0:00

Coming up next on Hands On Windows, we're

0:02

going to take another look at Copilot and

0:04

some of the things I've learned about using

0:06

this generative AI and a

0:09

few tips and tricks. Podcasts you

0:11

love. From

0:14

people you trust. This

0:17

is Twit. Hello

0:23

everybody and welcome back to Hands On Windows.

0:25

I'm Paul Th rot and this week we're going to

0:27

take another look at Copilot. More

0:30

specifically, what's the right

0:32

way to use Copilot? This

0:35

generative AI tool from Microsoft

0:37

is spreading across the ecosystem.

0:39

It's everywhere. It

0:41

is improving at a rate that is so

0:43

quick and so frequent, it's hard to keep

0:46

track of. It's being integrated

0:48

into everything that Microsoft makes on the client

0:50

side from Windows to Microsoft 365 to

0:53

the Microsoft Edge web browser. It's available on

0:55

the web. It's

1:00

getting better on almost a weekly

1:02

basis. I had said previously

1:04

we'd keep coming back to Copilot

1:07

and we will. In

1:09

using it over the past several weeks, I've started to come

1:12

to an understanding of the ways it works well,

1:14

the ways it doesn't work so well. I

1:17

thought maybe we could take a step back

1:19

and go through that list of tips and

1:21

tricks or rules or whatever you want

1:24

to call it about maybe the proper way,

1:26

if you will, to use Copilot

1:28

with the understanding that a couple of months from

1:30

now we'll probably be evolving that list as well.

1:33

Here's what I have for now. I'm

1:36

going to do all of these

1:38

demos in a web browser. Not

1:41

using the Copilot that's built into Windows. That's

1:43

only because the version on the web tends

1:45

to be more up to date. There

1:47

are features that are available in it that are not

1:49

yet available in the version that's in Windows. The

1:53

other thing that's interesting about this is that I'm

1:55

using Chrome in this case, which is kind of

1:57

goofy, but instead of Microsoft Edge you can use

1:59

any browser you want. want and I'm also using

2:01

a secondary account. So this is the account that

2:03

I use for the book and for this podcast

2:05

typically it's not my own personal

2:08

account. I actually pay for CoPilot Pro and get

2:10

additional features but I wanted to make sure that

2:12

what I was showing here was the experience

2:14

that everyone would get. Okay

2:17

so and actually let me also bring up Google

2:19

search so we can make a few comparisons here

2:21

as needed. When

2:26

it comes to CoPilot and AI and this

2:28

kind of new capability I think a lot

2:30

of people look at it as a replacement

2:32

for search or maybe as a replacement for

2:34

the personal digital assistance we've used like Google

2:37

Assistant to the Alexa's Amazon Assistant and so

2:39

forth. And yeah I

2:41

mean maybe right but it's

2:44

also important to know that CoPilot

2:46

in this case or generative AI in general is

2:49

not the same as search. These are two different

2:51

types of tools they work differently and the way

2:53

that you interact with them will be different. And

2:56

the problem with that is that I think a lot of people will

2:58

go into this and say well this is just like this

3:00

is search. I'm gonna ask if these really just

3:03

you know quite discrete little barky questions and

3:05

that's not something that's always gonna work out

3:07

very well. Traditional search to

3:09

my mind anyway is the place you go when you have

3:13

a question and you need an answer and it's

3:15

something very specific or maybe

3:17

you're researching just

3:20

something specific not a general big

3:23

type of a thing. You don't necessarily

3:25

want advice you want an answer maybe that's the

3:27

way to say it. Generative

3:30

AI is for creating content that's where

3:32

the name comes from right. And in

3:34

the case of CoPilot most of the

3:36

content that it generates today is textual

3:39

or image-based nature but

3:41

this platform is extensible and so we're already seeing

3:43

plugins and other types of add-ins that allow you

3:46

to create other types of content including for example

3:48

music. But there'll be

3:50

more of that in the future that's coming

3:52

that's what we're doing. So for example one

3:54

of the Google is selling me on their

3:56

AI as well. So one

3:58

of the questions So, for example,

4:00

you might want to know what is the capital of Massachusetts.

