Okay. So yesterday I came across this tweet,
which I thought was interesting, Around the idea that LLMs are really
going to be a threat to white collar professionals when they can generate
slick PowerPoints, better than humans. And while I found this sort of,
semi amusing, I also thought it's kind of true, right? That a lot of people, their jobs seem
to just be making PowerPoints and perhaps doing email on top of that. So what I thought I would do is
try out, just to see, okay, with what's the current state of building
a PowerPoint deck with an LLM. Now I've deliberately not gone for
doing an agent kind of framework. So maybe in the future, I will do this
sort of video again with agent frameworks And look at how well that
could be improved upon. But what I do is I just take the
sort of top LLMs that we've got out there and I've chosen three of these. so I've chosen ChatGPT
Plus, which is GPT 4. I've chosen Claude 3 Opus. And I'm going to use the Google Gemini
1.5 pro model in here to try these out. So where possible I'm going to use the
same prompts for each of these models. And I'm going to try and, see
if I can try and use the same prompts all the way through. All right. So we're kicking off with
a ChatGPT Plus or GPT 4. and I basically just asked it a very
sort of generic, simple question. What is the best way to make a PowerPoint
slash side deck programmatically? And you can see that what it's
coming back with, is it suggesting a number of different packages or
frameworks that you could do this. I'm not sure about markdown or LaTex
there, but okay, we've got some things going on in the ideas in here. If we give the same thing to Claude. and you'll notice I'm trying to
use, like I said, I'm trying to use the exact same prompt each time. I did the same thing to Claude. It's also giving me a list of
different kinds of frameworks. But one of the things that's nice about it
is that it's actually giving me some sort of code examples of how I can sort of see. this as I go through here. Okay. And then finally, when I give it to Gemini
1.5 Pro, you can see that we're getting back a similar kind of things, talking
about different frameworks is stuff. Although some of the frameworks
are different in here. and it's almost like they're all
sort of trying to clarify something like, what language are you using? that kind of thing. All right. So I'm going to follow up with the same
prompt for all of them to clarify this. And what I'm going to do is I'm going
to ask it to make a set of slides about the five good emperors of Rome. So if you, come and look at,
just doing a Google search. Google generative, AI search here
for the five good emperors of Rome. It clearly knows what emperors they
are and it's giving me, this is the generators stuff with an LLM before
it actually then goes on to give me search results and stuff like that. So we know that an LLM should know
what these five emperors are et cetera. so I've clarified that. ahead of time. What I've also done is I've gotten
a couple of books about these and converted them into text files so
that we can actually try uploading the books and getting the LLM M to
use this as a resource, maybe for a second round, going through, et cetera. Okay. So here's the next prompt
that I'm going to go with. So I'm going to basically
clarify that I'm using Python. okay, let's say I'm using Python. Give me a code example that will generate
10 slides about the five good Roman emperors with, I probably want to make
that with two slides for each of them. I want it to be creative and very stylish
as if it was designed by an advertising agency, creates ads for Paul Smith. for the info on the slides
have the emperor's top five achievements For the content. Let me just put a comment there. For the content approach it as
if you are a top four consulting company, making a pitch. Actually we'll say making a pitch deck. All right. So we kick this off in here. What I'm really curious to see
is like, where do they strongly go for the same kind of thing? And where do they strongly
differ in their outputs? So I've got. GPT 4 going there. Let me just kick Claude off. So I'm really curious to see
sort of where are they similar? Where are they different? Not just in the actual
sort of specific content. But in the way that they respond, to this. It's interesting that this one got the
subtitle, a pitch by top four consulting. So maybe I should have used the
actual, consulting names there. Let's give this one to
Gemini and kick that one off. And while that one's running, let's
come back and see what, GPT 4 has done. Okay. we can see here that, they've all
settled on this particular package. which I haven't used before. So eventually I'll try and build these
and see what these actually look like. Add emperor slide. So it's got a function there. Okay. It does look like it's got the
five emperors, correct here. All right. we've got Nerva, Trajan, Hadrian,
Antoninus Pius and Marcus Aurelius. Okay. so it's putting those together. I don't really see anything
about styling in there. So this code provides basic structure
to truly capture the Paul Smith style you should enhance this with this. Okay. So we'll see if the other ones
do the same kind of thing. Okay. So it's basically kicked off. It's got Paul Smith colors. Okay. So it's very interesting. just Googling quickly what
the Paul Smith colors are. You can get a sense of, this
kind of stripey color there. so it's definitely defined
some interesting colors. I wonder how it's actually
going to use them in here. And we can see that it's also defined
