Introduction

I thought before what I should be doing is clear, preparing these lecture notes, but I think I would give you some more background on my context.

Setting the Context

So what my challenges and problems were and what my constraints were.

And I think these are always very important when dealing with generative AI, thinking about the context.

where you will be operating.

Because depending on the situation, you might use different tools, or you might just approach the problem differently.

Utilizing Tools and AI Assistance

So in my case, I had a good idea of the content of the course, or where I would find the content.

And I also had to use some existing tools, like PowerPoint and Moodle, because those were required by my institution.

And so effectively, what I ended up doing was to use the AI tools as assistants.

So I didn't use them for research.

I used them maybe a little bit for brainstorming and then to effectively allow me to produce slides or content more effectively.

Role of AI in Supporting Tasks

And so I started by thinking about the process that I would go through.

And there would be four steps.

Find the topic.

put together the content, prepare the slides, and create an assessment for students.

Then I thought about how it is that AI can help me.

And I will share with you what I learned throughout that process.

Breaking Down the Process

So I will start by the first question I will ask.

So it's my step one.

So I am a lecturer in an engineering school.

and I'm preparing lecture notes for a course.

I will give you instructions on the different steps.

My students are first-year engineers.

As the first step, can you produce a list of potential topics?

Step One: Identifying Topics

So I missed one thing, which is the duration of the course.

Different steps.

So it's important to mention that you are focusing on different steps, because that signals to the model that you will ask different things at different times.

So this is very important contextual information.

Also, the lecture lasts an hour.

And it's the first.

It is the first.

Step Two: Personalization and Topic Refinement

so let's see what it comes up with what is economics core concepts introduction all right why is economics important for engineers

All right, that's not a bad result.

By the way, if you think that my ChartGPT is giving more concise replies, that's why because I have changed my settings.

And so under personalization,

you can actually change your settings, and you can insert some guidelines that basically allow you to tailor the way in which ChatGPT would reply to your messages.

How many people knew this feature of personalization?

OK, quite many.

Good.

So I'm asking, always explain your reasoning step by step.

Take a deep breath.

And I will give you a good mark for a great answer.

So there was actually some research suggesting that if you told your GPT that you would tip, the average quality of results improves, which was very surprising.

It was very interesting when I heard that.

OK.

Step Three: Creating Presentations

So now, can you produce an outlet for the PowerPoint?

All right.

Interesting.

A lot of the key concepts are there.

Okay.

Does anyone have any comments on this?

Yes.

How do you know that you are teaching economics?

Did I not specify that?

I think it's because of the personalization.

Maybe it saved some information about you?

Sorry?

Maybe it saved that information about you?

Possible.

But I think it comes from the personalization.

OK.

Now, how many of you know about the canvas function?

OK?

So you can actually edit directly the text here, which is very easy if you need to make some quick changes.

And

And then, in a sense, you could also export it.

This is where the real work begins, and you should be thinking about how you tweak the outline.

Fine-Tuning Content

So we're moving into, in a sense, the second step, which is putting the content together.

I want to show you a couple of tricks that I used and I found useful in this second step.

Using Tools to Enhance Content Quality

How many of you know perplexity?

okay so this is not going to be new but for example I didn't quite remember a quote by Keynes that I wanted to insert so I asked I'm looking for a quote by Keynes that speaks about practical men be

from the general theory of the unemployment.

So if you know perplexity, you know that

You can look at the quote here.

So it has basically come up with the quote that I was looking for, which can be very useful.

Then other things you can do.

You can ask to summarize PDF documents and links, perhaps from your favorites.

But I don't want to take too much time on this.

And what I actually would like to show you is how many of you know Gamma?

Not many.

OK.

So what Gamma does is basically puts together PowerPoint presentations for you.

And obviously, I would spend quite a lot of time editing this text.

But just for the purpose of this, let's actually copy and paste it.

into gamma.

And then I go on traditional format, presentation.

I continue.

Now you will have some options here.

I know I want my text to be medium.

You can decide on the tone, educational, clear, concise.

The images.

So Gamma could generate some images with AI if you wish it to.

For this purpose, let's do AI images.

And then we continue.

Now, you can choose a template.

I have my own custom template from that I imported.

So I'm going to use that.

And then it starts generating.

And so you sort of get the gist of it.

Test Creation and Automation

While gamma is generating, we can start thinking about the fourth step, which is about the test.

I said we could do a test for the students.

So can you give me 20 multiple choice

questions based on this outline, indicating the correct one.

So you see, I can make it a bit bigger because I'm still on canvas.

All right, they are right.

And in theory, I could take these questions, I could paste them on Moodle and use them to create a quiz, or I could use any other educational app that is out there to take care of these questions.

So I used, as for the outline, I would use these as a base and then change the questions.

And by the way, this is also a great tool for students.

Because students can actually test themselves

give themselves a test.

So a problem I had, or find the questions for the test.

But I would spend some time in changing them and making them also perhaps a bit less predictable.

But it's still saving me quite a bit of time in all the typing that I would need to do.

And I think that the point is not to really create the questions, but to get rid of a lot of the groundwork that goes into them.

And I remember a problem I had when I was a student was that it's very difficult to test yourself.

For example, when you're practicing an essay, you would

you would want to simulate the exam conditions, but it would be difficult.

If you know the question, you would start thinking about it before the 45 minutes that you would want to do your practice exam with.

And so this would have been very useful.

I would have given the syllabus and say, OK, give me an essay question that I can practice on for the next 45 minutes.