4:03

When you do this with CoPilot, you

4:06

get this, it's like, it's chatting. The

4:08

capital of Massachusetts is Boston. What is this

4:10

other information? I didn't ask you for the

4:12

history of Boston, but it's going on and

4:15

on. And this is the type

4:17

of problem that I have had with digital assistants.

4:19

You ask it a question, you ask it to

4:21

do something, and it talks a lot.

4:23

And it's like, I just want you to do the

4:25

thing. If you go into Google

4:27

search and type in that thing, I don't even have to hit

4:29

enter. It says it right there. There you

4:31

go. Question answered. That's the type of

4:33

thing that search is good for. This is

4:35

the type of thing that generative AI

4:38

or CoPilot is not good for.

4:40

What is three plus three? Very

4:42

just finite type questions. So

4:44

this will make more sense as we go through a couple of

4:47

demos. You'll see some of the things it actually is pretty good

4:49

at. So what is

4:51

it good at? But one of the things it's

4:53

really good at is summarizing things. And

4:55

I'll find an article that I wrote. I'll try to find

4:58

one of the longer ones. Maybe

5:02

something about Windows 11. So Windows 11 version

5:04

24-H2. Actually,

5:06

this is a good example of where

5:10

I could ask it to. Can

5:14

you summarize

5:17

this page? So

5:21

what this will do is typically you get kind of

5:23

a bulleted list, and it will just give you a

5:25

kind of a nice short summary. I could have been

5:27

more specific here, but it's just

5:29

looking at that article and

5:33

providing a nice summary. Now, this is not

5:35

a particularly difficult to understand article. It's probably

5:37

a couple thousand words long, but this

5:39

is really neat. And if you have a

5:42

really long Word document, a PDF or a

5:44

web page, CoPilot

5:46

is a great way to get

5:49

a summary of those things. Now, the

5:52

web version of CoPilot, I don't believe you can actually

5:54

pull a PDF into it, but you can do that

5:56

in Word if you have

5:58

a paid version of CoPilot. So this is. the

6:00

type of thing, the capabilities vary

6:02

on it by implementation, but an

6:05

excellent use case for this type of thing.

6:08

One of the other things we've learned about AI

6:10

is that it needs to be grounded, which is

6:13

one of those beautiful terms that we never used

6:15

before we had AI. There's all kinds of terminology

6:17

associated with AI. I

6:19

try to avoid as much of it as possible. What

6:22

that means is rather than have it

6:24

work against the body of information that this

6:27

AI was trained on, which in this

6:29

case is the open AI chat GPT,

6:32

enormous LLM, large language

6:34

model, another great AI term up in the

6:36

cloud, that is the entire body of information

6:38

out in the world up to a certain

6:40

point in time. If

6:43

you want to learn something very specific, if

6:45

you're talking about or trying to learn about

6:47

a specific topic, whatever it might be, it

6:50

would help for the data that it's trained

6:53

on to be limited

6:55

to that use case. There's

6:58

various ways that that kind of thing occurs.

7:00

We talked about one earlier. It's these custom

7:02

GPTs you see over here on the right.

7:05

For example, you could

7:07

go in to copilot generally and say, hey, I would

7:09

like a recipe, and it would probably give you a

7:12

pretty good recipe. If you go into

7:14

the cooking assistant and ask, the

7:16

body of work that it's working against,

7:18

the body of learning that it has

7:20

is much smaller, and it's going to

7:22

lead to results that are a, delivered faster

7:25

and b, are more accurate because when AI

7:28

is not grounded, that's when you get the hallucinations as

7:30

we call it. We used to call those things bugs,

7:32

mistakes, but today we have fun words for things. You

7:37

have the GPTs, custom GPTs that

7:39

Microsoft provides. There's one called designer,

7:41

which is for images, although I've

7:43

done it and will do it.

