Is A.I going to take your job?

If you enjoy these newsletters and think a colleague or a friend should read them, please  get them to sign up on this page . 📝

If it is your first time here or you want to binge-read past letters, here are all. 📰

If you found this newsletter interesting, I would love it if you could share it🔗!

👋 Hey folks, Panos Papaemmanouil here!

I'm taking over this week to tell you about all the exciting stuff happening in the AI field.

1/ Do you want to be a new-era "hacker"?

Then you need to learn prompt injection!

Prompt injection is a technique used in natural language processing and machine learning models that involves adding specific text prompts to guide the model's output. The purpose of the prompt injection is to direct the model toward generating more accurate or desired results.

Last week, I told you about the new Bing Chat by Microsoft, enhanced with AI capabilities. It turned out that some users had already found their way to "hack" the model, meaning they made it act in ways not intended.

A Stanford University student found he could circumvent previous instructions in the language model and provide new ones in their place. Other users worldwide did the same and revealed "hidden" information about OpenAI's model. 👀

This method of injecting prompts behaves like a social engineering exploit aimed at the AI model, similar to tricking a human into revealing confidential information. 🤯

After a couple of days, since all these became public, Microsoft confirmed that users could disclose "hidden" model information and released a list of rules governing the AI's behavior.

Do you want to go into the rabbit hole of prompt injections?

Then these links are for you:

• https://simonwillison.net/2022/Sep/12/prompt-injection/
• https://research.nccgroup.com/2022/12/05/exploring-prompt-injection-attacks/

Enjoy with caution!

2/ You should be convinced by now that prompt engineering is necessary if you want to use all these new AI tools to their full potential.

However, since prompt engineering is a relatively new field, finding quality resources online is challenging. But that's why you're reading this newsletter!

I've done the tedious work for you and found the best online resources. Here is a handpicked list of my favorite resources to take your prompting skills to the next level.

• Learnprompting.org
• Cohere.ai docs
• Midjourney docs
• Arthub.ai
• https://www.saxifrage.xyz/post/prompt-engineering

3/ Last week, Meta AI released LLaMA (Large Language Model Meta AI), a foundational, 65-billion-parameter large language model.

People who have been long enough in the AI field will remember that the latest Large Language Model released by Meta only survived the internet for 3 days. I'm talking about Galactica, which was supposed to be a tool for scientists and researchers, but its quality was insufficient, leaving people dissatisfied.

But this time, things are different. It's not that technology is far more advanced now than it was on November 2022.

However, now we have amazing models (ChatGPT by Open AI), so Meta AI knew that it was possible to create something comparable good. And they did it!

This is the same situation when people used to think that it was impossible to run a mile in less than 4 minutes until someone did it. And then a lot of people did it. This is the kind of conventional thinking you must overcome to create something pioneering!

Reportedly the LLaMA has comparable results with ChatGPT in various benchmark tasks.

And the most important part? It's open source!

You can find a minimal, hackable, and readable version of LLaMA in the LLaMMA GitHub repository. However, to get full access to these models, you will need to submit a request.

4/ Do you wonder what datasets are being used to train all these Large Language Models?

A lot of people think that it's "the entire web" . But what exactly does this mean?

Reading through the research paper on the LLaMA model (LLaMA: Open and Efficient Foundation Language Models), you can get an answer to this question.

Training datasets:

1. CommonCrawl
2. C4
3. Github
4. Wikipedia
5. Gutenberg and ThePile
6. ArXiv
7. StackExchange

As you can see, this collection of documents provides good coverage for general knowledge, literature, science, coding, and many more!

5/ Notion AI

If you know me, you know how much I ❤️ Notion.

Notion is an incredible piece of software that allows users to create and manage various types of digital content, including notes, tasks, projects, databases, wikis, calendars, and more. It can be used for personal productivity, team collaboration, and project management.

People who use Notion are so obsessed with it, so it's not a coincidence that there's a "cult" growing around it.

It makes more sense if you compare it to programming languages that have similarly enthusiastic communities. People love to build and share their own tools.

Last week, Notion launched the Notion AI, a new functionality that enables you to use AI inside your Notion workspace. And it's really amazing! It helps me work faster, write better and think bigger!

But here's the thing that it's important here. Notion is just an example.

In the next couple of months, you should expect to see the same paradigm in other software as well.

Integrating the capabilities of an AI model in your app can give you a huge boost to improve its quality! Think about it. It's like giving each one of your users their own personal assistant.

📥 I'm curious to know what you think about this.

That's all for now. See you next week! 🚀

- Jon V (as in Victory) & Panos 🚀

- Fri 24 February 2023