Brainpower - Best of replay: Prompt Fundamentals with ChatGPT

Summary

In this episode of 'Brain Power', Josh explores the basics of prompting in large language models like GPT and Bard. He demonstrates how to build effective prompts by adding roles, tasks, and detailed instructions to improve the quality of responses from AI models, using the example of creating learning objectives for cooking a smoked potato salad.

Takeaways

• 🧠 The episode focuses on the fundamentals of prompting in large language models like GPT, Bard, and Claude.
• 💻 It emphasizes the importance of compute power and data in the functioning of large language models.
• 📚 The script explains how large language models are trained and the role of human intervention in setting guardrails for responses.
• 🎲 The concept of probability in generating responses from language models is discussed, highlighting the variability of outcomes based on input prompts.
• 📝 The video introduces the idea of 'zero-shot prompts' where the model is expected to generate a response without prior examples or additional information.
• 📈 The role of adding a professional role to prompts, such as 'instructional designer', is shown to influence the quality of responses.
• 📑 The script demonstrates the process of building up a prompt by adding specific tasks and detailed instructions to refine the model's output.
• 🥔 A practical example is given using the task of creating learning objectives for cooking a smoked potato salad.
• 🔍 The importance of specificity, measurability, achievability, result-orientation, and time-bound criteria in crafting effective learning objectives is discussed.
• 🔗 The video mentions the availability of additional resources and prompts for learning and development, guiding viewers to access them.
• 📚 Lastly, the script encourages viewers to follow along with the examples and try using the prompts in their own large language model interactions.

Q & A

What is the main topic of the episode of 'Brain Power'?

The main topic of the episode is exploring prompt fundamentals and how they work in large language models like chat GPT, Bard, and Claude.

What are the two important aspects of a large language model according to the script?

The two important aspects are compute power, which is necessary to drive the model and provide results in a timely manner, and data, which includes the corpus of information used to train the model and the guardrails in place with that information.

Why does Josh mention the difference between GPT 3.5 and GPT 4 in terms of speed?

Josh mentions the difference to highlight that there are variations in the speed of responses between different versions of the model, with GPT 4 being faster than GPT 3.5.

What is a 'zero-shot prompt' as mentioned in the script?

A 'zero-shot prompt' is a simple request given to the model without any examples or additional information, relying on the model's training to provide a relevant response.

What is the role of 'control statement' in prompting?

A 'control statement' is used when prompting to ensure that the model understands the context and limitations of the request, preventing inappropriate or irrelevant responses.

Why does Josh create a new chat for each prompt in the exercise?

Josh creates a new chat for each prompt to avoid any influence from prior prompts, ensuring that each 'zero-shot prompt' is independent and not affected by previous interactions.

What is the purpose of adding a 'role' to the prompt?

Adding a 'role' to the prompt, such as 'act like an instructional designer', influences the quality of the information returned by the model by aligning the response with the expertise associated with that role.

What are 'SMART' criteria used for in the context of learning objectives?

SMART criteria are used to describe a learning objective in a way that is Specific, Measurable, Achievable, Result-oriented, and Time-bound, ensuring clarity and effectiveness in the objective.

How does adding detailed instructions to a prompt improve the results from the model?

Adding detailed instructions to a prompt provides the model with more specific guidance on what is expected, leading to more accurate and relevant responses.

What is the significance of the 'Plinko game' analogy used in the script?

The 'Plinko game' analogy is used to illustrate the concept of probability in how a large language model generates responses based on the input prompt.

Where can viewers find additional content and support for the episodes?

Viewers can find additional content and support at Josh Cavalier's website or by accessing the prompts provided at JoshCavalier.com/brainpower.

Outlines

00:00

🧠 Introduction to Prompt Fundamentals

In this introductory segment, Josh welcomes viewers to an episode of 'Brain Power' focused on understanding the basics of prompting in large language models such as Chat GPT, Bard, and Claude. He emphasizes the importance of compute power and data in the functioning of these models, highlighting the role of human training to guide responses and ensure safety. Josh also mentions his plan to post additional content to support the episode, inviting viewers to follow along with their own language models and to check out his resources for further learning.

05:02

🎲 Understanding Probability in Prompt Responses

Josh explains how prompts work in large language models by drawing an analogy to the Plinko game from 'The Price is Right'. He illustrates how the model's responses are based on probability, with certain words having a higher likelihood of being followed by specific others. This randomness can lead to varied outcomes even from the same prompt. He encourages viewers to experiment with free-form and structured prompts, demonstrating the difference between a simple prompt and one that is more refined with additional information.