Efficiency in Slide Creation

OK, we go back to gamma.

I think that the quality, I generally find that the quality of the first presentation good.

Of course, there is some work.

that needs to be done in the adjustment.

But all in all, what I find is that it takes a lot less time to do the slides.

You can export to PowerPoint.

You can work on them locally and in different ways.

The PowerPoint export would be here, for example.

User Experience and Practical Tips

So I hope this was useful.

Are there any questions, comments?

So gamma, some parts, the first cards are free.

And then you start paying after, I think, 20 or 30.

And I think it's 10 euros a month.

Yeah.

Yeah.

ChartGPT, you can probably do a lot with the free version.

the pro version, I'm paying for my version, yes.

Yes.

By tipping, oh, I will show you.

So you, in the guidelines here,

you specify that I will tip you $20 if you give the right answer.

Yes.

And it still works.

So it tells you something about the unpredictability of these systems.

And this was some research that was done with undergraduate students.

with a treatment and control.

It was quite robust.

It tells you a lot about the unpredictability of the systems.

And as someone said in a session that I led, also about the fact that it has been trained on US data to a significant extent.

And so it really has the patterns of behavior that you would expect in the US, perhaps you wouldn't expect in a lot of Europe.

Yes?

It's also known that ChatGPT uses this themselves

I'm always really polite.

I always say thank you to the AI overlords, but ChatGPT really screams at the model in its internal comments that it sends to the, or instructions that it sends to the models.

It really says some awful things.

It's literally threatening them to get the answer correct, and apparently that helps to get better output.

I would be curious to hear what Claude would do.

How many people know Claude?

So it's another large language model that, compared to ChugGPT, generally has more elaborate responses.

It's great, especially for writing tasks.

And Claude is, I find it generally more, I mean, the outputs are generally more polite, I think.

I see rude nodding.

So I would be curious to see the internal workings of Claude.

Any other questions?

You had a question.

Yeah.

From 0 to 10, how good would you say it is for the work that you do?

That's a good question.

Solid seven.

Seven.

So does it make you more productive, or does it give you more ideas how to be better?

Would you say you're more productive with it?

Yes.

Do you get more free time, or do you just do more time?

Ah, interesting, interesting.

No, no, I get more free time.

I'm very jealous of my free time.

That's good.

And is it OK to use it in an academic setting?

It depends on the different institutions have different rules.

I think EZAP has generally been quiet.

I mean, the students use it.

So whenever I teach, my students in the back are always cross-checking everything I say with ChatGPT or Google.

So I'm not really a teacher.

I am a facilitator.

I'm a coach.

Principles of AI and Practical Applications

By the way, I just wanted to say that the underlying principle to artificial intelligence is called reinforcement learning, which is pretty much what we do to teach dogs or animals to do certain tasks.

So we give them a reward every time they do well.

And so it makes total sense that tipping would actually give you a positive outcome.

Yes?

Have you played around this augmented generation?

No, yeah, I've not done it personally.

I think it's a very interesting idea.

We should.

I did use it to prepare the base of a test, but then I had to adapt it heavily to increase the difficulty of the questions.

Or maybe if you need to go through a lot of scientific articles, do you use it for some preferences, some resonance, or not?

No, I've not myself.

I've not personally done that.

Have you tried Gipsy?

No, not in a great deal.

So I've played a little bit with it, but not systematically.

I've mainly used CharGPT and Cloud.

And I would use CharGPT for these simpler tasks, where it's more about generating content or summarizing, translating.

Comparing AI Models

And I would use Cloud for writing tasks.

ChartGPT, to me, in a lot of cases, sounds a bit too American, as much as I've studied in the UK.

And to me, as much as I try to prompt it to get away from, to get towards more British spelling and a way of framing things, I personally struggle.

Yes?

No, no, I've not.

Not yet.

But I've heard it's for French output.

I've heard for French, it's generally, it performs generally better for most tasks than ChiaGPT, unsurprisingly.

Strengths and Weaknesses of AI Responses

Have you ever noticed big mistakes in the results?

Identified big mistakes?

Of ChiaGPT?

Yeah.

More than mistakes at this stage, I feel.

It's more that sometimes there is a bias.

I'm more worried...

It's just a general bias on, for example, when I was preparing the notes, I mean, we looked at this with the students.

And we looked at this with the students.

And so we were discussing the advantages and disadvantages of capitalism.

And the response that Chubb GPT gave us was very oriented towards the first response after a good prompt, was very oriented towards

liberal concepts, neoliberal concepts.

And so only when you prompt it, when you mention Marx or some other left-wing economists and philosophers, it would start giving you information from that word cloud.

But effectively, so the way it works, it's effectively that think about, I think, like the sky, this universe of information, and you need to point the model towards the right cloud.

in a sense.

And if you don't do that, you get exposed to the bias of the training data, the training data being we don't know where they come from, but being mostly we think of American origin, then they will introduce a bias in the response that you get.

So it's very important that.

Whenever you're doing research or you're asking a research question, you're very much aware of that.

Of course, it didn't apply to this demo that we just had, because as I told you, I assumed that I had some expertise in the topic and I would be able to deal with that.

Concluding Thoughts

Web Search Capabilities

By the way, you can also search the web with ChaiGPT in a fashion similar to perplexity.

It's this icon here.

Perplexity is currently better on average, I think.

But you can also do it with chat GPT.

So it would give you the links, and you would be able to cross-check the information.