Welcome everybody to the HPC Cafe.

Today we do some LLM topic again, prompting for passion, free yourself from the busy work.

So we start off with an introduction, then I go a little bit into the history of prompt

engineering, the early days with completion models.

Then we look into the breakthrough which made AI so popular, basically the chat GPT effect

and how this came to life.

After that we go into a small framework, five steps to better, to do better prompts.

And we also look into prompting strategies like chain of thought and view shot.

Then I show you some interesting prompt hacks, can also do some prompt hacking.

And then we go into emergence of tools that we use and agentic behavior in AI.

After that we look at the takeaways and discussion.

So really for the introduction, at the beginning of this whole AI boom I started prompting

for fun already.

So I did some like poems and song text with AI and it was super funny.

But now as it evolved I basically prompt to get rid of the busy work.

This is mostly coding or coding boring things, creating data sets, restructuring data sets,

stuff like this I like to offload to AI because I don't like to do it as much.

So other things I enjoy is taking photos for example, so I do less of image generation,

things like this.

So my motto is really I want to prompt to get rid of work I don't like, to focus on

my passions, to focus on things I'm interested in.

And that said, prompt engineering is also something that might not be super interesting

and is already done better by AI than by people.

So if you look at this left image we see some prompts that humans wrote and then the automated

optimized prompts they look a little bit different.

So our understanding of how to prompt in LLM might be different from what's actually like

the most perfect prompt for it.

And also this job prompt engineer is not like surviving I would say.

So there was a huge hype around this new job title prompt engineer.

So this could be something that's also obsolete in the future.

We also see it with the large reasoning models that oftentimes view shot is like it doesn't

make a huge difference if you do a chain of thought prompting or a few shot prompting

because these models have such high accuracy that you don't really see a difference anymore.

So we can also save some time knowing that.

Let's go into the early days.

So when I began with this topic, GPT-2 was the latest model.

It was a completion model.

So it just completed text that you gave to it.

There was no instruction tuning so it wasn't really tuned to answer your questions.

It might hit the right answer because it just completed the text and it looked like this

would be the next token.

But it was not fine tuned to really follow your instruction and give you a correct answer.

So the completion model is basically it will predict the next token based on the input

but the model just was trained on predicting the next token.

So it was trained on plain text, maybe stories, websites or books and those completions might

not be facts for sure.

We all know those phenomena, hallucinations or making something up that sounds convincing.

So these models did this a lot more.

So when I saw those, I thought they were pretty fun but I didn't, I wasn't able to assess

that this would be like the next big thing, right?