Inside the World of LLMs

Large language models (LLMs) such as GPT, one of the technologies that have revolutionized the field of artificial intelligence today, attract attention with their human-like text generation, language understanding and various problem-solving abilities. Although i argue that these are not real artificial intelligence, i will still try to explain how they work so impressively. At the end of the article, i will add a section with my own opinion on why they are not real artificial intelligence.

Large language models are deep neural networks with millions or even billions of parameters. These networks are trained on vast amounts of text data to learn the patterns, word relationships, and sentence structures of the language. During the training process, the model understands both the context and the connections between words, and generates appropriate utterances to follow up on a given starting text. LLMs are complex systems that process text in a series of steps, from tokenization to decoding. Understanding these steps gives insight into how these models work and produce answers. However, the topics i will discuss here are subject to change and evolution. Different methods and strategies may even emerge in the future.

Tokenization is the first critical step, where the input text is broken down into manageable units called tokens. In effect, it is the same method used when people were first taught to read and write. It is like teaching a child.

Modern LLMs often use Byte Pair Encoding or WordPiece tokenization techniques. These balance the size and scope of the vocabulary by splitting rare words into frequently used subwords.

Doesn't it sound so similar to what we were taught in our early years of education? The early years of LLMs, where we break sentences down into words and then syllables, tend to work in the same way.

The tokens are converted into embeddings, which are vectors of numbers representing their meaning. This numerical representation allows the model to mathematically process linguistic information. That is, it allows the machine to mathematically comprehend the language at hand. We can attach emotional meanings to words (not to mention memory) and remember them that way, but a machine has no emotions, so we need to give it mathematical meaning.

These are calculated in an algorithm in the background, and at this point the fact that it is not a real artificial intelligence comes to light, but it still starts to associate some things with each other using the network method in its own memory system, and this is where it starts to get closest to a real artificial intelligence.

One of the most important building blocks behind the success of LLMs is the transformer architecture. Transformers, especially thanks to a mechanism called “self-attention,” can evaluate the relationship of each word in a sentence in parallel. This method allows the model to produce consistent and meaningful output even in long texts. Self-attention helps the model better understand the subtleties of the language by determining which words it should focus on more. More modern structures can contain more complex attention mechanisms. Of course, it also needs a large memory store to do this.

LLMs are pre-trained on large datasets, with the aim of teaching the model universal language patterns. After this “pre-training” process, fine-tuning steps are applied for specific tasks or application domains. Fine-tuning makes the model more effective for specific tasks, such as text classification, summarization, or question-answering. This process allows the model to blend general language knowledge with the requirements of specific applications. This is one of the reasons why it has worked so effectively in the software field.

LLMs use advanced decoding strategies to produce consistent text. We saw the best example of this in the system developed by the Chinese. Even GPT has only recently taken a step towards this, it is a somewhat complicated strategy and is still open to development. Frankly, as of the date i published the article, i have not yet come across one that works properly, i have not seen one that can make a significant difference between generating smart answers and normal answers. I am sure we will see that this will change in the future. After the training process is complete, LLMs start producing output based on input from the user. During the output generation phase, the model uses various sampling techniques; these techniques help produce more natural and diverse texts by keeping randomness under control. For example, methods such as “beam search” allow the model to determine the most probable results, while settings such as the temperature parameter allow for creative and unexpected outputs because they control randomness.

These examples can be multiplied in large numbers. This is still an area that is very open to development, and the foundations of what we call intelligence lie here.

LLMs are used in many different areas, from customer service to content production, language translation to coding. The continuous development of technology will allow for even more precise and effective models to emerge in the future. However, ethical and security issues will continue to be on the agenda as an important part of this development process.

I need to evaluate this in two aspects. One is energy consumption and the other is humanly, this artificial intelligence craze that is currently pushing the boundaries of technology leads to incredibly high energy consumption, in order to keep these artificial intelligence tools alive, large data stores need to be created and powerful chips need to be operated. In addition to this benefit, it means a huge waste of energy for the world, so i think that we are still in the early stages of this system that they call artificial intelligence and today's technology is forced to keep this structure alive and causes such a waste. In human terms, i am in favor of using these products, which have the capacity to end many professions that will put many people out of work, within limitations, it is a good thing that information can be accessed very easily, but bad information can also be accessed very easily in addition to good information, although ethical restrictions are tried to be imposed for this, it is very easy to bypass and these artificial intelligence tools still provide you with easy access to everything you want. I do not support this system today, which i think has major disadvantages compared to its benefits. Yes, i use these tools too, but i dont upgrade to a higher package by paying money, in other words, i try not to provide my support directly to them, i think you understand what i mean, i dont know what will happen in the future, but right now, i try not to support it as much as i can because i think it has major disadvantages compared to its benefits.

This is my own opinion, but it may be far from the general opinion. It would not be wrong to think of these tools, which are currently being introduced as artificial intelligence, as functions that perform algorithmic calculations using large data stores and give you output as a result of input by basing them on certain relational results. I think that current chip technology and data technology are one of the biggest problems that prevent us from reaching real artificial intelligence. There will be those who say that we are still in the early stages, but i am not talking about this, we can use basic methods and structures, we now know how to train, but i dont think that we will use these structures that we have built in today's technology in the future, i think that we will reach real artificial intelligence with a new or yet unknown technology, i see quantum as the biggest potential right now. These are currently auxiliary tools, but they are really costly tools for the world, when i look from the past to the present, i get hopeful when i see how fast technology is advancing and i think that the future will not be dark.