Home » Uncategorized » LLMs Demystif-AI-ed: A Gentle Introduction


Econ One’s expert economists have experience across a wide variety of services including antitrust, class certification, damages, financial markets and securities, intellectual property, international arbitration, labor and employment, and valuation and financial analysis.


Econ One’s resources including blogs, cases, news, and more provide a collection of materials from Econ One’s experts.


Get an Inside look at Economics with the experts.
Data Engineering Director

M.P.A. specializing in Data Analytics & Security Policy, Columbia University

B.A. in Security Policy, Duke University

Econ One (Los Angeles, CA), 2020-Present

International Trade Centre (Geneva, Switzerland), 2018-2020

Duke Social Science Research Institute (Durham, NC), 2016-2018

Share this Article
October 5, 2023

LLMs Demystif-AI-ed: A Gentle Introduction

Author(s): Miles Latham

In recent years, the realm of Artificial Intelligence has seen remarkable strides, ushering in an era of unprecedented potential. Advancements in training techniques and the ensuing leaps in performance have propelled Generative AI and Large Language Models (LLMs) into the spotlight. The speed at which ChatGPT, a product of OpenAI, garnered its vast user base upon its release in January 2023, underscored the growing prominence of these innovative technologies. As other tech giants scrambled to introduce their own iterations, it quickly became evident that LLMs will critically shape the technological landscape of the coming years.

What are LLMs?

LLMs are AI systems that are designed to process and analyze massive amounts of natural language data and use that information to generate responses that closely mimic human conversation. Given their impressive capabilities, they are among the most sophisticated and accessible solutions in the realm of natural language processing (NLP) today. LLMs fall in the domain of Generative AI- a broad category of AI systems focused primarily on content generation, in contrast to other AI functions like data classification, pattern recognition, or choosing actions (like steering autonomous vehicles). While other forms of Generative AI like image and audio generators have also found utility in the workplace, LLMs or text-based content generators are at the forefront of business applications due to their adaptability to diverse job functions. The versatility of LLMs is undeniable, with common use-cases spanning content creation, Q&A tasks, code completion, language translation, and text summarization.

From Definition to Operation

A ‘Language Model’ uses probability distributions to predict the next most appropriate word or phrase with the given context. Instead of just focusing on grammatical validity, the models learn the essence of human communication to be able to construct coherent and contextually appropriate sentences. Underpinning this process of analyzing input text and formulating an output are parameters. These are internal components of the model that numerically encode or capture a wide array of linguistic patterns and relationships learned during the training phase. For example, parameters enable the model to distinguish between the meaning of “apple” when referring to the company versus the fruit in your input prompt and generate relevant responses. Emphasizing their significance, the ‘large’ in Large Language Models points to the vast number of parameters which can range from millions to billions. Research has consistently found that endowing these models with more parameters, training data and computational power escalates performance. However, the term remains somewhat elusive as there is no established consensus or accepted threshold for what qualifies as a ‘large’ model., LLMs often stretch beyond the computational limits of a single machine and therefore are usually provided as a service through APIs or web interfaces. For instance, GPT-3 (Generative Pre-trained Transformer 3), one of the largest LLMs with 175 billion parameters, is made accessible to the public via ChatGPT, an AI chatbot web application.

Model Architecture and Training

LLMs are Transformer-based neural networks. Conceptualized by Google engineers in the seminal paper “Attention is All You Need” (2017), Transformer models apply an evolving set of mathematical techniques, called attention or self-attention, to capture dependencies between distant data elements in a sequence, such as words in a sentence. This allows for better contextual understanding compared to other types of neural networks, rendering these models highly effective for NLP. This intricate architecture consists of multiple layers which work collaboratively to ingest input text and yield output text predictions, constituting the foundation of LLM functionality.

The training process can be broken up into the ‘pre-training’ and ‘fine-tuning’ stages. During pre-training, the model is exposed to massive datasets so it can learn and develop proficiency in fundamental tasks such as language comprehension, analysis and generation. A notable advantage of these models lies in unsupervised learning, where LLMs discern hidden patterns within unlabeled datasets that may not be readily detectable or even intuitive to humans. This circumvents the need for the often expensive, time-consuming and challenging task of data labeling, i.e., associating each data point with a known target or output, which is one of the biggest impediments in model development.  Pre-training requires massive computational power and cutting-edge hardware. However, after this stage, the model evolves into a versatile entity, capable of servicing myriad requests. Fine-tuning, a subsequent step, involves introducing task-specific data to optimize performance for particular use-cases. For example, feeding the model more financial data, such as from 10K filings, than what it already saw in the pre-training stage can help it refine its understanding of financial concepts and parse financial statements more effectively. This stage is notably more efficient and cost-effective as it only requires a fraction of the data and computational resources employed during pre-training. This two-pronged training methodology amalgamates the model’s broad knowledge base with sophisticated task-specific prowess.

Navigating the Future of LLMs

The surge of LLMs has heralded a transformative era of communication, problem-solving and innovation. As these technological marvels permeate industries and our everyday lives, finding an equilibrium between their unparalleled potential and a critical assessment of their risks will be imperative. It is through this delicate balance that the synergy of human ingenuity and AI advancement can unlock a new world of possibilities.


Sources: https://aws.amazon.com/what-is/large-language-model/; https://developers.google.com/machine-learning/resources/intro-llms; https://www.databricks.com/sites/default/files/2023-06/compact-guide-to-large-language-models.pdf; https://www.nvidia.com/en-us/glossary/data-science/large-language-models/; https://mark-riedl.medium.com/a-very-gentle-introduction-to-large-language-models-without-the-hype-5f67941fa59e; https://cset.georgetown.edu/article/what-are-generative-ai-large-language-models-and-foundation-models/

Latest Related Resources and Insights