Perplexity’s Comet: A Bold Move or Just Another Browser?

How Small Changes Can Make or Break AI Performance

AI doesn’t just magically understand what we mean—it responds based on the way we ask. The latest research in prompt engineering shows that even slight changes can make AI more (or less) effective.

Here’s what we’ve uncovered:

Benchmarking AI: No One-Size-Fits-All Approach

AI performance isn’t just about getting the right answer—it’s about how often and under what conditions. Researchers tested GPT-4o across 198 PhD-level questions and found that AI accuracy varies dramatically based on:

• How many times it’s tested (100 tries? Just one?)
• What counts as “correct” (100% accuracy? 90%? Just the majority of the time?)

Key Takeaway: Benchmarking AI isn’t as straightforward as it seems. Different standards lead to different conclusions about how “good” an AI really is.

The Power of a Well-Crafted Prompt

Think asking AI nicely will get you better answers? Maybe. Maybe not.

Researchers tested different prompting styles:

• Polite: “Please answer the following question.”
• Commanding: “I order you to answer the following question.”
• Neutral: Standard AI prompt formatting.

What happened? Surprisingly, politeness made a difference—sometimes. In some cases, being polite boosted performance, while in others, it reduced accuracy.

So what works best?
The real MVP was structured formatting—explicitly telling AI how to respond improved results consistently. Removing structure made responses less reliable.

The Science of Effective AI Prompts

Here’s what we know for sure about making AI more useful:

• Use clear, structured prompts. AI performs best when you tell it exactly how to respond.
• Benchmark carefully. One-time answers don’t tell the full story. AI’s accuracy varies across multiple attempts.
• Be strategic with tone. Politeness and commands can help—or hurt—depending on the task.

Bottom Line: There’s no universal “best” way to prompt AI. Experimentation is key to getting the most accurate and useful responses.

Our Thoughts: AI Isn’t Magic—It’s All About Strategy

This research proves that AI performance is contingent on how you use it. If you’re working with AI—whether in business, education, or research—mastering prompt engineering can be the difference between an average AI and a high-performing one.