What Microsoft Discovered About Asking AI Questions

Small tweaks to your questions that lead to dramatically better answers

Improve Your Conversations with AI

The other day I was reading a Microsoft Research paper and was absolutely blown away 😮. A research team managed to get GPT-4 (you know, the brain behind ChatGPT) to give much better answers. Their secret? Simply changing how they asked their questions.

In their research, "Can Generalist Foundation Models Outcompete Special-Purpose Fine-Tuning? Case Study in Medicine," the Microsoft team showed they could cut errors by 27% just by improving how they asked questions. And the best part? These techniques work for absolutely any topic, from business analysis to philosophy.

Three Ways to Ask Better Questions

1. Dynamic Few-Shot (Example Selection Technique)

This technique involves showing the AI a few examples of questions and answers before asking your own question. But here's the key - these can't be just any examples; they need to be directly related to what you're about to ask.

Let's say you want to ask ChatGPT about digital marketing strategies. Instead of simply asking:

What are the best strategies to gain followers on Instagram?

First, give it 3-5 examples of other questions and answers about social media marketing:

Q: What are the best times to post on LinkedIn?
A: The best times are between 10-11am and 4-5pm on weekdays, when professionals are most active...

Q: What type of content works best on TikTok?
A: Short and entertaining videos, tutorials under 60 seconds, and content that follows current trends...

[And so on with 2-3 more related examples]

NOW your question:

Why does this work better? Because you're giving the AI clear context about how you want it to respond. It's like saying, "Hey, this is exactly how I want you to explain social media strategies to me." The researchers found that simply by adding these related examples, the accuracy of responses improved dramatically.

2. Self-Generated Chain of Thought (Thinking Out Loud)

This second technique gets the AI to walk through its own reasoning process step by step. It's essentially asking the AI to think out loud.

Let's see a practical example. Instead of simply asking:

What would be the best price for my new digital marketing online course?

You prompt it to reason through the problem methodically:

I need to decide the price for my new digital marketing online course.
Please think step by step:
1. What factors should we consider when setting the price?
2. What are typical market prices for similar courses?
3. How does price affect the perception of value?
4. Based on all this, what price do you recommend?

The difference is night and day. Rather than getting a quick, oversimplified answer like "charge $199," you receive a comprehensive analysis that weighs multiple important factors. The researchers discovered this technique significantly improved accuracy because it forces the AI to justify its reasoning process.

3. Choice Shuffle Ensemble (Multiple Perspective Approach)

The third technique involves asking the same question in multiple different ways and then identifying patterns across the responses. Think of it as gathering several expert opinions on the same issue.

Why is this so effective? Because each time you rephrase your question, the AI approaches the problem from a slightly different angle. When you combine these varied perspectives, you end up with a more comprehensive and reliable answer.

For instance, if you want advice on your email marketing strategy, you might ask:

Version 1:
What is the best frequency for sending newsletters to my email list?

Version 2:
How often should I send emails to my subscribers to maintain engagement without overwhelming them?

Version 3:
For a B2B email list, what would be the optimal sending schedule to maximize open rates?

After collecting these responses, look for the common threads between them. If all three responses suggest that 1-2 times per week is ideal for B2B audiences, you can be much more confident that this is solid advice worth following.

When Techniques Combine: Exponential Results

What's truly remarkable is that these techniques don't just work well individually - they amplify each other when used together. When the researchers combined all three approaches, they achieved that incredible 27% reduction in error rates. And this wasn't limited to a single domain - they tested these techniques across wildly different fields including electrical engineering, philosophy, accounting, and psychology, seeing consistent improvements across all of them.

Rethinking AI Interaction

This research completely transforms how we should think about AI interactions. Here's the game-changer: you don't need specialized or fine-tuned models (I'll dive into that in another post) to get expert-level responses - you just need to know how to ask better questions.

It's like discovering your smartphone has had powerful features hidden in plain sight all along. The capabilities were always there at your fingertips; you simply needed to know how to unlock them.

Universal Principles for Any AI Assistant

The real advantage of these techniques is their universal applicability. No matter which AI assistant you prefer - GPT, Claude, or any other - these principles work consistently across all of them:

1. Provide relevant and specific context

2. Ask the AI to explain its reasoning

3. Get multiple perspectives on the same question

Looking to the Future

This research opens up a world of fascinating possibilities. The researchers believe we've only scratched the surface - there are likely many more techniques waiting to be discovered. This is just the beginning of our journey to unlock the full potential of AI assistants.

What excites me most about these findings is that they confirm what I've always believed: the real power of AI doesn't come from having the most cutting-edge tools, but from mastering how to use them effectively.

See you in the next one,
G

---

Hey! I'm Germán, and I write about AI in both English and Spanish. This article was first published in Spanish in my newsletter AprendiendoIA, and I've adapted it for my English-speaking friends at My AI Journey. My mission is simple: helping you understand and leverage AI, regardless of your technical background or preferred language. See you in the next one!