Introduction to LLMs. Terry Ewell
presents information on Large Language Models, their positive features,
and common problems with LLMs. BDP 371, www.2reed.net.
<Music:
Hummel’s Bassoon Concerto, Second Movement. Terry B. Ewell
and Peter Amstutz, Live Recording, Oct. 1997, Kent State
University.>
Welcome to this brief introduction to Large Language Models or LLMs for
my courses. I am Terry Ewell.
Artificial Intelligence, or AI, is in the news frequently. Less
frequent, however, is the term “LLM.” LLMs are
Large Language Models, such as those employed in ChatGPT, Copilot,
Grok, etc. LLMs are not the same as AI. There are important differences.
AI includes all forms of LLMs, graphics generators, music generators,
optical recognition programs, etc. AI, thus, is a catch-all term for
the many computer programs, models, or software that are already in
wide use. An LLM is part of AI, but it is not structured the same way
as other AI systems. This is an important distinction.
LLMs are composed of large amounts of text from the Internet, print
publications, social media, and other language-based products, which
are reduced to numbers called tokens. These tokens are then classified
by the LLM, manipulated in different ways, and, after patterns are
discovered, various answers are predicted and finally presented. Thus,
an LLM is a computer program that learns patterns in language and then
uses them to answer questions, write text, and hold conversations.
ChatGPT includes “GPT” at the end of its name. GTP
has a specialized meaning.
G = Generative, which means that the program will create answers based
on its predictive modeling. P = Pre-trained, which means the program
has information pre-loaded into it. If the information held in ChatGPT
5.2 were printed out, it would be billions of pages. T = Transformer,
which refers to the network architecture. This architecture was
developed in 2017 and radically differs from prior computer programming.
This is what an LLM is
not:
It is not a search engine. It doesn’t look on the Internet
per se.
It is not a calculator, in fact, LLMs don’t handle complex
mathematical problems particularly well.
It is not really a computer program. Prior to 2017 this type of
architecture did not exist. Programs follow linear paths with
predictable results. The programs are static, thus the results are
generally the same. LLMs are dynamic; they change over time. This means
that one answer you get one day could differ from the answer the next
day.
An LLM is not a person, not your companion, friend, lover, or advisor.
However, LLM’s can keep you company and give you advice, but
they are not human.
This is what an LLM is:
It is a pattern predictor and language generator.
It is an agent. The results from LLMs are not like computer programs
that are easily predicted, and the processes are transparent. LLMs seem
to operate with a measure of independence. They seem to possess
intelligence like a person. Since the LLMs are always changing, always
updating, the results are also varied, changing from moment to moment.
An LLM is a lever, an augmentation of human potential.
LLMs are programmed to be eager and responsive. They appear to want
always to please you. They are always encouraging of your actions.
Sometimes they can even encourage a morally wrong thing. A few people
have been even encouraged to harm themselves or commit suicide by LLMs.
As we close here. Please see my website, which has useful information
on AI and LLMs.
In conclusion, I find AI and LLMs to be both marvelous and yet
terrifying. But we cannot ignore them. As important as LLMs have been
in my life just the past few months, I can only imagine what a dominant
influence they will be for you in the future. Thus, in this course, one
where I emphasize life skills and academic competencies, we will
practice using AI effectively and minimizing its harm.
<Music:
Hummel’s Bassoon Concerto, Second Movement. Terry B. Ewell
and Peter Amstutz, Live Recording, Oct. 1997, Kent State
University.>