blank

Reasons why using Large Language Models will not Work

Posted on by Toma Velev

While they’ve achieved impressive results in many areas, Large Language Models have some reasons why they might not be suitable or effective for certain tasks and will not Work for you. Here are some potential drawbacks.

Computational requirements

Training and using large language models require significant computational resources, which can be a barrier for many organizations or individuals. I’ve experimented with several Open Source Large Language Models and tools that run them. They all require significant computer and Good Nvidia GPU. There are options to run things on CPU, but things will run slow and unusable for public usage.

Biases and unfairness

Large language models can perpetuate existing biases in the training data, which may lead to unfair or discriminatory outputs. This was demonstrated during US Elections 2024.

Inability to handle nuance

Large language models often rely on surface-level features and patterns, which can make it difficult to capture subtle nuances in language, such as sarcasm, idioms, or figurative language. My girlfriend that is not techy had a conversational fight with it multuple times. It takes a lot of prompting to get the value in specific topics and areas.

Insufficient context understanding

While they’re great for processing text, large language models might not fully grasp the context in which the text is being used. This can result in misinterpretation or misapplication of the information. You may extract info from LLMs from things that it has parsed. There is still information that has not yet been digitalized, indexed by search engines nor processed through Machine Learning.

Vulnerability to adversarial attacks

Reasons why using Large Language Models will not Work – for Good. Large language models can be susceptible to intentional manipulation or spoofing, potentially leading to misinformation or deception. Deep fakes could be recognized – fast – by people that work with them, but – it may fool non-tech individuals.

Lack of transparency and interpretability

How do you test that the LLM works for you. Human Wisdom and understanding is complex and not totally mathematical. Additionally – the complex internal workings of large language models make it challenging to understand how they arrive at their conclusions or predict outcomes.

Dependence on specific architectures and tuning

The performance of large language models is often highly dependent on the specific architecture, hyperparameters, and tuning used during training. This can make it challenging to reproduce results or adapt models to new tasks.

Lack of control over output generation

Large language models can generate text that may not always be accurate, relevant, or appropriate for a particular context. This requires testing a lot with prompting and figuring out the appropiate system message – passed to the LLM engine.

Overfitting

Large language models can easily memorize the training data rather than learning generalizable patterns. This leads to poor performance on unseen data.

Lack of common sense

Despite their vast size, large language models often struggle with understanding common sense and real-world knowledge. They might not generalize well to situations they’ve never seen before.

Limited domain knowledge

While large language models are great for general-purpose text analysis, they may not be specifically designed or trained for a particular domain (e.g., medicine, law, or engineering). This can lead to inaccuracies and misunderstandings.

While these limitations are important to consider, they don’t mean that large language models aren’t useful tools. Rather, it’s essential to understand their strengths and weaknesses to effectively apply them in various domains and applications.

You could check out my tools that I’ve build on top of self-hosted GPTs: