Thoughts on AI and LLMs

What is AI?

The term AI actually refers to a broader term, artificial intelligence, and has been heavily overused by many major companies to mean something a little more sophisticated than what it is. The product that everyone is pushing so to speak, is large language models. Large language models are basically in essence a large static file that contains highly compressed information. Like a file on your computer. Those AI chatbots are actually just querying a database file, but unlike other databases like SQL it doesn't easily change. You can of course re-train the model, but that takes quite a lot of time, energy, money. It's like oh we've got some new information to add to the database. With SQL it's effortless. With an AI LLM, it can take significant amounts of money for the next information dump.

Biases

As such, as these LLMs cannot learn new things easily, it's very easy to accidentally, or on purpose add biases. Political correctness for example. For example AI's that prefer the resumes from men over women https://www.digital-adoption.com/ai-bias-examples/ and other slightly worrying prejudicial biases.

Good uses and bad uses

What a lot of people don't know is that the concept of AI or LLMs, is not new at all. ChatGPT did not invent it. LLMs are trained neural networks, and neural networks have been around in practice since the mid 20th century (70+ years ago) https://en.wikipedia.org/wiki/Neural_network. So when people might say they're anti-AI, they're essentuially condoning a technology that's 70+ years old.

Good uses

Some might argue there are very few. From what I am starting to see, LLMs by themselves are usually not good enough at doing most things. They almost always require human oversight, plus fact checking of some sort. However, they do seem to have a trend for doing things like, summarising information and creating a natural language query for large amounts of text. But also creating less biased text, such as for programming. However even with programming, it can be biased Copilot stops working on gender related subjects #72603. It can also be used for things like image and text recognition, audio, speech recognition. These are quite valuable uses that have already been in use pre-LLMs.

Bad uses

I'm starting to think that outside of more neutral, logical uses of LLMs to create less expressive things. Almost anything expressive, artistic, musical. LLMs can be used to create images, video and music. Despite how impressive some might be, I am very unfavourable to using LLMs for anything remotely artistic. People will still use it. There are however two issues with using LLMs for expressive puropses. One is that it's not from a real human with a unqiue perspective. It's my opinion that the whole point or purpose of art is to express yourself. From the perspective of the artist, but also from the viewer. The artist is channeling a part of themselves, which is inherently extremely unique. And from the viewer, the experience of seeing unique artworks is usually more enriching than from the perspective of a static LLM binary file.

Neutral uses

One example of a controversial use is language translation. In some cases it can be very convenient to rely on a LLM to translate a language over a traditional language translation technology, or a human. On the other hand, using a human or a traditional translation service that's been trialled and tested (by humans) often picks up on unique nuances each language has. Remember Engrish, a old past time of sharing funny Japanse translation fails. These were when someone would try to translate words literally.

https://www.pairaphrase.com/blog/llm-translation-advantages-disadvantages

Environmental Friendliness

From the major online AI products such as ChatGPT, Gemini and Claude, we don't actually know how much energy they use, but we can guess https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/. They do prefer more reliable power sources such as coal stations, over wind and solar which fluctuate. With how eagerly companies are shoveling AI into all their products, for the questionable duration that it will lead to profits, it feels a little irresponsible that AI tools are so ubiquitous when they are consuming more power, with the high chances of increasing. Overall, without transparancy of more numbers, the sense is that it's not terrible but it's not great if there is profit incentive to increase AI/LLM use which in turn increases power demand, which is currently preferable to coal or continuious power supplies.

Some of my uses

I am inherently curious, I pulled apart my own tech gadgets when I was younger, like my GameBoy and casette player (I'm that old). You give me a free website with a highly interesting text prompt, I will tinker with it unless someone stops me. Over time my uses have change a bit. Originally a lot of text queries, asking for answers to random questions that I would have otherwise use a search engine for (I use DuckDuckGo currently). Moving to more and more technical questions, less and less non-technical questions. I'm finding that it is healthy to develop your own ability to solve problems, using existing open data, which the LLMs are largely trained on anyway. Or just asking friends, professionals who more often than you might think can offer very decent answers. I find myself gradually shifting towards using LLMs for more technical problem solving and programming assistance. Even things like recipes, a classic example, I prefer to do my research on what people have used IRL. It may take a tiny bit more time, but the output is a lot more reliable. It's starting to cotton on I think, that LLMs are kinda BS machines sometimes. They hallucinate or just come up with very unimaginitive ideas. It's a skill though, like learning to Google or web search. What are the sensible uses for LLMs.

Software prototyping

Very recently I have been tinkering with Claude. I'm not too partial to it, I know there are other aletneraives, such as host your own. It's been pretty successufl in producing quick prototypes of ideas. For example, a project I've always wanted to see exist is a modern version of iTunes notably taking the smart playlists and device sync features, which seem to have largely disapeared from modern music players. You can't even create a smart playlist, which I used to do, for example the top rated tracks from the last rolling 12 months. I got Claude to make a Flutter app with a visual node programming interface for creating smart playlists . This impressed me a lot. It will not create final products and you also need to be a little competent at programming. I also use AI assistance while coding, but fiind that I don't need to use it all the time. It's been sort of like a dopamine hack, I have had a significant boost in motivation to work on all my projects again. That can't be a bad thing. However it might not last forever. I do prefer to find more ethical alternatives with transparency about energy use for example. And, supposedly the better you get at software engineering, the less you require AI or LLM asssistance.

Special note - licensing issues

Open source projects sometimes use a very strict license that means that the code used from open source is not meant to be used in certain ways, for example commercial uses, GPL for example, which if used you are required to provide your source code openly. It is currently being debated whether LLMs can be used without any license concerns any source software development, as there is a chance that LLMs could reproduce licensed code https://itmedialaw.com/en/ai-code-tools-and-open-source-licenses-risks-for-developers/. It is currently recommended to consciously ensure that your code is not using licensed code, with provided and external tools. My feeling is that LLMs for software development should be used sparingly, for example helping me get out of a programming writers block. Light code assistance for boilerplating and repetitive tasks. But not for heavier use. The legal issues are still unresolved.

I'm indecisive though

I have a tendency to be indecisive, I still can't decide what my favourite Linux desktop distro is for example. Sometimes I see worrying things online about LLMs that drive me away for a while. But then I come back. I'm also very open to discussion about AI and LLMs, if you think strongly one way or another, it's something I will not ignore. Everyone has an opinion on "AI" so no need to flood my inbox. I find face to face talks are a pleasant way to discuss complex subjects.