LLMs: areas of excellence and limitations

Large language models (LLMs) are foundational models trained on extensive datasets, equipping them with the ability to comprehend and produce natural language as well as other content types for a broad spectrum of purposes. Thanks to  ChatGPT, artificial intelligence has caught the attention of a wider public. And now, generative AI has the peak interest of major organizations and small businesses alike. LLMs are making their way into organizations that are focusing on adopting artificial intelligence across numerous business functions and use cases.

When approaching any new technology, a good question to ask is whether it has any blind spots and in what areas it really excels. To give an answer to this question, we will explore what LLMs are, where they fit in the field of AI, and briefly go over the mechanisms of their work to understand why they excel in some spheres better than in others.

Artificial Intelligence (AI) is a rather wide-ranging discipline and consists of several branches. In order to properly understand how LLMs function, one should have an understanding of where they stand in the wide world of Artificial Intelligence. Let us illustrate the scope of the AI field as layers, illustrated in the image below. From the outer to the inner, these layers are:

In order to grasp what LLMs can do and where their limitations lie, let us briefly consider the workings of each of these layers. Aside from Artificial Intelligence, which is too broad for our purposes, we will commence our journey right from Machine Learning.

Machine Learning

The purpose of Machine Learning is to identify any patterns in the data, especially the dependency of output on input. The complicated nature of the latter requires ever more advanced algorithms. In general, the increase of inputs and classes increases complexity, and more complicated dependency needs greater amounts of data to learn from. This is where Deep Learning comes into play.

Deep Learning

As mentioned previously, complicated interdependence between inputs and outputs, along with the high amount of variables involved, requires us to have a more robust and versatile model. Thus, we come to neural networks, which are loosely based on the structure of the brain and consist of numerous layers – hence, the name Deep Learning, which refers to their great depth. With their depth, these algorithms can become very big – for example, ChatGPT uses neural networks consisting of an astonishing 176 billion neurons, exceeding even the human brain's 100 billion neurons.

Their design is quite simple in essence. Consisting of layers of interconnected "neurons", they use inputs and predict their outputs. In theory, one can think of deep neural networks as linear regressions with non-linearities, stacked layer by layer.

Large Language Models

The definition of a Large Language Model may be summarized as a highly complex neural network that is trained on large volumes of data (text) to solve language-related problems. Due to their ability to work with large volumes of text data, the Large Language Models can predict subsequent words in a sentence and can thus be used for text generation.

There are several stages in the training of Large Language Models, which include pre-training, instruction fine-tuning, and reinforcement learning from human feedback. The stages mentioned above are designed to improve the model's performance and allow it to act not as a text predictor but also to be able to assist the user properly.

Additionally, it is necessary to note that because of its high performance in working with sequential data, the transformer model architecture forms the basis for the largest and strongest LLMs. The transformer model is what stands for T in Chat GPT; its core idea is to focus on the most relevant aspects by ignoring the others – just as humans do.

LLMs are quite advanced tools for natural language processing based on Deep Learning that allow generating text similar to human one.

Nowdays, large language models are making progress in numerous spheres, ranging from search engines, NLP (natural language processing), healthcare, robotics, to programming code generation. Given the way in which their functionality is provided, it is clear why LLMs outperform in all language and text-related tasks.

• Text generation. It should be stated that LLMs perform exceptionally in generating linguistic expressions that are fluent, concise, and correct.

• Language comprehension. LLMs demonstrate high-level skills in understanding the language, which includes tasks such as sentiment analysis, text classification, and factual knowledge processing.

• Reasoning in its simplest form. Assessing the capacity of the model, one should pay attention to how successfully LLMs manage problems related to arithmetic, logical, and temporal reasoning. Although at times LLMs cope with such tasks in case if the context is simple, it cannot be ignored that large language models are capable of remembering relations between repeatedly encountered word pairs and making simplistic generalizations based on previous experiences. The consequence of this process is difficulty in understanding something radically different from what has been processed before.