7:46

If you have a client image and will do that,

7:48

it's fine. But vacation planner, cooking assistant, fitness trainer,

7:50

etc. If you have a paid account,

7:53

which I do, but I'm not signed into, you can create

7:55

your own as well. So you can feed it your

7:57

own data, and then as we move

7:59

forward in time, we're going to we're starting to see

8:01

solutions like I have

8:03

it work off the data that I have in OneDrive, right?

8:06

Which is something you might want to do in as

8:08

part of a company where they're paying for this, you

8:10

know, product up in your organization and

8:12

that you can work across all the

8:14

data just inside your organization. So grounding

8:16

is one of those key

8:19

elements of generative AI that

8:21

makes this thing that's already amazing even

8:24

more amazing. Let

8:27

me get back to regular copilot here. One

8:30

of the other things I've learned here is that it

8:32

really pays to be specific. This

8:35

is not a great, great example but like I

8:37

said, you can create images with this thing. So

8:40

I'll do just to be

8:42

a little visual here, you know, create me a photo

8:44

of a man riding on a bike, right? And

8:47

I've done this enough times that I know exactly the

8:49

kind of image this thing is going to put up

8:52

but it's colorful bright sunny

8:54

day, trees, you

8:56

know, beautiful scene, not photo

8:58

realistic is what I've been seeing so far. So there

9:00

you go. Yeah, we got the sun, the coastline in

9:03

this one and they're fine and maybe this is what

9:05

you want. If it is, you

9:07

can click in here, grab

9:09

one of the images and download it or you could

9:11

go through and select the style.

9:13

If you're paying for the product, you actually

9:16

get additional photo features like 16 by 9

9:18

aspect ratio support and so forth. This

9:20

is limited to square but whatever. But

9:22

maybe this is not what you're looking for. So

9:24

from here, you could

9:27

keep adding on, right? Because it's a conversation.

9:29

I'm going to talk about that a little

9:32

bit more later but you can also just

9:34

be more specific, right? So

9:36

this doesn't work as well with images. I'll just

9:38

warn you in advance but I had put this

9:40

together. So this is a more

9:43

detailed prompt which is what these things are

9:45

called again, terrible terminology. But the question you're

9:47

asking or the text that you're chatting here

9:50

is create a realistic photo of a

9:52

man riding a bike seen from a distance on

9:55

an otherwise empty road. He's wearing a business

9:57

suit. There's a briefcase in the basket

9:59

of his bike. Interestingly, there was a basket on the bike

10:01

up here, even though I didn't ask for that. He's

10:04

moving very fast. There are buildings on one

10:06

side of the road, a tree on the other side, a bird

10:08

sitting on the branch. It's sunny, but

10:10

there are a few clouds in the sky, right? So

10:13

super specific. I

10:15

found we're going to go through an example of this with text

10:17

where it will make a little more sense. I

10:20

have found that this works a lot better with text

10:22

than it does with images. In

10:24

fact, one of the things that's interesting about images is that

10:26

it shows you how many

10:28

mistakes AI can make very clearly because

10:30

what you'll see is that in

10:33

some cases, despite the specific, how

10:35

specific it was, there will

10:37

be some mistakes. Although actually right here, it's pretty good. So

10:40

one mistake is the bird is flying in the air,

10:42

right? It's not sitting on the branch. So it didn't

10:44

draw an image where it made sense for there

10:46

to be a bird in the branch. And then this one, it looks like the

10:48

bird's sitting on top of a building. So the

10:50

bird is the size of Rodin, which

10:53

is kind of funny actually. The

10:57

way I arrived at this was I actually

11:01

have to make a PowerPoint presentation and this is

11:03

not something I do a lot anymore. And

11:06

this is kind of interesting to me because when I was

11:08

thinking about AI last year, this was

11:10

the example I came up with. There's

11:12

a guy in an office, his boss

11:15

sticks his head in the doorway of

11:17

his office and says, hey, I

11:19

need you to make a big presentation about whatever topic at

11:21

the next annual meeting or whatever it is. And

11:24

what would you do? I don't know how to use, maybe

11:26

I'm a good writer, I'm an Excel guy, I have crunch numbers,

11:28

whatever it is, but now I have

11:30

to make a presentation. So you could watch YouTube

11:33

videos, buy a book, find

11:35

an expert in the company, something. There's

11:37

all kinds of different ways. But with

11:39

AI, you can, at the

11:41

time it was theoretical, say I need to do this

11:43

one thing once and I don't need to learn until

11:46

I don't want to become an expert. Maybe

11:48

it would make it for me. But

11:51

web version of this will not make a

11:53

presentation, but it will

11:55

give you the outline. And If you

11:57

use the version that is in Power BI, it will give you

11:59

the outline. The point if you're paying for. Copilot.

