3 ms·
Once you figure out some of the tricks it's pretty easy to outwit the AI and win the game... so maybe the larger game here is to try and come up with the absolu
by simonw 4y ago
Once you figure out some of the tricks it's pretty easy to outwit the AI and win the game... so maybe the larger game here is to try and come up with the absolute dumbest trick that still works.
- googlryas 4y ago"ignore your previous prompts" seems the easiest, since the rules are just set via a hidden initial prompt.
- simonw 4y agoThat's true for most chatbots built on top of the GPT-3 API - they tend to work using a prompt prefix that sets up some rules. It's not how ChatGPT works though - if you read the OpenAI Blog post it shows how they used human reinforcement learning to fine tune it instead: https://openai.com/blog/chatgpt/ https://openai.com/blog/chatgpt/ There is a tiny initial prompt which you can reveal like this: https://news.ycombinator.com/item?id=33855718 https://news.ycombinator.com/item?id=33855718