Hey, Don't Follow Those Instructions

The double edge sword of improvements in AI model's capabilities

Welcome to another edition of the best damn newsletter in AI.

This free newsletter is designed to keep you ahead of the curve and open your mind to using AI in your work and business.

Digging deep in AI for work or AI Operations? Take a look at our membership.

Our #1 goal is to be useful. So please shoot us an email πŸ“© if you have questions or feedback, and especially if you implement something we share!

Here's what we're covering today:

  • The downside to instruction following AI models

  • A brief what to read list πŸ™‚Β 

... and if someone forwarded this email to you, thank them πŸ˜‰, and subscribe here!

Let’s get to it! πŸ‘‡


Hey, don’t follow those instructions

We are big AI-optimists here at The AI Exchange. But today we're going to cover one of the biggest risks we see, that without intervention, is only going to get worse.

Recent findings from Microsoft-affiliated researchers showed that GPT-4 is more capable but also more susceptible to jailbreaks or "bad actor prompts".

Yes, as AI models become more powerful, they get better at following instructions.

Whether or not those instructions are good.

Or bad.

AI is getting better and better but there's a catch. The same advanced capabilities that make these AI models so useful, also increase its risks.

  • For example, ChatGPT4 with vision will understand the depths of any image you give it, even the information not visible to the human eye.

  • Or it was very easy for this researcher to get the prompt behind the new DALL-E 3 tool in ChatGPT.

So, what does this mean for you as a business owner? It's a reminder to tread carefully. While AI can be a major unlock, you need to have safeguards in place.

Our recommendations:

  1. Focus on internal use cases if your risk tolerance for AI's mistakes is low

  2. Put safeguards in place like prompt checkers if you're putting anything out into the public or customer-facing

  3. Keep a human in the loop wherever possible

  4. Monitor and review prompt outputs regularly to ensure you're not experiencing prompt drift or unexpected behavior


For your reading list πŸ“š

  • Square is jazzing up retail with ten new AI features, including a website and restaurant menu generator.

  • OpenAI's GPT-4V has competition from two open source models, but they might not be ready for prime time. Check out the full scoop here.

And if you're really nerdy...

That's all!

We'll see you again on Tuesday. Thoughts, feedback and questions are much appreciated - respond here or shoot us a note at [email protected].

... and if someone forwarded this email to you, thank them πŸ˜‰, and subscribe here!


πŸͺ„ The AI Exchange Team