• Understanding of context. There is also an excellent understanding of context by the large language models. As a result, it allows them to produce coherent outputs that relate to the context of their input. Nevertheless, one should remember that guaranteeing the correctness and contextuality of information is a difficult task.
• Performing NLP tasks. In addition, it can be stated that the large language models perform successfully on different NLP tasks.

Within the context of Natural Language Inference (NLI), LLMs exhibit subpar performance, and they struggle in properly representing disagreements among humans. NLI requires one to establish if a particular "hypothesis" is true based on a "premise."

• Credibility. LLMs may have issues with credibility since, at times, they create fabricated information or inaccuracies during interactions. In instances where an LLM does not know the answer to a posed question, it may create fake facts.
• Bias. LLMs can learn and even promote offensive and derogatory aspects of language found within the dataset used to train them. This consideration should be taken into account while using LLMs in any consumer or science-related application since they may introduce biases.
• Citations. LLMs can come up with texts that contain citations that may seem valid. The fact is that the LLMs do not know anything about the Internet because the Internet does not exist for them, nor can they memorize the sources of the training data they have been learning from. This problem can be at least partially addressed with the use of search-augmented LLMs that are able to search the Internet and other sources for providing more accurate information.
• Mathematics. LLMs can answer simple mathematical questions incorrectly. They have been trained on huge amounts of text material, whereas expertise in mathematics may need a totally different approach to training. Here the use of tool-augmented LLM can help.
• Prompt hacking. Prompt hacking can be defined as manipulation of LLMs in order to make them generate certain, often inappropriate or harmful content.

• Abstract reasoning. As far as abstract reasoning is concerned, LLMs are quite limited in their capabilities and prone to mistakes in complicated situations. Nevertheless, recent prompt engineering solutions have helped to mitigate some of these problems.
• Relevance. Another shortcoming of LLMs lies in their inability to process timely or dynamic data and incorporate it into the analysis of different phenomena.

By analyzing the strengths and shortcomings of LLMs outlined above, as well as the underlying principles that help these models succeed, it is possible to suggest a number of guidelines for interacting with LLMs and other text-based applications, including GPT:

Exercise autonomy

While reducing LLMs to the next iteration of predictive text might seem simplistic, it is still useful. Asking a cutting-edge predictive text application to write a new piece of writing character-by-character might prove fun but not necessarily practical. After all, any application is only as good as the person who uses it.

Augment, rather than replace

In the vast majority of cases, it would be wise to consider LLM automation a cybernetic augmentation—an extension that multiplies effectiveness through proper usage rather than a full replacement.

Workflow clarity

Timing plays a major role in determining AI’s effectiveness. For example, in image creation, AI can be used to help create thumbnails, but not necessarily in the final creation process. In addition, when writing an academic paper, while AI can help in outlining the structure of the paper, it may only apply to certain parts of the paper. There are also cases where the AI-generated content can be used straightaway, such as automating activities in a roleplay game or generating visuals for a blog post.

Embrace iteration

Every AI experience should be treated as an experiment, with every result being an unfinished project. Although techniques may differ depending on the program used, iteration should always come first in all forms.

Confirm

Large language models usually produce unpredictable and occasionally new outcomes, which include made-up facts. Relying on these outputs uncritically could lead to serious issues, which could range from comical to grave depending on the use. When automating processes, make sure that you compare what the AI produces with your goals and make adjustments as necessary.

In conclusion, although large language models provide immense possibilities and capabilities, there are also some difficulties associated with them that users need to understand.

Does the LLM only revolve around the prediction of the next word, or is there something else involved? Some scientists believe the latter to be true. In their opinion, to perform effectively in the prediction of the next word under various circumstances, the LLM needs to gain a compressed perception of the world inside itself. Thus, it differs from the concept of a machine's ability to mimic the world without fully understanding it, the languages, or any other topics.

Nowadays, no clear distinction can be established between the two theories since they might be talking about one thing from different angles. The LLM has indeed been proven to be exceptionally useful by being able to show excellent knowledge and reasoning skills as well as general intelligence. Nevertheless, the similarities between human intelligence and the ability of an artificial entity to demonstrate its intelligence still remain unclear.