10:02

📝 Crafting a Zero-Shot Prompt for Learning Objectives

Josh demonstrates how to create a zero-shot prompt by asking the language model to generate a list of learning objectives for cooking a smoked potato salad. He discusses the limitations of such prompts, which rely heavily on the model's training data. Josh critiques the initial response for its lack of specificity and measurable outcomes, emphasizing the need for clear, detailed objectives. He then shows how to refine the prompt by adding more context and detail to improve the quality of the model's response.

15:04

👨‍🏫 Role of an Instructional Designer in Prompting

Josh explores the impact of adding a role to a prompt, specifically that of an instructional designer, to influence the language model's response. He shows how this role can guide the model to provide more relevant and targeted learning objectives. Josh then further refines the prompt by adding specific details about the task, such as making a visually appealing and tasty smoked potato salad, to elicit more detailed and relevant objectives from the model.

20:06

🔍 Enhancing Prompts with Detailed Instructions

In the final segment, Josh emphasizes the importance of adding detailed instructions to prompts to achieve more precise and actionable learning objectives. He uses the SMART criteria (Specific, Measurable, Achievable, Result-oriented, Time-bound) to guide the language model in generating objectives that are clear and measurable. Josh demonstrates how these detailed prompts can lead to more effective outcomes, showcasing how the model's responses evolve to include time-specific goals and more precise actions.

Keywords

💡Prompt Fundamentals

Prompt fundamentals refer to the basic principles and techniques for effectively communicating with a large language model. In the video, this concept is central as the host discusses the process of building and refining prompts to elicit better responses from AI models like chat GPT. The script illustrates this by showing how simple prompts can be expanded with additional information to obtain more accurate and useful results.

💡Large Language Models

Large language models, such as GPT 3.5 or 4, are AI systems trained on vast amounts of data to understand and generate human-like text. The video emphasizes the importance of these models in modern AI applications, particularly in how they process and respond to user prompts. The script provides examples of interacting with these models to achieve desired outcomes.

💡Compute Power

Compute power is the computational capacity required to run large language models efficiently. The script mentions that this is an important factor in the performance of models like chat GPT, as it affects the speed and responsiveness of the AI system when processing user inputs.

💡Data Inputs

Data inputs are the corpus of information used to train a language model. The video script discusses the significance of the quality and type of data used for training, as this directly influences the model's understanding and the accuracy of its responses. The guardrails placed on the data inputs are also highlighted as a way to control the model's behavior.

💡Probability

In the context of the video, probability refers to the likelihood of certain words or phrases being chosen by the AI model in response to a prompt. The script uses the Plinko game as an analogy to explain how AI models generate responses based on the statistical likelihood of word sequences, emphasizing the variability in AI-generated text.

💡Free Form Prompt

A free form prompt is an open-ended request given to an AI model without specific constraints or structure. The video script contrasts this with structured prompts, showing that while free form prompts are simpler, they may not yield as refined or targeted results as more detailed prompts.

💡Structured Prompt

A structured prompt is a more detailed and specific request that includes additional information to guide the AI model's response. The script suggests that while not the focus of the episode, structured prompts can be used to achieve more refined responses by incorporating advanced techniques.

💡Learning Objectives

Learning objectives are specific goals that outline what a learner should be able to do by the end of a lesson or training session. The video script uses the creation of learning objectives for cooking a smoked potato salad as an example to demonstrate how to interact with an AI model to develop detailed and measurable educational goals.

💡Zero-Shot Prompt

A zero-shot prompt is a request made to an AI model without any prior examples or additional context. The script describes the use of zero-shot prompts in the video's experiments, highlighting that they rely heavily on the AI's pre-existing knowledge and training to provide a relevant response.

💡SMART Criteria

SMART criteria are a set of guidelines for writing effective learning objectives, standing for Specific, Measurable, Achievable, Relevant, and Time-bound. The video script incorporates these criteria into the prompts to demonstrate how detailed instructions can improve the quality of the AI model's responses.

💡Instructional Designer

An instructional designer is a professional who creates and manages educational content. In the video, the role of an instructional designer is added to the prompt to influence the AI model's responses, aiming to generate learning objectives that align with best practices in instructional design.

Highlights

Introduction to prompt fundamentals in large language models like chat GPT, Bard, and Claude.

Explanation of the importance of compute power and data in the functionality of large language models.

Discussion on the role of human training in models like GPT 3.5 and GPT 4 to establish response guardrails.

Illustration of how large language models operate based on probability using the Plinko game analogy.

Differentiation between free form and structured prompts for interacting with large language models.

The concept of zero-shot prompts and their reliance on the model's inherent training.

