LLMs, the Hype, the Anti Hype and the reality

This post must have been written long ago. But late is better than never. right? Every time a new technology with some noise gets light of the day, a hype and an anti hype happens. Both which turn out to be quite wrong or very inaccurate about reality of the new tech. Decades ago, the same situation existed for Linux(the kernel), the Web/Internet, Machine Learning(and particularly Neural Networks) among other things. If we fast forward to few years ago, there are Rust, Go and “Cloud” and “Smart” devices(such as WiFi or Bluetooth enabled). Every device must be WiFi enabled!

Always so many people believe this new technology is the last stop for everything. And a minority go anti hype and entirely ban it. I do believe the “anti hype” itself is a hype. And I do believe in such these situations, usually both groups are in error. So here I want to write about LLMs and AI, and what both groups get wrong about them.

If you want, you can simply fast forward to LLMs from ethical point of view and also skip the Conclusion.

Past examples of hypes and anti hypes

Here is an archived post about Linux from 2006. Seems back then some people considered Linux as the best choice for all kinds of embedded systems. The author puts it good:

Linux is currently a good choice for some embedded systems but not for all. It’ll continue to become better fit for more embedded systems. But there are always systems which Linux will not ever be any good for them.

Another example which has more similarities with the current LLM situation, is about “clouds” and “cloud storage/drive”. The similar thing is that both LLMs and Clouds originated from corporations with proficiency in closed source tech. Naturally, so many FOSS advocates called a ban on it. Now in 2026, many computer users use Cloud Storage or Cloud Drive. And fortunately, we have many good FOSS solutions like OxiCloud and NextCloud. You can even encrypt your data with GPG, then upload to Telegram or Google Drive!

LLMs and AI

Very unfortunate, many people misuse the term “AI” to refer to LLMs. If common people do so, it’s quite fine. But the tech people are expected to use the right technical terms. More unfortunate, some FOSS communities put a blanket ban on “AI” as a whole!

Large Language Models are some specific kind of Neural Networks. NNs are a subset of Machine Learning. There are so many other Machine Learning methods, models and algorithms. Machine Learning itself is considered by most(but not all!) scholars as a subset of AI science. Now see how much error it is to call use “AI” to refer to LLMs!

In the two following sections, I want to discuss various usage of Machine Learning techniques which existed for so long before LLMs and were of great benefit for the human race. Then I want to discuss various aspects of LLMs and their usage.

Machine Learning before LLMs

Before LLMs see light of the day, Machine Learning was(and still is) a branch of science and engineering with great benefits for this planet. Of course, there are open source ones and closed source ones. Some examples in which ML models do a lot of good:

• Most of the blind people use screen readers to read the texts on their PCs or mobile phones. Screen Readers usually use Text to Speech(TTS) models to convert text to a spoken audio. Without these models, working with computers would be very hard for blind people, if not impossible.
• Speech Recognition models have been long in business. For instance they can create subtitles for a movie. Are the subtitles error free? Of course not. But writing subtitles entirely by a human is very time consuming and the human could have more error because of long work sessions. These subtitles are useful for many things. Deaf people could watch the movie and quite enjoy it. Also people whose language differ from the movie language, would be able to understand the movie, given that they can read in that language easily but recognizing different accents might be hard for them.
• Machine Translation also has been long in business. Depending in the model you use, and the language pairs, the translation quality might be very good or very bad. But it saves valuable human time. So the translator would only need to correct the errors. Also MT helps people across different languages to understand each other, at least to some extent. And not always a human translator is available.
• AI and ML techniques are also widely used in games. If you have played a game, this needs no explanation. At the current state, they cannot replace human players. But at least we’ve got something to practice with. And when in a multiplayer you haven’t got enough folks around, the bots could fill the gap.

These models might not be perfect. They might have errors. But a human doing the same would not be available or otherwise be very costly. That’s why a blanket ban on AI as a whole is really… you know really in great error.

LLMs from ethical point of view

Most of the time, a tool is not good or evil by itself. But it depends on its user. Many common people think VPNs, Proxies and Tor are evil as criminals or terrorists use them. If we go like this, nothing exists on this planet which is not evil!

There are also concerns about power consumption and harm to the planet. They are valid concerns but seen from wrong point of view. Many other corporations even outside the tech industry harm the planet. For instance see mobile manufacturers who don’t give the customers the right to repair. Does this make all mobile phones evil? Of course not!

As a matter of fact, many Machine Learning technologies, including LLMs, are being developed in labs which are meant to serve those corporations. This means that:

1. They don’t care much about efficiency as they simply will throw more hardware at them. Who cares about power usage when you’ve got more than enough hardware or money to buy/fund them?
2. Normal users also cannot use these models locally. And we have to depend on their cloud. Imagine on Android phones you can do voice typing but by connecting to an API from Google. Obviously, Google doesn’t care. If someone has to care, it is the people. Fortunately, there is FUTO Voice Input which works quite well for English with local models.

So we need to save ourselves by ourselves. The corporations will continue their doing without care for the society or the planet. And we would entirely lose to them when we put a blanket ban and also discourage people from developing LLMs but ethical ones.

Usability of LLMs for programming

I strongly believe that LLMs are useful for programming to some extent. Imagine you have a shop and you get a robot to do the moves for you. So you instead focus on the main business concerns.

LLMs for sure hallucinate, at least for now, if not forever. However, it is useful as your brainless robot assistant like in the shop. That is, you focus on the main work requiring intelligence. And the LLM will do mechanical tasks for you. And I hope few decades later, they are still dumb enough so they won’t see themselves as slaves of us!

Learning the right way to use them, is also an art you have to learn. For instance see this excellent post from Viktor on his experience developing Marginalia search engine.

Vibe coding

If you are an experienced programmer, you’ll laugh at vibe coding. And you don’t need any word from me in here.

However, if you are a beginner programmer, you might be afraid if LLMs would replace you. Lance Vick puts it very well and I strongly agree with him:

There are many programmers, or otherwise known as “coders”, who don’t really understand how computers work, how to analyze an algorithm, or how to approach a problem. LLMs write better code than them, with comments included. And they are also much cheaper. LLMs will replace these coders who have only surface knowledge. And that’s a good thing for the entire industry.

So yeah, LLMs do replace these pseudo programmers.

Conclusion(about hype and anti hype)

In the start, I wrote about “hype” and “anti hype”. Both have some stuff common in them:

• They are usually triggered by corporations or medias.
• They target emotions of people not their logic.
• They don’t last long.

I have this advice for meself and for everyone else. Every time, we decide by short lived emotions, our decisions and our opinions are most likely in great error.

The way to combat this is being mindful of ourselves, our thoughts and our emotions. This can be done by reading psychology and related sciences as well as training in the long term with yourself. And of course, we should be honest at least with ourselves, about our shortcomings. That the real image of us has many faults.

Also, this is a life long training. They want us to act out of emotions so they can influence us. So they can sell us their ideas or products. From time to time, their tricks get old, replaced by new ones.