Large Language Models (LLMs): The Foundation of Artificial Intelligence

“Magic is just science that we don’t understand yet.” – Arthur C. Clarke.

Many things have the power to astonish humans, at least until we

unravel their inner workings. While most of us readily embrace

various technologies, the intricacies of what lies beneath often elude our

interest and curiosity. This phenomenon holds true for 99% of us. The

unknown has a captivating allure that can keep us in its thrall.

Recently, OpenAI’s ChatGPT has taken the world by storm, amassing a

staggering 100 million users within just a couple of months of its launch.

Those who ventured into its world couldn’t contain their amazement and

excitement. Social media platforms were flooded with tales of unique and

exuberant experiences. Children rejoiced at the prospect of completing

their homework without the watchful eye of their parents. Specialized

knowledge, like computer coding and more, suddenly became child’s play.

It seemed as though everyone had become a poet or a writer of elegant

prose. The excitement rivaled the euphoria of humanity’s first moon landing.

Beyond the euphoria, one might wonder how ChatGPT accomplishes

these seemingly unimaginable feats in the blink of an eye. These AI

chatbots will likely be the ordinary person’s first foray into consciously

experiencing and integrating advanced technologies into everyday life.

So, let’s break it down.

The concepts are straightforward at the core, and ChatGPT is no

exception.

The foundation of all AI applications, including ChatGPT, is the Large

Language Model (LLM). But what exactly is an LLM?

Think of it like a library full of books. LLMs are vast digital information

repositories. Advanced AI algorithms meticulously shuffle and sift through

this organized digital text data (the LLM) to produce astonishing answers

to queries. It’s akin to a librarian quickly locating a specific book from the

shelves upon request. However, the librarian is no match for the speed

and efficiency of an LLM. The capabilities and versatility of LLMs are

attributed to their size and the diverse information they contain—much

like an extensive library housing a treasure trove of various book genres.

Creating LLMs marks the first step towards implementing and integrating

AI technologies into the workplace.

Libraries can be either public (general) or private (specialized). Similarly,

LLMs can be general or specialized. ChatGPT from OpenAI, LLaMA

from Meta, and PaLM2 from Google are all examples of general LLMs.

Due to their extensive and diverse data storage, general LLMs are

larger and capable of answering various queries. ChatGPT, for instance,

is trained on a wealth of general internet data.

Specialized LLMs, on the other hand, contain very specific pieces of

data or information tailored to particular requirements. In a business

context, these smaller LLMs might house a company’s HR information,

production-specific data, sales data, and more. Companies often opt

for a gradual integration approach due to caution and other constraints

associated with existing technology.

It is important to note that even the largest LLMs to date, such as

GPT-4, has a limited range of information stored. Their ability to respond

to queries is also limited in context, although it still amazes us. Currently,

inputting data into LLMs is mainly text-based, which is time-consuming,

labor-intensive, and relatively expensive. As with any technology, costs are

expected to decrease as adoption rates increase.

Parameters play a critical role in determining the capabilities of an LLM.

In simple terms, parameters help LLMs make decisions between different

probabilistically evaluated answer choices based on the query. Larger

LLMs have more extensive parameters. The smallest LLM boasts 350

million parameters and holds 40GB of text data, while ChatGPT features

a staggering 175 billion parameters and 570GB of text data. OpenAI’s

latest creation, GPT-4, is said to be over 500 times more substantial and

effective than the earlier ChatGPT. GPT-4 can even interpret images and

sounds, although OpenAI has yet to disclose the exact specifications.

To put things into perspective, the human brain is still believed to be

millions of times more efficient than GPT-4 and similar advanced LLMs.

One day, when LLMs can store the entire internet’s data, they may claim

superiority over humans. Maybe.

Next edition on 26th Feb'24: How AI works - An Introduction

Book trivia:

·???? ‘Future of Work – AI in HR’ received a cherished endorsement from the HRD institution himself, Prof (Dr.) TV Rao : “I am reading Sreejith Sreedharan’s book. It is very well written and I recommend it to be read by everyone HR or non-HR. It portrays the possible future of work…”. Thank you once again, Prof.Rao. Your endorsement of the book will remain the most valuable.

·???? The book has secured another prestigious perch at the 北京大学 library in Beijing, where it will soon be accessible to students and faculty. I shall intimate the availability soon for the China audience.

·???? The book is cataloged and available for students and professionals who have access to the Singapore University of Social Sciences (SUSS) library.

Ref: https://lnkd.in/gniBwUuN