model from scratch so I can write in my natural handwriting with a stylus on my Linux surface tablet. This is only possible with AI. I am also looking forward to autonomous driving (if it works reliably and is safer than people driving around).
Even in medical image analysis, I think the AI can do a lot of things, and correct errors that might have occurred without it, like early detection of cancer.
My opinion
Now this was all a lot of chitchat, but didn’t really get to the point. What is my opinion now?
I think, you should reduce LLM usage to a minimum. You can’t really fully avoid it anymore. Every google search is now producing an answer with Gemini. Some people are using it in the jobs (or have to, even).
Use LLMs as what they are: capable models of natural language. Translations are now much more accurate than ever before. Finding documents in large databases based on semantic similarity is also a really cool thing you can do, even though this requires embedder models, and not generative models.
I wouldn’t use LLMs though where one would expect a human to do things, like in recruiting, customer service, or especially art. Do not use them to write articles for you, take over thinking for you, and I wouldn’t use them for research as long as the hallucination problem isn’t solved. In science? I also wouldn’t use them there. People are using LLMs way more than it could ever perform, and they are using it as a replacement for themselves, and not as an enhancement.
Do I hate AI? No. Do I hate generative AI? Yes. Even though I have used it already. It is like having to drive with a car while being an environmental activist. You can’t escape it if you do not have the possibility for public transport.
I personally will continue to learn myself. Write code myself. Enjoying the fun of solving problems with my mind that I was quite gifted with. Writing these articles, myself. If someone asks me about my opinion on generative AI, I now have this article that I can link to.
I have seen its capabilities, and especially what it isn’t capable of. AI won’t replace me that fast. It will replace those who are content with the quality of the output of generative AI models much, much earlier. Generative AI is trained to produce the average output, and as long as you are not average, AI won’t replace you.
In the end, do what you want. It is your decision and your life after all.
Stay special.