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.

The field of AI is quite broad and encompasses several branches. To better understand how LLMs work, we need to be able to comprehend where they fit in the vast world of Artificial Intelligence. Let's visualize the AI field in layers, as seen in the image below. From the outer to the inner, these layers are:

To understand the capabilities and limitations of LLMs, let's take a brief look at how each of these layers works. Apart from Artificial Intelligence, as this layer is too broad, we'll start directly from Machine Learning.

Machine Learning

The objective of ML is to detect patterns within data, particularly the correlation between input and output. The intricacy of this relationship calls for increasingly sophisticated ML models. Typically, as the number of inputs and classes grows, so does the complexity. Additionally, more complex relationships demand larger datasets for effective model training. That's when Deep Learning makes its entrance.

Deep Learning

We've discussed how complex relationships between input and output, coupled with a large number of variables, demand more powerful and flexible models. That's where neural networks chime in, loosely inspired by the brain and often spanning many layers, hence the term Deep Learning, indicating their immense depth. This depth allows them to be extraordinarily large. For instance, ChatGPT is built upon a neural network boasting a staggering 176 billion neurons, surpassing the human brain's approximate 100 billion neurons.

Their architecture is deceptively simple. Comprising layers of interconnected "neurons," they process input signals to predict outcomes. Conceptually, they resemble multiple layers of linear regression sprinkled with non-linearities, enabling the modeling of highly intricate relationships.

Large Language Models

A Large Language Model is basically a sophisticated neural network trained on a vast amount of data (text) for language understanding tasks. By processing vast amounts of text data, these models learn to anticipate subsequent words in a sentence, which allows them to perform text generation tasks.

The training regimen for Large Language Models involves several phases, including pre-training, instruction fine-tuning, and reinforcement learning from human feedback. These phases aim to refine the model's ability to understand and respond to user input effectively as well as to be an assistant rather than a text prediction generator.

It's worth mentioning that due to its computational efficiency in processing sequences, the transformer model architecture serves as the foundation for the largest and most powerful LLMs. The transformer is the T in Chat GPT, and the main principle behind it is focusing the model's attention on the most important task at hand and tuning out everything else, much like the human brain does when concentrating.

Large Language Models represent a significant advancement in natural language processing, using Deep Learning techniques to comprehend and generate human-like text. Despite their complexity, ongoing research and development efforts continue to enhance their capabilities and applicability.

Large language models are currently helping make advancements in various fields, including search engines, natural language processing, healthcare, robotics, and code generation. Taking into consideration the methods behind how they work, it's not surprising that they excel in all language and text-related tasks.

• Text generation. LLMs excel in generating linguistic expressions that are fluent, concise, and accurate.

• Language comprehension. LLMs exhibit notable proficiency in comprehending language, including tasks such as sentiment analysis, text categorization, and processing factual information.

• Simple reasoning. Benchmarks for assessing LLM's capabilities are how well they tackle tasks like arithmetic reasoning, logical and temporal reasoning, and more complex tasks like mathematical reasoning and structured data inference. Though sometimes LLMs are able to take on these tasks when the context is simple, the reality is that Large language models remember connections between constantly reoccurring pairings of words and make simple generalizations based on what they have learned. This leads to a struggle to understand things that differ immensely from the processed data.

• Grasp of context. LLMs show a strong understanding of context, enabling them to generate coherent responses that are aligned with the provided input. However, it's important to note that ensuring the accuracy and contextual relevance of information during inference remains a challenge.

• NLP tasks. LLMs also achieve impressive results across various natural language processing tasks, including machine translation, text generation, and question answering.

In the sphere of Natural Language Inference (NLI), Large Language Models (LLMs) demonstrate performance that falls below average and encounter challenges in accurately representing human disagreements. NLI involves determining whether a given "hypothesis" logically follows from a "premise."

• Credibility. LLMs may lack credibility, sometimes generating fabricated information or inaccuracies within conversations. When asked a question, the answer to which they don't know, LLMs can generate false information that sounds confident but is incorrect. 
• Bias. LLMs can learn and potentially amplify harmful content sometimes presented in their training datasets, containing toxic language elements such as offensive, hostile, and derogatory language. Consider this when using LLMs in consumer-facing applications or scientific research, as this issue may lead to biased results.
• Citing sources. LLMs can generate text that cites sources that seem plausible. However, these sources may be entirely made up since LLMs don't have access to the Internet and the ability to remember where the data they were trained on came from. This issue can be partially fixed by using search-augmented LLMs that can search the Internet and other sources to provide more accurate information.
• Math. LLMs can give incorrect answers even to simple mathematical tasks since they are trained on large volumes of text, and excellence in math may require another training approach. Using a tool-augmented LLM can help with this issue by enhancing the capabilities of an LLM with specialized tools for tasks like math.
• Prompt hacking. Users can manipulate LLMs to generate specific content. This can be used to trick the LLM into generating inappropriate or harmful content and is known as prompt hacking. 
• Abstract reasoning. When it comes to abstract reasoning, LLMs have constrained abilities and are susceptible to confusion or errors in complex scenarios. However, innovative prompting techniques have partially addressed this issue.
• Relevance. LLMs have limitations in integrating real-time or dynamic information, making them less suitable for tasks requiring up-to-date knowledge or rapid adaptation to changing contexts.

If we take a look at the areas where LLMs excel and their pitfalls discussed above, as well as the principles that make LLMs work, we can anticipate several rules that can be applied when working with LLMs or text-based applications like GPT:

Exercise agency

While labeling LLMs as the next generation of predictive text may oversimplify, the comparison remains informative. Requesting an advanced predictive text system to generate a novel one word at a time, while entertaining, could be more efficient. Remember, the effectiveness of a tool depends on its context and the user's proficiency and awareness.

Augment, rather than replace

In most cases, it's advisable to view LLM automation as a cybernetic enhancement—a tool that magnifies effectiveness when used appropriately—rather than a complete substitute.

Clarify workflow

The timing of AI implementation significantly influences outcomes. For instance, in visual work, AI might assist during the thumbnailing stages but not in the final execution. Similarly, in writing a research article, AI might aid in organizing an outline but may only be suitable for some of the article's content. There are tasks where raw AI output may suffice, such as streamlining tasks in a private roleplaying game or creating illustrations for blog posts.

Embrace iteration

View each interaction with AI as an experiment and treat outcomes as ongoing projects. While methods may vary across apps, platforms, and applications, continuous iteration is crucial across all formats.

Verify

LLMs tend to generate unexpected and sometimes novel results, including fabricated facts. Unquestioningly trusting these outputs can be risky, with potential consequences ranging from amusing to severe, depending on the application. When automating tasks, always compare the AI's output against your objectives and adjust accordingly based on accuracy.

In summary, while Large Language Models (LLMs) offer significant power and versatility, they are accompanied by challenges that users must acknowledge. Recognizing and addressing these challenges is essential to utilize LLMs more effectively and responsibly.

Is the LLM solely focused on predicting the next word, or does it encompass more? Some researchers argue for the latter, suggesting that to excel at next-word prediction across various contexts, the LLM must have developed a condensed understanding of the world internally. This perspective contrasts with the notion that the model simply memorizes and replicates patterns observed during training, lacking genuine comprehension of language, the world, or other subjects.

Currently, there is no definitive answer between these viewpoints; it may represent different perspectives on the same concept. LLMs have undeniably proven to be highly valuable, showcasing impressive knowledge and reasoning abilities, and perhaps even demonstrating hints of general intelligence. However, the extent to which this mirrors human intelligence remains uncertain, as does the potential for further advancements in language modeling to enhance the current state of the art.