nicely, the top achievements of each of these, which is, I think I said
the top five achievements from memory. Yep. I said the top five achievements. So this seems to have paid attention to
that and responded quite nicely in there. And it's basically telling us
that it's going to make this out. So this is almost like that it's saying
this is already done and ready to go. I think this one, we will
sort of try building it. Gemini has been going. so Gemini we've got, oh, we got an
interesting thing here where it's been done it for two of the emperors,
but then it's just sort of sit our ad details for, the others in here. And then it hasn't really got anything
about the Paul Smith colors or anything. Although it's got a font name,
Paul Smith's handwriting. Again, kind of interesting
because that there is like a, some kind of font like that. My guess is that it's probably
not easy to get that font. Okay. I'm not seeing a direct link to that font
when I actually come and look that up. what else would they have got in here? it's got for this, okay. The first key thing I think is I need
to tell it that you need to give me the details for all five emperors. Okay, so I'm going to give this same
prompt to the other ones, although I think the Claude one may have
already done that, but let's just say. Let's just pass that in for each of these. this one actually It looks like this
one did do something along the lines of what we actually wanted in here. And this time, we've certainly
gotten some nicer formatting coming out of it for that one. For Claude, okay, suddenly here is the
complete code with the details for each emperor and the content for each slide. Okay. So it seems I messed up, copying
it across, where for some of them, I had a slide for each of them. for the ChatGPT one I had
two slides for each of them. Although it doesn't seem like
it's made the extra slides there. it seems to be sticking to
one slide for each of them. let's continue generating. All right. So I think now what I'll do is I'll
copy the code over, install the packages and we'll start to see, actually
how well do these actually work? So I've managed to get the code
into a VS code for each of these. You can see I've basically got one for
the chat GPT slides, one for, the Claude slides and one for the Gemini's slides. I will say that the Gemini one, did
have errors in it and it required me going back to it multiple times,
asking it to fix its own errors. Cause I deliberately didn't
want to fix any of themselves. It was just basically for getting
to import different things in here. And it was only when I basically said to
it, come on, like you need to get your act together kind of thing that it was like,
I sincerely apologize for repeated errors. You're absolutely right. These mistakes aren't acceptable. And then it actually did fix
all of the code in there. So I've gone through and basically just
run each of these, and given them, just change the name of the front to be. in it so that we know where they
came from for going out here. And we're going to do next
is basically upload them. I don't have PowerPoint on this machine. so I'm going to upload them into
Google slides and we can basically have a look at what we get up
there, knowing that's probably not the perfect way to evaluate these. but at least we can sort of see
and get a sense of, what they were doing, what they were thinking. in here. Okay. I guess another way I can basically check
them is just as sort of use the preview tool In the Mac to go through them. So this is the ChatGPT
ones, they come out. pretty much normal sort of boring slides. the content, I haven't gone through
the content thoroughly, but it seems reasonably sort of right. Yeah, focuses on Marcus
Aurelius meditations. it's got some different
sort of things in there. If we have a look at the Claude one. The Claude one has the colors,
or certainly is getting at some of the color ideas from the Paul
Smith thing in there, very sort of boring layout of the slides, right? It hasn't really sort of nicely
proportioned them in any way. but we do have like, when they
actually ruled, you know, the ruling dates for each of these. Although we've got, things overlapping
each other, that, which is also not ideal. And then we can see with the Gemini one. We've basically got some sort
of repetitions of things. And I'm guessing we're missing something. This could be because we're looking
at something that's black on black. Okay. So If we bring these over and bring
them in, we can actually see what they will look like in a Google slides. So we can see the ChatGPT one, has
gotten stuff, although it looks like it's bleeding across the edge
on Google slides, although that is probably a Google slides issue rather
than a PowerPoint issue in here. again, we've got, I guess
looking at these, we've got the content in these, but we haven't
exactly got nice layout in here. okay. When we look at the Gemini one. And it's interesting that the errors
that the Gemini one was getting was it was trying to put in images. and it just was getting errors
because those images didn't exist. So you could imagine if we had
prompted it that, okay, here are the five images we've got, use those,
in here or something like that. We could've gotten some better results. this is not good though,
where it's basically giving us Trajan's accomplishments. And one of the five accomplishments
is he was considered one of the five good emperors. That's not, really holding the
context of what it is that we're sort of going for, in here. Again, with Marcus Aurelius the idea
