Transcripts
1. INTRO + WELCOME: Everybody has been
talking about chat GPT. What does this mean? What does it do and
how can it help you? This is a six part
class about chat GPT. How to effectively use it, how to get it to help you
in your daily routine. And we'll explore
some best practices. This class talks
about everything from inputting and formatting data to how to fine tune and train chat GPT to learn what
you're looking for. We'll talk about a couple
of advanced features. There's also going
to be a really interesting class
project where you can try the skills you
learn in this class. On chat to PT, will look at some use cases. And we'll talk about
best practices. Whether you're a novice at AEI or really experienced
in the field, this class should help you. This is a quick class
intended for all levels. I'm looking forward to
having you as a part of it.
2. INTRODUCTION TO CHATGPT: In this lesson, we'll cover
the basics of chat GPT, what it is, and how it works. We'll also discuss
its capabilities and how to access
and use the model. What is chat GPT. Gpt-3 is a large language model developed by a company
called Open AI. It's a variant of
the GPT model and specifically designed for natural language
generation tasks. One of this GPT mean CPT stands for generative
pre-training transformer. I know that sounds
very complicated, but that's what the chat
GPT-2 is all about. Chat is the ability to talk with a generative pre-training
transformer model. Charge. If it is a software
that's been trained on a wide range of Internet text. And it's capable of generating texts that's pretty human-like. You can ask questions and
they will answer you. You can phrase things in really human terms
and write them down and you'll get answers
back that are pretty good. Okay, So how does
the chat GPT work? It's quite complicated, but
it uses something called a neural network architecture
known as the transformer. This architecture
allows the model to understand the context of the input text and generate appropriate responses
to each question. This Models pre-trained on a massive data set
of internet text. It allows it to understand
a wide range of topics and language patterns. What are some of the
capabilities of TBT? This software can be used
for a wide variety of natural language
generation tasks like language translation, text summarization, texts
completion, content creation. It can also be fine tune for specific tasks like
question answering, text classification, and
dialogue generation. The power of this
software is madness. So how do you access it? To access or use chat to beauty? It's really easy. You can access it
through Open AI API, which allows you to send requests to the
model and receive, generate texts and return. In order to use that model, you have to sign up for an
API key on OpenAI website. There's also open
source implementations of the model
available on GitHub, and those can be finally
tuned and use locally. Once you have access
to that model, you can input a prompt and
receive regenerative response. The prompt can be
any texts at all, but it's best to
provide some context to help them model generate a
more appropriate response. But the easiest
way to access chat GBT is by going to
chat.openai.com. You have to sign up. You can sign up with a Google account. Then you can start asking
questions to the software. You can play around with it. The raw power of this free
tool is absolutely amazing. In the next lesson,
we'll dive deeper into how to input and
format data for the model.
3. INPUTTING AND FORMATING DATA: This lesson is called
inputting and formatting data. In this lesson, we're
going to discuss the different ways to
input data into chat, to beauty and the
proper formatting for different types of tasks. We're going to cover topics like context, prompt,
and temperature. One of the most
interesting things about chat UP T is the way
that you input data. You inputted a thing
exact same way you speak. You can just write questions
in and it will give you answers back in relatively
human-like form. Data can be inputted into this model and a
variety of ways. Use Open AI's API, use open-source
implementation, the model, or you can go to the website. But the most common way to input data is by
providing a prompt. And a prompt is the text that the model uses
degenerated response. The prompt can be
any text at all. But it's best to provide
some context to help them model generate a more
appropriate response. Additionally, you can
provide a context, which is a string of
texts that allows you to provide additional
information to the model. This context allows the model
to understand the context, the prompt, and generate a
more appropriate response. It's really important to format your
information properly. For your input, you want
to make sure you're being as clear as possible. The clearer you are, the better the model will be generated in response to
that meets the criteria. So proper former dean
is really important for ensuring that the model
generates response. It's relevant to the prompt. It's important to use proper grammar and spelling
when I'm putting this data. The model is trained on clean
and well formatted text. For certain tasks like
texts completion, It's important to format
the prompt as specific way. E.g. you want to provide a specific number of words,
there'll be completed. You can say, generate a ten word lists of
synonyms for raccoon. I find it very difficult
to find synonyms for RecA. And then there's temperature. Temperature is the most
interesting part of this because temperature is a parameter that controls the randomness
of the model's output. A higher temperature generates more diverse and
creative responses, but a lower
temperature generates more conservative and
predictable responses. So you can play around
with this quite a bit. You can experiment with different temperatures
to try to help you find a sweet spot that best suits the task
you're trying to do. By understanding the
various ways to input data and understanding
temperature parameter. You can fine-tune chat
to PT and then the model to generate more accurate
and relevant responses to your specific tasks. In the next lesson,
we'll cover how to fine tune and train chat to PT on specific
tasks and datasets.
4. FINE TUNING AND TRAINING: Okay, so let's talk about
the benefits of fine tuning. Fine tuning, fine tunes, the process of training and retrain model on a
new task or data set. Fine-tune it allows you to
take advantage of the models. Pre-trained knowledge will also adapting it to your
specific task. It leads to improved
performance on your tasks. The model already
learned many of the language patterns and common phrases that are
relevant to your task. So how do you do it? The process of fine-tuning,
fine-tune chat, too pretty neat to
have a labeled set of data for input and output pairs that are relevant to your task. And the labeled data set will be used to train the model
on your specific tasks. So once the fine
tuning is complete, you can use the model to
generate output from new input. There's a lot of
guessing to this. You try things out
and see if they work. It's important to have a large and diverse set
of data for training. Also, you don't just want to
throw in ideas and numbers. You want to preprocess the data to ensure
that it's clean and well formatted so that when
you input into chat GBT, It's able to understand
it pretty clearly. Experimentation is key. When fine-tuning, it's important to
experiment with different hyperparameters, like the learning
rate or batch size, number of training steps. It's also good to use GPU to speed up the
process of training by understanding and benefits and processes of fine-tuning, as well as following best
practices for training, you become able
to really improve the performance of charge
CPT on your specific tasks. The next lesson we'll cover some of the advanced features of chat to beauty that can be
used to improve performance.
5. ADVANCED FEATURES: Let's talk about some of the advanced features
of chat to beauty. In this lesson,
we're going to cover some of the advanced features of chats you could
do that could be used to improve performance. We'll also discuss how to use these features to improve the
performance of your model. Number one, controlling
the length of output. By default, chat GPT
generates texts, that's a variable length. But you can control this. You can control the
length of gender texts by specifying the maximum number
of tokens to be generated. It can be done through
the OpenAI API or by using an open source
implementation of the model. Controlling the length of
output can be useful for tasks that are things like
text summarization. You can say things like make a one paragraph long summarization
of this piece of data. Number two. Generating multiple
responses. By default, church, if you teach, generates only one response
for each prompt. But you can actually
change that around. You can generate
multiple responses by specifying the number of
responses to be generated. You can say, give me
a five-point list, build a civics lesson class, generated multiple responses
can be actually very useful for tasks like
dialogue generation, where multiple responses are
needed for a single prompt. Number three, you can
use conditional inputs. So conditional
inputs allow you to provide additional
information to model, to control the output. E.g. you can provide on
label or category to the model that will control
the topic of the output. This feature becomes useful for things like text classification, where the output needs to
be in a specific topic. So when having chatted
me to write a document, you can specify what the classification that
document's going to be odd number for
using beam search. Beam search is this
algorithm that generates multiple
output sequences, each with a probability. And the algorithm generates
multiple outputs and keeps the best one a coordinate
specific criteria. Beam search also helps improve
the quality of the output. By understanding these
four advanced features, you're able to improve
the performance of Chad GPT on specific tasks. In the next lesson, we're going to explore some of the most common use cases for chat to beauty, including
language translation, text summarization, and
texts completion will also discuss how to adapt chat
GPT for other use cases.
6. USE CASES: This lesson is all about
use cases for church GPT. What have people not use
chat GPT for nowadays? Everything from
recipe designed to song creation, to
writing emails. In this lesson, we're
going to explore some of the most common use
cases for chat GPT, including things like
language translation, text summarization texts
completion will also discuss how to adapt using
chat GPT for other use cases. And it's not an exhaustive list. This is an AI that you can use for an unimaginable
number of things. Here are some
language translation. Chat GPT can be used for
language translation. You can find in the model on a data set of bilingual texts. You can fine tune
this model to them to use to translate text from
one language to another. This use case requires
a large data set of bilingual texts and
fine-tune the model for a specific pair of
languages, but it works. And it's an interesting way
of doing that type of thing. You can use chat to PT
for text summarization. Text summarization is
incredibly interested in. You can take a huge
amount of texts like a large thing,
cut and paste it, put it into chat to PT and say, summarize it in three
paragraphs are in two lines, and it does pretty well. Chat GPT is used for
text summarization by fine-tune the model on a data
set of texts and summaries. You can also how to
summarize the text in the voice of a
character or a song. The fine-tuned model
can then be used to generate a summary
of the given text. And this use case also requires a large datas that
have texted summaries. And you can fine tune that model for specific tasks
of summarization. But it's a pretty
incredible use. It's something that we
haven't seen before, at least not on the
commercial level. Chat to critique can be
used for texts completion. You can give a little
bit of a text. You can tell chart UP te, here's the beginning
of a song, right? Another verse. Chat to be taking these
for task completion by providing a partially
completed texts as a prompt. And then the model
generates remaining test. I don't really know how it
does this, but it does. And this use case doesn't require a fine tuning
right out of the box. It works. And you can try it and
it gets it pretty well, learns the rhyme, the
cadence, the alliteration. Chat GPT can also be
used for other cases like answering questions test. And again, it does
it quite well. And you can say answered this question in the
voice of Shakespeare, or as if it was a
Seinfeld episode. Those are the things people have commonly been using as ways of showing you how smart
this software can be. Chat to be taken with adaptive also for dialogue generation, create a script and you can fine tune the model on a data set
is relevant to the task. The larger and more
diverse your data set is, and the more you find a model
of better it gets at it. The number of use cases
for chat UP T is, in my opinion limitless. You can adapt chat GPT
for your specific task. And it gets better over time, which is kind of crazy. In the final lesson, we're going to
talk about some of the best practices for using
chat too busy right now, and some tips for
avoiding common mistakes. We'll also discuss
some limitations of the model and look
at the future of AI. Text generation.
7. BEST PRACTICES: In this final lesson, we're going to provide some
best practices for using chats you beauty and tips for
avoiding common mistakes. We'll also discuss
some limitations of the model and the future of
AI powered text generation. Best practices always use proper grammar and spelling
when inputting data. The smallest mistake
can fool the system. The models trained on clean
and well formatted text, I recommend use bare metal at different
temperatures because then you can find
that sweet spot that best suits your task. What you're using chat to P24. And you can play
around with it a bit and you can also say to it, I mean, it doesn't
have feelings. You can say, no, make
this more funny. No, make this more serious. Know, make this
more alliterative. Use a large and diverse data set when fine-tune the model
for specific tasks. The more information you feed
into it, the better it is. Feeding, use them in
back out preprocess the data to ensure that it's
clean and well formatted. You can use a GPU to speed
up the process of training. There are lots of
common mistakes. Here are some tips
you can use to avoid making these mistakes. Number one, you want
to be careful using chat GBT for anything that's
sensitive or high-stakes. Because occasionally
it generated things that are pretty
offensive or harmful. Even though it's been
trained to limit certain categories of
thought and anything that's violent or
racist or offensive. It's still couldn't
get it wrong. You have to be aware
of the model is not perfect and it may make mistakes or generate irrelevant
or nonsensical text. I've been playing
around and chat to take quite a bit
and occasionally produces things that are complete mumbo
jumbo gobbledygook. You also have to be
aware that the model is not able to understand
the meaning of words. It only associates them with a probability of appearing
in certain contexts. So occasionally words will pop up out of nowhere
and you're like, why is this here? It's a big limitation. Chat GBT as a language model, it's not able to understand
the meanings of words. It only associates
them with probability, um, and so it's not perfect. It may make mistakes and may generate things that are irrelevant and may
repeat itself. So you have to be careful
around certain topics. This is a very active
field of research. There's so many new advancements and techniques being developed. Even chat to PT every couple of days it will give you a
new version, a new model. The version that was
most recent that I'm using words from
my day or two ago. But that will change in three
or four days in the future. We can expect to see much more sophisticated models with improved performance
capabilities. But it's still going to be ai is still not going to
understand meaning. It's just gonna be able
to give you things that things will be there based on all the things that seen before. There's lots of ongoing research into the field of ethical AI. Making sure that the
results that are communicated from an AI
system are safe, pair, and reliable by understanding
the capabilities and limitations of the model and
by following best practices, you're able to really
effectively use chat to PT for your
specific tasks. Try it out. Sleigh with a whole
bunch of variables. Try to figure out how
you would respond. If you will respond
to that question. It's software that is amazing and will only
keep getting amazing. It's a limitless
sandbox where you can play with things in
a variety of ways. Some of which are quite useful, and it's only gonna get better. You have to keep in mind, this is a rapidly
evolving field and new advancements and techniques are being developed
all the time.
8. CLASS PROJECT: Okay, Now it's time
to do an experiment. I would like you
to use chat GPT to create an original song in the style of your
favorite artist. And make the song about something that you wouldn't think the song would
be about naturally. E.g. write a song
about chocolate and unicorns in a style
of Depeche Mode. Write the lyrics
to the song about Disneyland in the
style of Metallica. And then once that
song is written, have chat GPT, add additional versus what chords to the song. Be as creative as possible. I'm really interested to see
what you've come back with.
Daniel Berkal, Consumer Research
