Do LLMs Really Solve Problems?
I want to share three small personal stories with observations. A recent Codeberg update regarding LLM usage inspired me to share some thoughts.
I hate writing bureaucracy letters
I love that I can give 2-3 sentences in English and get a proper official letter in German bureaucratic language. It saves my sanity, saves my time, but does not really solve the problem. The real problem is that German bureaucracy is overcomplicated and unfriendly to expats. LLMs work more like painkillers - a symptomatic treatment. Root cause stays untouched.
The proper solution would be digitalization and simplification of the system, but doing it and contributing to it is so much harder than just using LLMs to merely overcome the problem. Although, i see that digitalization actually happens in Germany, but with limited success.
I believe LLMs are massively used to write bureaucracy letters nowadays. Therefore, inefficient bureaucracy became even more inefficient due to additional electricity and water impact it creates. And RAM/hardware prices of course. Add the time-waste created by LLM hallucinations on top.
Meanwhile, I admit that returning to writing such letters manually will be a huge pain in the ass for me. A good idea for a Halloween costume, by the way - a yearly tax-filing form or a folder with documents for a visa application.
On a more serious note, did LLMs solve overcomplicated bureaucracy, or make the problem worse?
My time is too precious for reading slop
So I use Kagi Assistant a lot when searching the Internet. And I highly recommend it over Perplexity, Google, etc.
Even before LLMs, Google's search algorithm favoured verbose sites. Even if you want to share a couple of sentences, you should make a 3-5 paragraph article out of it for the sake of better visibility in search engines.
With the help of an LLM, you no longer need to write those 3-5 paragraphs. An LLM can write you a whole book! (That I do not want to read.)
As a result, I use Kagi Assistant so another LLM can read through all that generated slop and point me to an article that makes sense. And now my search queries are far more expensive and can hallucinate. Yay!
But without an assistant, I'd be directly exposed to the worst slop.
Did LLMs solve the problem of searching the Internet, or make it worse?
Now I can one-shot simple tools
Some time ago I decided to migrate from Spotify to Tidal. I couldn't find an official way to migrate my music collection, so I found a script on GitHub. Also, I wanted my Spotify "liked songs" to become a playlist in Tidal, not "liked songs". I had to adjust the script. To not spend too much time, I used an LLM. I had a working and tested result in 30 mins.
I haven't pushed it as a PR because I haven't read the code properly. I know that the changes are shitty and may contain disgusting bugs. If I push such a PR - it's an act of disrespect towards the maintainer.
A nice story indeed - I saved some time for things I really like to do (not a fan of Python scripts, sorry) and got my problem solved. I believe many, many people did something similar. Even non-engineers.
In the pre-LLM era, I'd need to understand this script at least a bit to be able to change it in a way I desire. This increases the chances of me sharing my work with others: I already understood what I did and the act of sharing this work is not disrespect at all - I do not move the responsibility of understanding the code from me to a maintainer.
As a result, everyone builds their own shitty tools instead of collaboratively working on a shared solution. And unnecessarily burns a lot of electricity in the process.
On the other hand, more people can afford to solve their problems with software by generating one-shot tools without understanding things.
In this and similar stories around the globe, did LLMs make our relations with software better or worse?
Is it about LLMs?
Even if you answered all three questions like "LLMs make it worse" - is it correct? Are LLMs really inherently bad?
Personally, I do not think so. It's a social and political problem, not a technological one, unfortunately. Arguing about type systems never led to unbearable RAM and GPU prices.
That's why I want to support Codeberg even if Codeberg is not really stable nowadays. While I'm using LLMs here and there, I see it as my moral responsibility to counteract negative LLM-related influence when I can afford it.
Otherwise, I may start to hate myself.
If Codeberg's current problems are unbearable for you - please continue supporting them even if you left the platform. Or create a git mirror somewhere as a fallback when Codeberg is under pressure. It's a non-profit organization, not even a business. It's made out of passion, not greed. Help them via donations or by utilizing your media influence.
After all, their mission is really cool!