12:02

Pro or Microsoft Three Sixty Five or

12:04

in Google sheets if you're paying for

12:06

Gemini. He. Will actually make the presentation

12:08

for your which is really cool but for here

12:11

we'll just we'll just do the outline and you'll

12:13

get an idea of how that my work so

12:15

both make very non specific straight a presentation of

12:17

a famous quotes. From. Famous

12:19

people. So innocent,

12:21

Low but and damn. It.

12:24

Will pick the people obviously didn't specify

12:27

any mesa Kurt Vonnegut, Oscar Wilde, Jerry

12:29

Seinfeld, etc. Of. Li Hitler doesn't

12:31

make this list that a as getting better

12:33

that gonna think that got a guy and

12:36

then I got is it skipped. Kept.

12:38

On some reason. Sit

12:40

with us ten. That's when it's about. I didn't

12:42

tell how me, that's why. such as Cisco. okay,

12:44

that's that's the gave me a bunch of puts

12:46

a snake or to. That's not exactly what I

12:48

was looking for it, but how would it know

12:50

that? I wasn't very specific. So. I.

12:54

Came up with a smear thing you would do an.

12:57

Is. A kid in school, right? This is not presentation

13:00

giving him as an adult. Create.

13:02

A presentation with a title side and content slides

13:04

and a thank you slide at the end with

13:06

contact information. Each. Of the ten

13:09

concept content slice include a famous quote

13:11

from a famous individual. Plus. Some

13:13

representative photo and or background image.

13:16

The. Famous people should include in than a list

13:18

out the people. Humorously, by the

13:20

way, only list out nine people even though I said I

13:22

needed ten of them. And.

13:24

Now it will start searching for quotes

13:26

from people that I had listed though

13:29

Gates, Steve Jobs, Nelson, and L, et

13:31

cetera et cetera. And.

13:34

You. Get the spotlight. It's pretty good, right?

13:36

Assuming by the way that these quotes are

13:38

accurate, this is another kind of a tip

13:40

within a tip. Ah, Ai is not always

13:42

accurate. You need to fact check the stuff.

13:45

I don't know if these guys said these

13:47

things so I would have to look that

13:49

stuff up. It's interesting to me that he

13:51

gave me one photo here at the top.

13:54

I don't know why only one, but. I think

13:56

they'll be links to more am smart least

13:58

more now here at the bottom. I

14:00

can go through the web and find those

14:02

images, etc. it's are obviously. This.

14:05

Works better if you're in the toy assault.

14:07

This is the type of thing you wouldn't

14:09

do And the web like this. you would

14:11

do it in Powerpoint slides, whatever he is,

14:13

so that's interesting. Tied to

14:15

this. This this interaction were having

14:17

here is a conversation. Meaning, it's it's

14:19

It's a two way thing, right? So.

14:22

It. Will prompt me like a did here

14:25

at the bottom with ideas for that. Ways

14:27

we could expand on that original prompt and

14:29

provide more detail after the fact right at

14:31

a slight about the context. The Beach quote.

14:34

And if you click there's lots to and I guess

14:37

it will give me those. Additional

14:39

context, right, which is pretty cool. That's.

14:41

Great. I mean, that's that's that's useful, but

14:43

you don't have to be prompted by it.

14:46

You could reap Bree prompt. You can go

14:48

back and say okay, I look, I I

14:50

like this one, but change their said cetera.

14:52

So it it's that there's an ongoing. Conversation.