Demonstration of creating a learning objective for cooking a smoked potato salad using chat GPT.

The impact of adding a role (e.g., instructional designer) to a prompt to refine the model's response.

Enhancing prompts by detailing the task to improve the specificity and quality of the model's output.

The significance of using specific instructions within prompts to guide the model's response.

Example of refining a prompt by incorporating SMART criteria for learning objectives.

Observation of improved results when adding detailed instructions to prompts in chat GPT.

The influence of prior prompts on subsequent model responses and the strategy of creating new chats for each prompt.

Josh Cavalier's offer of additional content and resources to support the episode's topics.

Promotion of Josh Cavalier's prompts for Learning and Development and the associated worksheet.

Conclusion summarizing the process of building effective prompts using role, task, and detailed instructions.

Transcripts

in this episode of brain power we are

going to explore prompt fundamentals

let's go ahead and jump in

[Music]

hello everyone it's Josh thanks for

showing up for another edition of brain

power

today we're gonna get really basic and

talk about prompting and how it works in

large language models like chat GPT Bard

and Claude so today's episode we are

going to start with some very simple

prompts and then build them up and talk

about the reasons why you would want to

go in and modify your prompts with

additional information to get back

better results now if you're following

along at home this is actually a

recording typically I will do this show

live live but because of prior

commitments you're now watching a

recording which is happening on January

19th

2024 okay but you could still follow

along with me if you have a large

language model open up like chat GPT

today I'm going to be using GPT 3.5

model so even the free model you can use

here today or if you want to use

Microsoft co-pilot Bard Claude whatever

your flavor of large language model is

you can go ahead and follow along with

the prompts also I will be posting

additional content to support these

episodes they are not up yet but if you

go to Josh

cavaliers. brainpower or Josh

cavalier.pdf

not up yet but coming very soon so you

may want to check it out over the

weekend right okay without further Ado

let's go ahead and let's jump into

prompt fundamentals the first thing I

want to talk

about is how a large language model

works there's really two things that are

incredibly important when it comes to a

large language model the first one is

compute power

right so you need to have the

computational power to drive the large

language model to give you results in an

appropriate period of time if you've

worked with chat GPT you understand that

there are differences between the speed

of 3.5 and four okay so that's the first

one the second

one the data right so the inputs are

incredibly important what information

or Corpus of data was used to train the

model once you have that information

what guard rails are in place with that

information and in regards to chat GPT

and the 35 and four models humans were

used to help train the model and the

responses that were coming back again to

put Protections in so that it isn't the

wild west and you can really ask it

anything that you want uh I know that

when I first started working with chat

GPT you're going to see this in a moment

but I have a control statement that I

use when I

prompt how do I cook a smoked potato

salad well when I first started

prompting I was asking it how do I smoke

a potato and it thought I was really

trying to smoke the potato like a

cigarette and wouldn't allow me to do it

so yes these models are trained and they

are

based on probability I know that for

some of you when you're working with

chat GPT it feels like a Google search

or it's grabbing the information from

the internet and that's not the case at

all um you have to keep in mind that we

are working with a vector database and

everything that the model was trained on

are just points of information and so

when you type in a prompt there are

words that are coming back and it's all

based on probability now what I want to

do is go ahead and show you an example

here and if you have ever watched the

prices right my gosh I hope that you at

least caught one episode The Price is

Right youve probably seen the Plinko

game and the Plinko game is when they

have a chip and the chip drops down and

it's going to go ahead and you know hit

the bottom of the board with an

amount so in this addition of the Plinko

game this is the probability example of

a prompt in chat GPT or really any large

language model and you'll notice that I

have a prompt over here on the Le hand

side that says as I cross the street I

noticed from out of nowhere a

huge what

like what's going to come up next in

regards to the response from chat GPT

well based upon probability if I use the

word

black or if I use the word huge the next

word could be dog it could be black it

could be rain it could be bird all these

various different options are possible

and again it's all based on probability

now some words have higher probability

than others and in this instance dog so

if I get a response back from chat GPT

from out of nowhere a huge

dog was running towards

me again that could be completely

different based upon various factors and

what the model comes back with and so if

you have worked with chat gbt and you're

wondering well why

am I getting a different response from

the same

prompt well this is the reason why okay

again probability kicks in and all it

takes is a single word or a different

word to appear to give you a completely

different

result all right so now would be the

time that you would want to open up chat

GPT or whatever large language model

you're working with and I'm going to put

some prompts in here again you can type

these prompts in if you want or once the

documents are uploaded you could

download the sheet and copy and paste it

from the

worksheet so this concept I first want

to talk about is a free form prompt

versus a structured prompt if you take a

look at the prompt over on the left hand

side you could see that it's just simply

sentences it's just like plain language

to the model that you're giving it

within the prompt over on the right hand

side is a structured prompt I'm not

going to get into that today but there

are some Advanced Techniques that you

can use to get a more refined response

with additional information within the

prompt for today because we are talking

about prompt fundamentals we just want

to keep it simple and begin with a free

form prompt just by writing

sentences all right so today we are

going to be working with a learning

objective and learning objectives are

extremely important if you're an

instructional designer uh again just

giving Focus to the uh instructional

content uh that you are creating and

making sure that it's very specific and

measurable and that you have guidance uh

towards those objectives again whether

it's knowledge skill attitude Behavior

whatever the case may be so that is the

topic for today is creating a learning

objective but we're going to do it in

chat

GPT all right so now out of the gate

let's just go ahead and open up chat

GPT and

just start with a very simple request

all right so I'm going to go ahead and

switch

over to chat

GPT and again it's the it's the 35 model

here right

so uh it's not the four it's the again

the free model so that's

fine and now we're going to give it a

very simple

prompt now remember I I do have a

control statement that I use so I'm

going to add that in here and so for

this

prompt I'm going to go ahead and

say create a list of learning objectives

to cook a smoked potato

salad seems like it's going to work I

mean you know there's nothing um crazy

about this prompt it's just a simple

request

when you uh when you create a

prompt and you give no examples or

additional

information about that

prompt uh this is what we call a zero

shot prompt okay what we're doing is

we're leaning really hard on the model

to again we hope that learning

objectives was trained in the model

somewhere

right uh if it wasn't you're going to

get back really horrible results but uh

you know with 175 billion parameters in

the 35 model and over a trillion

parameters in the gp4 model odds are

learning objectives and its definition

are in there but now you're going to see

the results that are going to come back

and how we are going to coax the model

to give us better results as we prompt

so out of the gate let's go ahead and

try this simple prompt create a list of

learning objectives to cook a smoked

potato salad and let's let it rip and

see what

happens all right so you know one of the

first things that I see here uh with

these learning objectives is one it does

give me many different bits of criteria

uh about this smoked potato salad if I

move back up to top you can see that the

topics include safety cautions selecting

ingredients potato preparation smoking

techniques and so on but the language in

here it could be way better um things

like or words like learn and

understand uh that doesn't really give

me anything specific about the task at

hand uh as far as it's measurement right

and so I'm looking for some learning

objectives again that are very specific

that are measurable achievable result

oriented time bound uh if it's specific

task and I want more detailed learning

objectives so I mean this is a good

start actually I was it's pretty

impressive how um it has all these

different uh criteria in here listed

okay so not too bad now for this

particular exercise I am going to create

a brand new chat for each prompt and the

reason for that is I don't want any kind

of influence from a prior prompt when

you're prompting in chat

GPT you have to keep in mind that when

you use words and a prompt and then you

get a result back all of those words

your prompt and the return result are

going to influence the

conversation the words that are in there

and related words have a higher

probability of showing up so I want to

again for this exercise create zero shot

prompts every single time so you could

see the difference between the results

now if I were to go ahead and attempt to

ask it and coax it towards better

results in here I could do that in the

same conversation this is what we call a

prompt chain but again we want to remove

that and just see

with the zero shot prompt what we get by

modifying our prompts each time all

right so let's go

back now I want to go ahead and create a

brand new

conversation and this time we are going

to add in

here a

roll we're going to add a role to the

prompt and in this case it's going to be

act like an instructional designer

now you know when we're talking about

you know adding or uh you know modifying

a

prompt in this case by adding a role the

words instructional designer beside

themselves is going to influence the

quality of the information coming back

just because there are words related to

instructional designer that deal with

learning objectives hopefully it's

trained that way in the model but

because you know these models are

trained on all kinds of information on

the internet odds are that relationship

is there right so again we are now going

to add a role in here and take a look at

the results all right so we'll go back

over to chat

GPT and we'll modify the

prompt in here and say act like an

instructional designer create a list of

learning objectives to cook a smoked

potato salad and let's take a look at

the

results all right

well very interesting it looks

like we have objectives in here and it's

going to go in and give us certain

outcomes

so we do have

some modifications of the learning

objective by the end of the lesson okay

that's that's good um Learners will be

able to identify and gather all

necessary ingredients for making a

smoked potato salad all right not too

bad I mean we could be better here with

the learning objective uh but again we

can still see language in here Learners

will

understand all right these are a little

bit

better but we can continue continue to

modify our prompt and uh to get better

results in there right and we want to

continue again with our experimentation

and building up until we have extremely

good quality learning objectives so the

next item

here that we want to talk

about is the task

itself now I mean you could keep it

simple in regards to the task but the

more

information the more details that you

give about the task is going to give you

better results so let's go ahead and go

back and we're going to add some

additional details to the task and take

a look at the results from chat GPT so

back over we go we're going to create a

brand new chat and for this particular

prompt we'll keep in there act as an

instructional

designer create list of learning

objectives to cook a smoked potato salad

that is visually appealing and

tastes fantastic to your

guests all right so there is

the

prompt and I apologize for not switching

the screen but there it is now that you

have

it let's run it and see what we get

okay so now we have each you know bit of

criteria in here and we have identify

and select key ingredients explain the

role of each ingredient demonstrate safe

and efficient knife skills describe

select but you can still see we still

have like understand in

here I'm looking for better verbs than

that but these are getting better all

right you can see that you know if we

actually could combine uh some of these

bullet points together into one sentence

uh we probably would

have uh some decent learning objectives

again depending upon the criteria in

here all right so there's that but again

the additional words of visually

appealing and tastes fantastic is going

to

influence the criteria and the

information within those learning

objectives all right now let's take a

look at this last bit of information

here

so one additional thing that you can add

to a prompt in addition to the task are

very specific

instructions and so in this example it

says learning objectives are only one

sentence learning objectives contain a

goal Behavior Criterion and conditions

now depending upon where you learned how

to write a learning objective um it

could vary depending upon again what you

are looking for in that learning

objective um now in our prompt today we

are going to use the smart which is

specific measurable achievable result

oriented in time bound uh criteria or

the description of a learning objective

and so that's what we're going to be

using here

today and those are going to be our

detailed instructions for the learning

objectives so back over to chat GPT we

go we are going to create a brand new

chat and for this last prom

prompt let me go in and grab that

prompt and paste it in

here act like an instructional designer

create a list of learning objectives to

cook a smoke potato salad that is

visually appealing and tastes fantastic

to your guest learning objectives are

specific measurable achievable result

oriented and time

bound learning objectives are only one

sent

so those are our detailed information

and you know it

really doesn't matter what task you are

trying to perform within chat

GPT the

role the task and the detailed

instructions alone if you add those

items in there are going to give you

fantastic

results okay and you'll notice that

again as we've built this prompt up and

we've added that additional information

in there we have gotten back better

results now it should be interesting

here when we run this prompt especially

the time specific I think you're going

to see in these learning objectives that

there's going to be time values

associated with some of these tasks or

some of these learning objectives well

let's check it out and see if it happens

so back we go to chat GPT let's go ahead

and let's run this

yeah so we definitely have instances

where time is now a factor by the end of

this cooking lesson Learners will be

able to identify and gather the

necessary ingredients for smoked potato

salad within 15 minutes Learners will be

able to properly wash peel and diced

potatoes ensuring uniformity and size

after following the provided

instructions Learners will be able to

prepare a flavorful marinade for the

potatoes consisting of measurable

quantities of herbs and spices you get

the idea

H well well well so now we have some

decent learning objectives in here uh

you could see that verbs like understand

and learn uh and even describe are not

even in there anymore U and we have some

decent learning objectives that we can

then use in our project so hopefully

today this gives you a really good idea

of

how to use

prompts first by just going in and

requesting or having a conversation with

chat GPT but then building

up building up the prompt by adding a

role of any professional doesn't have to

be instructional designer right it could

be an accountant it could be a marketing

professional whoever out in the world is

very proficient at that task you can put

in for the role

then we go in and we have a very

specific task in there right and you and

you want to be detailed with that and

any criteria or any specifics about that

task and in the case of the learning

objective we had those smart uh criteria

in there for the learning objective you

want to add that in

there and that's going to go ahead and

give you a better result so hopefully uh

today you were able to go in and and see

how to create a prompt and build it up

using roll task and detailed

instructions I want to thank you for

joining me today uh just a

couple items here uh before we roll out

uh again you can find me up at Josh

cavaliers. or Josh

cavaliers. uh also if you want to go

ahead and download or get access to some

prompts I do have 150 prompts right now

eventually within a few weeks that's

going to change to 250 prompts uh but

this link is going to work right now uh

you can just put your name in email in

there and you get access to a notion

site with 150 prompts for Learning and

Development and finally you'll be able

to go ahead and go to josh.com

brainpower and download the associated

worksheet with these episodes so with

that again thank you so much for joining

me and I hope to see you here in the

next episode take care