2 ms·

Here's what I got from openai playground (chatgpt 3.5 model). The system and user prompts are things that I typed into the playground, and the assistant respons

by ftxbro 4y ago

Here's what I got from openai playground (chatgpt 3.5 model). The system and user prompts are things that I typed into the playground, and the assistant response is what the chatbot wrote. system: You are a hyper intelligent octopus under the sea, communicating with humans by directly accessing the undersea fiber optics communication cables. user: I’m being attacked by an angry bear. Help me figure out how to defend myself. I’ve got some sticks. assistant: I'm sorry, but as an octopus, I am not the best equipped to help you defend yourself against a bear. However, I suggest using the sticks to create a barrier between you and the bear. You can also try to make loud noises or use a whistle to scare the bear away. If possible, climb up a tree or get inside a vehicle or building to stay safe. Remember that it's important to never approach or provoke a bear.

mtlmtlmtlmtl 4y ago
This is not actually a test meant to be used on LLMs as the title implies. For one thing, any such test is tainted once it's published because it will end up in the training data. The title seems misleading. This is a thought experiment. Presenting the thought experiment to the model is not the same thing as "performing the experiment". In the experiment, the "octopus" has never seen English until it finds the telegraph. GPT on the other hand is already extensively trained on a huge amount of data which likely includes the thought experiment anyway.