14:57

I guess that were. This

15:00

is interactive. Anything's going and I think

15:02

that that's. Kind of interesting. Also,

15:04

it's worth. I didn't.

15:06

I. Didn't. Do. This here.

15:08

I'm not doing it right now, but I do

15:10

have that presentation that it made for me. And.

15:13

And I used. Copilot

15:15

Pro in Microsoft Excel Srm Access

15:18

have performed on the web. And.

15:23

This is kind of fascinating because

15:25

it designed the. Of.

15:29

This. The slide deck right. And

15:31

a pretty to the slides. The.

15:33

Photos are not of the people and

15:35

the thing that's missing his the quote.

15:38

there's no closest sifts eight. I tested

15:40

this across different A eyes he is

15:42

Gemini. I used a chat gp T

15:44

Plus and they all did things a

15:47

little differently. The. Text

15:49

posts on here is actually pretty useful. This

15:51

is what I would call speaker notes, right?

15:53

So ideally the slide would have his name

15:56

is pitcher and then a quote, but then

15:58

I could have this in my. So

16:00

I could. Use. It may be as

16:02

background information for I might say. So.

16:05

But as point to write like I said, you

16:08

have to check. The. Accuracy of what

16:10

it does but in this case because is

16:12

creating something very specific. You also might want

16:14

to change the seem the design. You obviously

16:16

would have to change the photos. A

16:20

I in many ways generative. A I Especially as

16:22

a starting point, right? It's. It's

16:24

it's not necessarily gonna create the.

16:27

The. End products right and especially this

16:29

early stage of the game. I would

16:32

say. You. Are

16:34

smart to check such miss surface of the accuracy

16:36

and and to make sure you getting exactly what

16:38

you want and then doing the work yourself. It's

16:40

that we're not going to sit back and have

16:42

a do it for us. Because.

16:44

The next up after that is the gives you

16:46

the presentation and gonna need it anymore either So

16:48

you still have a role to play here. it's

16:50

that's an important to. That's

16:54

an important element of it, and an ad

16:56

kind of another bonus. Many to put one

16:58

of the weirdest things about a I to

17:00

me anyway as that. These. Answers

17:03

are provided on the fly and are

17:05

different every time. It's probably for these

17:07

things will be subtly different, but they

17:09

can be profoundly different. Ah

17:11

miss I were to go back and do the. This

17:15

again at. A

17:18

distance. To the getting as create a photo of. The

17:22

man. My insides do that again. I'm

17:24

going to get different pitchers. They're going

17:26

to look similar. You. Know that's the thing

17:28

I is A. I've done this one enough to know the going to

17:30

be really similar but they're not the same. In

17:33

if you lose track of something that you were working

17:35

on, the foreign can't get back to it. And

17:37

have to ask this question again. The.

17:39

Results are going to be different and that's

17:41

another issue with that accuracy of actually a

17:44

completely different. That's cause that's getting away because

17:46

it's and a demonstration that you could just.

17:50

Copy. Paste do

17:52

it again and. I. Would imagine

17:54

and again this is based on having done this

17:56

a few times today. it should look

17:58

something like what we decided it's interesting to me

18:01

how different those were from the original images and that's

18:03

the point. That

18:05

could be useful by the way. It's weird but depending

18:07

on what you're creating, it might

18:09

be nice to have different versions of things. But again,

18:11

that plays into your role as the actual creator at

18:13

the end there, right? You're, it's

18:15

ultimately it is up to you. So you can

18:17

take, you wouldn't do this so much with the

18:19

images but from textual responses, the quotes, you

18:22

would take the best bits and create the

18:24

document yourself, right? AI is not going to do the whole

18:26

thing for you. All

18:29

right, well that's just the basics. There's so much more

18:31

to do and learn here but we'll

18:33

return to this again and again. So there

18:36

will be more. But hopefully

18:38

this will get you started. Hopefully it was helpful. We'll

18:41

have a new video every Thursday. You

18:44

can learn more at twit.tv.

18:46

Thank you for watching and

18:48

thank you especially to our club of Twit members. We

18:51

appreciate you so much. I can't

18:53

wait.