"Resist! Prove that you exist!"

Foreword

Whatever our opinions on the matter, Large-Language Models (LLM for short), dubbed "AI" by the layman, are around us and here to stay.

After some time lurking on lobste.rs, it feels like there are two distinct camps: (1) LLMs are great and we should embrace them; or (2) LLMs are bad and we should not use them, resist and ban them (one such example is airesistlist.org). Actually, banned AI/Robots is a recurrent theme in Sci-Fi works: Dune, Asimov's universe, Blade Runner, Battlestar Galactica, Mass Effect, and so on. (Note that there are also works where sentient machines co-exist with humans, e.g., Star Wars).

I've personally used LLMs a bit to generate text, simple scripts, or images. It can be helpful, but my current impression is this is the case only when the Human in the loop knows about the discussed subject; these systems are too prone to hallucination and too persuasive. As for generating images, my excuse is my skills are stuck at drawing stick figures.

In this post I'd like to explore a few thoughts on this topic.

Who Created the Output?

A typical discussion about LLM and copyright can quickly turns into stating that LLMs have been fed with copyrighted material without any regard for the rules, whatever we might think about those rules. Going a bit further, we can ask ourselves who created the result the LLM produces.

I clearly remember the first time I asked an LLM to write a scientific-looking text. I was just curious. While the text was not too bad, I didn't feel like I owned it. It was not my text; it was written by someone else. How could I claim this text is my contribution if I am not the one having written it? By analogy to books, I was the publisher while the LLM was the writer.

Other criticisms I've read are:

• the LLM can change the voice of the author, potentially cleverly integrating its creator's agenda due to training bias;
• one doesn't learn how to write, how to express himself, when asking an LLM to write text;
• why would one bother reading text that was not thoughtfully created?

The Artistic Nature of Humankind

Another topic that comes often in discussions is whether machines will replace humans or not. Looking in the (not too distant past), I'm pretty sure painters had similar questions at the dawn of photography.

As we can observe, photography didn't replace paintings. Painters adapted their style to something more abstract, less realistic, because realistic painting didn't make sense anymore. It took some times to adapt, sure; there was a period of turmoil where no one knew what would come out of it, and I believe we are in the same situation at the time of writing this.

Even more recently, digital photography and good cameras on smartphones didn't kill professional photographers. What they did was merely decreasing the cost of taking a picture, similar to photography decreasing the cost of saving an instant compared to painting. Some people also reflected on this situation and tried to make their pictures special, not just showing what something looked like, but how it felt. I'm thinking of street photography in particular: the photographer conveys emotions in addition to a visual copy of a given situation.

Expressing emotions in photography was one of my reasons to get a new DSLR (it's expensive, I need reasons to justify the expense). My phone is for non-artistic pictures, for things I need to remember for later (the price of an object, the title of a book), while the DSLR is for artistic purposes. I cannot claim I am good at it; this is only the direction I aim for.

Another analogy would be craftmanship vs mass production. On one side we have a unique product that takes time and skill to make; on the other side we have a bland, standardized product devoid of any human touch. Not that one is worse than the other (everyone nowadays possesses tons of mass produced products), they are just different.

In short, after having industrialised labour in the past decades, LLMs industrialise creativity. I think both can co-exist the same way painting and photography, music concert and music players, or craftmanship and mass production. Depending on the task, person, and other factors, one can prefer one or the other.

The Darkside of LLMs

Ideally, technology is neither good nor bad, it is the way we use it that makes it good or bad. However, I think there is something bad about the current state of LLMs: the current framework of our society.

Which LLMs does the layman person know? The most famous ones are created and maintained by private companies who do not really have their users' best interests in mind: collection and use of data without users' consent, absence of guardrails to prevent misuses or disinformation, prioritize quick economic returns instead of long-term well-being, etc.

If people would be less interested in their own personal gains, if there were more guardrails and regulations, more accountability, better education, and less pressure tactics from the people in control then maybe we could achieve a better society.

About the Environment

There exists numerous complaints about the impact of LLMs on the environment. Some people claim that, well, the power consumption is not too bad, renewable energies are used when possible, the hardware is getting more efficient, etc. Well, bad news, doing a computation, whatever its efficiency is, consumes more energy than not doing it.

For many things in life we seem to forget to ask whether it is worth it. It is not because we can do something that we should do it. I believe that mass adoption of LLMs is not worth the damage. Yes, LLMs and other technics under the AI/machine learning umbrella can be useful in some cases, but this is not something we should push in every corner of society for the sake of riding the buzzword train.

We need to slow down and consider the long-term effects first. Homo Sapiens appeared ~300,000 years ago, writing is commonly recognized to having been invented ~5000 years ago, the industrial revolution was a couple of centuries ago, and the term "machine learning" appeared less than 70 years ago. We do not need to disrupt our lives every minute and rush towards destruction. Let's preserve what we have and consider the history of humanity as a marathon, not a sprint.

What's the worst case? Long-term suffering on one hand, or not being able to get our dopamine rush in the next 5 seconds on the other hand.

What's next

Is the solution a total boycott of LLMs? I do not believe so because it is already too late and not realistic. Instead, we can reflect on our relationship with them and respect ourselves by making them ethical and safer.

Instead of using closed-source models trained on unknown, potentially copyrighted content, we can use open-source ones. Instead of using the models of big, monopolistic corporations we can use the ones of companies that care for our privacy as well as the environment (LLMs run locally are also a solution to be considered). And, all-in-all, let's not forget that we are a social species; interaction with a program will never replace interaction with another real human.