I went to see an exhibition today. The theme was technology’s affect on people. Of course, the times being what they are, there was a lot of allusions to AIs. Well, that’s just myopic. I’m 48. That’s old for much of the history of humanity. That is enabled by technology. I can see, because I have glasses. That’s enabled by technology. I’m overweight, because technology allows us to have such an easy access to food.

Will AI have such a profound affect on humanity? Maybe some day, but it isn’t now.

The first mistake people do is they don’t understand that artificial intelligence fails when you leave it alone. More and more people seem to understand this. However, artificial intelligence can work as support for humans. This can actually facilitate better results and make things easier… Well, except that LLMs are not designed for this. They are designed for the greedy, for the people who want to get rid of humans, not the people who want elevate humanity. Sure, maybe some day these tools will be better than humans but considering how bad they are, this will probably take decades.

What people often miss is that we don’t write despite it being difficult, we write because it is challenging.

So, if you don’t want to use them, just don’t. It’s okay. It’s more than okay. The reason I use them is that I need to understand them for my job. I teach IT, so in order to do my job, I need know where their limits are. And turns out, writing is a good case study.

Don’t Let the LLM Write a Word or Tell You How Something Should Be

Brain rot is real. Think about it this way: a long time ago people mostly worked in the fields, so we got the exercise we needed that way. When that stopped because of technology, we have lost our muscles. If we don’t do the intellectual work, we will similarly lose our ability to think. Our brains are not constants. They need stimulation. If we are just to oversee the machine, we don’t get that stimulation, so we lose our ability to think critically.

Also, you will lose your connection to the text. If you haven’t put the necessary thought into text, you won’t know what it’s about or what you are trying to communicate. So, you need to do the work. Otherwise editing will be impossible and any kind of continuity will be just a dream.

Remember the Limitations of LLMs

A professor, who’s name I can’t remember right now, stated in a Wired video that LLMs are very good at producing good looking text, so they are good tools for creating good looking text. Note that there’s subtext here. Your text should be more than good looking. It should also have continuity, message, ideas and so forth. The LLM can mimic creating those, but it can’t actually achieve what you should want from your text.

Each LLM has what is known as a context window. This is how much it can remember at one time. It’s measured in tokens. Usually this means that each word is a token, but in some cases a word can be represented by multiple tokens. The LLMs make big claims about their context windows, but that is often dependent on the current load on the system. If the context window isn’t enough to handle the full input, it will be summarized by the LLM based on some arbitrary rules.

This might not always be obvious, but at some point the LLM can very well summarize your discussions without you realizing it. When you have longer discussions, you will notice that it will start getting things wrong. In some cases, this is because of the context window. It has summarized the text in such a way that it can no longer use certain information.

Of course, it can also hallucinate. This is common with themes and concepts where there isn’t much information on them in the “corpus” (the data fed to the LLM in the training). Even the best models hallucinate about 10 per cent of the time.

Now, these two combined can be devastating to a discussion. If the model hallucinates while summarizing a text, it will get to very weird conclusions.

This means that you might have restart discussions often, because when the discussion goes off the rails, there is no way to return it to the right path.

LLM Is Not a Good Copy Editor But It Can Help You Get Started

Writing is re-writing. This is where LLMs can be helpful. You should do the first draft without any of this, but after that, toy around with them if you wish. An actual copy editor will cost you much more than an LLM. At the same time, they will also be much more useful.

That being said, LLMs can be good for certain specific things. For example, being Finnish, personal pronouns have always been challenging for me. This is because we don’t have multiple. We only have one. So, I have this weird problem where I’m forming a sentence and I somehow skip from character to character and then I do it too early or too late and get the wrong pronoun. Like if a man is thinking about a woman, I might skip to feminine pronoun accidentally too early. Since my current project has a trans character, this is a real problem. So, I asked an LLM to check for this.

You can also ask it to analyze other things from a draft like character arcs, plot lines and so forth. You will notice it getting things wrong, but it can also give you a rough idea about those things and thus give you something to work with. I often try to come up with very specific questions, like “Are there any Chekhov’s Guns that have been left unfired?” or “Does the evolution of X as a witch feel organic?” Without getting into details, you will not get good answers.

You can even try weirder things. Like, I’ve asked an LLM to identify the favorite movies of specific characters. This might not mean much, but it gives me an idea on whether these characters read the way I want them to on the page.

The point being that if you want to use LLMs in a meaningful and productive way, you really need to get creative with how to use them.

Fixing Grammar

Do not use LLMs for this, use tools specifically made for checking grammar instead. Using an LLM for checking grammar is like using a truck to get to the grocery store in the next block when you could have easily just walked. The energy required by the LLM is immense while grammar tools are relatively simple.

Sure, neither of these are especially good at this. They tend to enforce mediocrity in style, so be careful with them. Still, you will be able to fix most problems without having to pay someone for that. And you will have problems in your text.

Research Assistance

Sometimes you come across things that are hard to research with search engines. One problem being that they are also so infested with AI now that they are becoming less and less useful. You should probably still try them first for the lesser energy consumption, or ask an expert if you have access to one. I mean, that is very helpful, because they will often give you perspectives you (or the LLM) wouldn’t have thought yourself.

Anyway, in my current project, among many weird things, I needed to figure out how omnivorous animals experience food. Yes. I had an idea how this works, but I wasn’t sure. I also couldn’t find good sources for this, except for one in a book I own about how rats are able to balance their diet, but it doesn’t really explain how they do it, just that they do it. So, I asked an LLM and got a good enough answer. I don’t know if it’s true, but it sounds true enough, so I went with it. A biologist might feel differently and if my work was aimed at biologists, I would have asked one, but since it isn’t, I went with what I had.

Another situation where I needed research assistance was an action scene. I needed to know various details on how snipers work and how that would work in this specific situation. That is hard to tell from a bunch of texts and I didn’t have a sniper available to me, so I used an LLM to explain things to me. Now, I know something about this, not much, but something from my days in the army (which was almost 30 years ago now), so I trust my intuition that all of this is realistic much more than with the omnivorous animals.

This works much better when you need to apply knowledge to a situation. Sure, you can get theory by reading, but applying that theory is much more difficult. So, you can get help with that from an LLM. Returning to my earlier comment, you still should not let the LLM do any of the writing. You should still do the writing yourself. Just ask a lot of questions on how something should work.

Finally

Again, you don’t have to do any of this. You can just write. It’s so much fun and so creatively fulfilling that you shouldn’t be worried about using AIs. You should just do the work. I could go into philosophy here, but I won’t. Just remember that you should not let AIs take the fun out of your work. You have chosen to write because you like it. So, keep liking it. It won’t be as fun if you don’t do it yourself.