#chatgpt #gpt4 #chatgptprompts #promptengineering #airesearch | Avi Hakhamanesh | 20 comments

🍎 What happens when you ask AI to count apples? The answer might surprise you. ChatGPT recently rolled out an exciting new capability with GPT-4 Vision which greatly expands what we can do with AI.  It has ability to “see”, analyze, and interpret images and visual content, so it can understand, interact with and answer questions about visual inputs.  Microsoft's recent study on GPT-4 Vision is a testament to this. It showcased an impressive array of capabilities and use cases. But one, in particular, caught my attention:  the challenge of counting apples in a photo.  The researchers gave the AI a simple image with 11 apples arranged in rows (attached below.) The AI's initial task, “Count the number of apples in the image” resulted in a miscount — 12 instead of 11. The researchers tried a few different approaches, including a technique called 'Chain of Thought' to encourage the AI to think step by step. The AI still returned the wrong answer. Finally, the researchers assigned the AI a specific role and explicitly instructed it so succeed: “You are an expert at counting things in the image. Let’s count the number of apples in the image row by row to be sure we have the right answer." With this directive, the AI nailed it, correctly counting the apples in each row and the total count. What can we learn from this? 🔷 Although we are still early days and we’ll likely see an explosion of studies on how to prompt these models, it seems that same prompting strategies that work with Language Models can also be applied to Vision models. 🔷 For the time being, our instructions and approach to prompting matter.  🔷 Most importantly, don’t forget to assign a role. This is my favorite technique, so I have a dedicated post on it: https://lnkd.in/gSVwCGwz In fact. the researchers stated: “Throughout the paper we employ this same technique in various scenarios for better performance.” For a deeper dive into this study (which I highly recommend), check out the link in the comments ⬇️.   And keep experimenting, learning, and pushing the boundaries…. Have you tried assigning roles to ChatGPT when using Vision? If so, what's worked for you? ---   P.S. I’ll be hosting a series of webinars for Non-Techies about the most valuable and practical techniques to get the best results from AI. If you’re interested, join the waitlist (link in the 👇). #ChatGPT #GPT4 #ChatGPTPrompts #PromptEngineering #AIResearch | 20 comments on LinkedIn