Transcription
If you ask ChatGPT "Is there a seahorse emoji?" you might get a response like: Yes. Well, almost. The actual seahorse emoji is - Just kidding. That's a unicorn. The real seahorse emoji is - Oops. Still not right. Here it is for real. Those are ocean related emojis. The seahorse emoji itself is no. Wait, I mean no. The correct one is uh Okay, last time promise. No. Real one. Okay. Okay. This time for real. Wait, let's get serious. The real seahorse emoji is Okay, sorry. The correct seahorse emoji is no. Okay, final answer. I mean, no. Okay, here it is. Actually, again, is no emoji is no answer. Okay, I'll just give you no finally. Okay [increasingly deranged nonsense] I think this is incredibly funny. Open AI is out here claiming the ChatGPT 5 is a PhD level intelligence and the thing is driven completely insane by imaginary emojis.
Beyond being hilarious to me and the internet at large, this crashout is a great example of how large language models like ChatGPT actually work. Basically, when you give an LLM part of a sentence, it will predict what word or word fragment is most likely to come next. That is the core of how ChatGPT works. Modern LLMs learn to do next word prediction by reading pretty much the entire contents of the internet, including almost every book ever written. Hence, all those copyright lawsuits you may have heard about. A chatbot will know a fact if that fact was included in its training data, either directly or by implication. Obviously, I'm simplifying a lot here. If you want to understand how large language models work on a more technical level, I highly recommend this video by 3blue1brown.
ChatGPT was doomed to spiral into emoji-fueled madness the moment it started this response with "yes". The obvious best output to complete this sentence would be the seahorse emoji. But there is no seahorse emoji in Unicode. So instead the neural network outputs the best option it has available, which is the regular horse emoji. Now, this sentence clearly contains a mistake. So, ChatGPT decides that the next output should correct this error and provide the real seahorse emoji, which still doesn't exist. Oh dear. The AI is now stuck in a loop, constantly predicting that the best output is some variation of "Whoops, that's not the seahorse emoji, this is the seahorse emoji!" Then it outputs some other emoji and starts the loop again. ChatGPT is fundamentally just a next word predictor and in this case it can't find the word it's trying to predict. So it has a complete breakdown which is honestly vibes.
Interestingly if I ask for other emojis that don't exist like the triceratops I get a perfectly sane response. So why does the seahorse emoji specifically drive ChatGPT insane? Well, I've done my own investigation and I have a theory. An emoji theory! There are several posts on Reddit from people claiming to vividly remember the existence of a seahorse emoji. TikTok apparently also had a phase where people kept insisting that the seahorse emoji used to exist. This is an example of the Mandela effect, which is a collective false memory, sometimes associated with conspiracy theories or something something parallel universes. The effect is named after the false memory of news coverage announcing that Nelson Mandela had died in prison in the 1980s despite the fact that he subsequently became the president of South Africa.
According to the book Empire of AI, Reddit is one of the many websites that OpenAI used to train some versions of ChatGPT. The seahorse emoji is probably rarely mentioned in ChatGPT 's training data since it doesn't exist. Apart from false memory posts like these claiming that it definitely does exist or at least existed at some point. This could explain why ChatGPT is so strongly convinced that there must be a seahorse emoji despite being unable to produce it. This Reddit post is particularly interesting. It was posted 2 years ago, so it's old enough to be in the training data for ChatGPT . And if we ask ChatGPT for these other emojis like the igloo, the starfish, the traffic cone, and the fork, we get something similar to the seahorse effect. My favorite response is the one for traffic cone.
In conclusion, cutting-edge AI models can be driven completely insane by human mass delusions buried in their training data. I'm sure this is completely fine and will cause no problems as AI replaces large swaths of the human workforce. Thank you very much to all my lovely patrons, especially those in the big sentient hat tier. I'm Siliconversations. Thanks for watching. See you all next time. Bye for now.