My Love-Hate Relationship With LLMs
27 May 2025I have a love-hate relationship with Large Language Models (LLMs). I feel like they are bad citizens of the internet and will have bad effects long term. At the same time, I feel like I have to use them to not be left behind.
Why is there this Catch-22 with LLMs? Why do I have such a difficult relationship with LLMs?
The Good
LLMs are, without a doubt, amazing tools. I can ask them very detailed questions, and it will somehow come up with an answer. They do make things easier. They help me write. asideHeck, while I wrote the original version, what you read here was reviewed and improved by an LLM! While coding, it is nice to have a very friendly, but also very junior pair coder at my fingertips. It is nice to have a kiind of “living documentation” that not just tells me how to write a module in Angular, but actually adapts the module directly to my needs.
LLMs are amazing!
The Bad
At the same time, I notice a big problem: I just forget to think. It’s easier to just ask an LLM than to do the work myself. I see this with others too. They start coding with an LLM, but because they never have to fully understand the APIs, they do not understand the bugs. They do not understand why something is not working, or is slow.
I do not want to say this is all bad. A calculator is not bad either. asideAnd I like my calculator, especially since I often make stupid mistakes when doing math by hand. At the same time, I do see the value of having learned how to do the math myself. It allows me to apply these formulas to more places, and I can spot wildly wrong result when I mistype something into the calculator.
There’s real value in knowing how to these things without the help of tools.
Another issue: LLMs are not as good as we think. Writing a few acceptable paragraphs isn’t actually a hard task. And honestly, that’s mostly what LLMs do. If you think they’re doing more than that, go read a book by a truly good author.
It does not produce maintable code. Yes, I tried. With different models. But they learn from Stack Overflow, which contains just a few snippets of code. Writing maintainable code is more of an art than a science. I can already imaging that in 5 - 10 years, I’ll be charging a major premium to clean up AI-generated “slop code.”
How many serious security issues will be introduced by people who just “vibe code”?
It also works best if you already know what to ask for. For example, I cannot create a good image with an LLM. Just look at the image for this post! I can’t explain exactly what is wrong with it, so I don’t know how to ask the model to fix it.
It is the same with code. I can tell an LLM to not store passwords in plaintext and instead use a cryptographically secure hash. But if someone doesn’t know the words “cryptographically secure hash”, how would they know what to ask? Or that there’s a problem at all?
The Ugly
Now, here comes the ugly part. The rest is all fixable with time. But here is the real Catch-22, the real crutch of the problem.
LLMs are trained on stolen content. Yes, I know, people argue if it really is stealing, but it is! Just look at the crave around the Studio Ghibli-themed images! The companies behind models admit that their business model would not work if they were not allowed to steal.
This is the “Uber argument”. Uber also claimed that their business model wouldn’t work if they would have to employ (and with that, pay fairly) their drivers. I strongly believe such business models just should not exist.
Using an LLM also seems to turn off our brains. There is something magical having a tool tell you what the odds are. Studies have shown that when an AI tells us something, we just believe it. And even worse, we absorb the biases of whatever model we use! Of course we also absorb the biases of the people around us. But that at least it distributed. If we have 3-4 AI models in the future (which I think is highly likely, as the network effect helps the AI get better. As it helped Google Search get better), then 3-4 companies control our thinking.
Even if the companies do not try to be biased, bias can sneak in. A famous example is when an AI model was racist. So the creators removed “race” as an input. What did the AI model do? It just ended up using ZIP-codes as a proxy for race. asideI know, shocking!
And I wonder: who will train the LLMs in the future? If more and more internet content is made by LLMs, and then LLMs train on that content, the biases will just repeat and grow stronger. The dead internet theory says ‘Hi’.
Will an LLM go and take a walk around a lake with a confidential source as investigative journalist do today?
And I haven’t even mentioned the environmental cost of training and running these models.
So…
Should we throw the baby our with the bathwater? No. But we, as a society, need to decide what is OK and what isn’t. Where should we use LLMs? And where should LLMs be forbidden, even if that means accepting inefficiencies?
The problem is: The cat is already out of the bag. Even if “we” (as in, the west) decided to stop, other parts of the world with weaker privacy protections would keep going. Should we just leave the playing field to them? Probably not. But we must find the right way to move forward.
Another Catch-22.
Should you be worried about LLMs? This is for you to decide.
Me? I am not. At least not for my livelyhood. There will always be a space for people who are willing to go the extra mile, to deeply understand things. But I do worry what this means for society. And I am not so sure that in the future we will look back and agree that this was a positive development. But we will all use LLMs. Because we have to.
And if everything goes wrong, I’ll just spend my days cleaning up AI-slop code. For a ridiculous daily fee.