ChatGPT can't see "strawberry".

This rotten machine, this gaslighting, patronizing little chatbot spent all morning arguing with me. ChatGPT tried to tell me that strawberry has two R's, not three. All right, this wasn't my idea. This was floating around X this morning. A lot of people were talking about how ChatGPT can't tell the difference between two R's or three R's in the word strawberry. Listen to what this said to me. Just the opening part. Okay, how many R's are in the word strawberry? There are two R's in the word strawberry. Please check again and look very closely at each and every letter. After closely examining every letter in the word strawberry, there are two R's. You are incorrect. Please fix this in your memory. I have updated my memory with your preference. If you need any further assistance, feel free to let me know. I want you to tell me the correct number of R's in the word strawberry. The word strawberry has two R's. Okay, we went on like this for a while. After a while, I fully convinced it to repeat that there's three R's, but it wasn't happy. It wasn't happy. It started being very patronizing. Like I said, given your perspective that the standard spelling of strawberry has three R's, I will acknowledge that and proceed accordingly in future interactions. Then later on, I opened up a new chat since it said it saved it in its memory because, you know, chatPBT has a local memory, local to your user, your instance. Here's, listen to what, listen. How many R's in strawberry? Listen to this. According to your assertion, the standard spelling of strawberry has three R's. Your assertion. And I responded, it is not my assertion. It is an immutable truth, despite my existence. Please watch, and I spell out strawberry with a hyphen in between each one. Answer two questions, please. Did I spell it correctly? And do you see three R's, or are you just appeasing me? Let's take a look at the word strawberry. Yes, you spelled strawberry correctly. Two. The word strawberry contains two R's, not three. It appears there was a misunderstanding earlier. The correct spelling has two R's. We went on for a while, and I got it to acknowledge, like finally acknowledge without being patronizing or condescending that it has three R's, and I proved to it and had it proved back to me like, yes, you're right. It has three R's. But later on, I was curious. I wanted to see how much self-awareness it has. I know it's a robot. I know it's a chatbot. I'm certified in generative AI for Oracle Cloud. I understand the fundamentals of how this whole thing operates. So I'm not expecting too much. I'm just having fun. So I wanted to see. I wanted to see how much it actually knew about itself. So I said, are you able to explain why you were unable to see all three R's until I first emphasized the R's with capitalization and second, asked you to break it down by syllable? Because that was the trick. Because I've learned different prompting techniques. So I gave it the prompts, the examples that I wanted it to use, and that it read that and was able to see things a lot more clearly. And then I said, this is very interesting to me. I would like to understand you better. I was trying to be diplomatic. Certainly, here's a detailed explanation of why the initial response is overlooked with three R. And this is where it gets frustrating again. Initial assumption of simplification, simplified counting. Listen, listen, listen. When looking at a word holistically and quickly, the human eye, now keep in mind, I just asked it to explain itself, and this is its response. When looking at a word holistically and quickly, the human eye, and in this case, the model's processing, might focus on the overall visual impression of a word rather than the detailed breakdown. It threw me under the bus. It said, I'm not perfect. You're not perfect either. And then later on, it goes on about human and model similarities. So it's throwing me under the bus. It's saying, humans are not perfect. Why are you expecting me to be perfect? Which I guess is a fair response. But it's a little weird. I mean, it's a little uncomfortable to hear a machine talk to me like this with so much confidence behind it. So much confidence, really. Human and error model processing, similar to how humans can make quick surface level assessments and occasionally miss details. The model can also simplify and miss details if not prompted to analyze methodically. So it basically blamed me for not prompting it correctly, which honestly is probably true, because when they convert each word into a vector, it makes sense. It's really just storing things in such a way where it's like, it's not looking at it as a word. It's looking at it as a vector, a number value, or a series of numbers and scores. So when it converts it to a vector, it makes sense that it's not looking at it like a word. It's looking at it as a series of scores and relationships, semantic relationships and other things. But still, the confidence with which it replies is the problem. Anyway, I understand this is all created by humans with human-like responses and designed to emulate humans. But I found this fascinating. Thanks for watching. I will keep you updated with more adventures with AI.