of, that one of his achievements was the thing that we're actually making
the the deck about really perhaps signifies to me, it didn't really
understand what I was going for in here. It does seem though that the main
issue is the issue around design. what I want to do is go and have a
look at a product, which actually I use quite a bit, Which is all
about getting LLMs to create nice designs by using certain components. And while it's not actually
doing sort of powerPoint slides. it's doing websites. so let's jump in and have a look at that. Okay. So another service that is not
exactly PowerPoint, but is kind of, I think, in the same realm. is that a website's right. And this is a service called
v0.dev created by Vercel. And you can see, I basically
put it in a prompt in there. and the prompt was create a website
about the five good emperors, have sections with pictures and profiles
for each emperors showing their name. now I didn't tell her anything
about the five imprints. It's basically using the
large language model. That's running this. Guessing it's open AI for this. but it's either got a lot of
good things in context in here, or it's got a fine tune model. So what basically v0.dev
is creates a websites. and you can come through and look
at them so we can see here that this one we've got five good emperors. Trajan, Hadrian, Antoninus Pius
okay, so it's got one of them wrong. So it's missed out, Nerva the
first one and it's added Marcus Aurelius' son who was probably not
considered one of the good emperors, He was a bit of a crazy guy. if we look at, some of the other ones,
though, here, we can see, again, it's getting some of the content wrong. But definitely stylistically it's
much more kind of interesting here. Now having put in the stuff around,
the advertising agency and stuff like that, we could certainly
come through and try that out. and we can see this one hasn't done
a very good job at all on that one. but the whole idea with V0 is
that once you've got something, you can then reprompt it. so now I can basically ask it, use
the colors of a Paul Smith design. let's see, what we'll basically
do is take what we had there and then sort of remake it. so obviously we looked before
at the Paul Smith colors are very flamboyant, very colorful. And it's kind of gotten
the colorful thing. I'm only seeing two colors in here
though, so that's not exactly ideal. but one of the things that I'm
kind of thinking is that, you could imagine a service or an agent that
allows you then to look at the slide output, and then, sort off reprompt
it, and basically improve on it and say, okay, I like how you did this
slide, but change this other slide. I need more information about this. I need more information
about that in there. And I think that, it's
probably not that far away. my guess is that there are
probably startups out there already working on this. if there aren't already, products
that exist really for doing this. it's kind of an interesting
problem to do in that, I think it's definitely a tractable problem
that you could do with LLMs. The issue becomes, how do you
deal with design features? and putting in colors, et cetera. So I'm just taking the
colors from the Claude one. And I'm just gonna say,
Remake it with these colors so you can see that even
as it's remaking it here. it wants to stick to a limited
number of colors on the page. Now that could be something in its prompt
about, telling it not to be sort of just too many colors or something that
stick to sort of two or three colors. and hone in on those, which it seems
to have done with both the yellow version and this other version as well. Anyway, I thought this would be a
fun little exercise, just to get a sense of how, different or how
similar the ChatGPT models, the Claude models, the Google models are when we
actually start putting these together. as you can see from the slides
that we got out still got a while to go before we could just do
this with the off the shelf, LLM. We're probably going to need an agent
we're probably gonna need, some kind of functional prompting and, some system
that understands about design a bit more and understands about how to take it. So I think it's interesting that
the LLMs have been able to get the facts, that we're sort of after. and you could imagine if you preloaded
them with a context for your company, for your project, that kind of thing. it will be able to get that. and especially if you imagine it,
that in that context, you've got things like a recording of a meeting
that you had about the project. a set of sort of agreements notes
that you came about for that project, that kind of thing. it should be able to do
those sorts of things. I think the challenge is going to be
making the design look really sleek. And also getting the style in a way that,
people are going to be happy with it. So I think There are some slide creation
tools out there that once you've got your facts in there, you can sort of just
click the button and keep changing styles until you get something that you like. And you could imagine that
would work really nicely here. Anyway, this is just a fun,
little play full thing that I wanted to try out on the weekend. let me know what you think. let me know your ideas on
how to make this better. if people are really interested,
maybe I look at sort of writing an agent to do this or writing
some kind of thing to do this. And have a go at making it
better as we go through. Anyway, as always, if you found
the video interesting, please click like, and subscribe, and I will
talk to you in the next video. Bye for now.
