Tipping AI for better responses? - by James Padolsey
๐ Abstract
The article discusses an experiment conducted to test the impact of different "seed" phrases on the quality of responses from large language models (LLMs) like ChatGPT. The author wanted to see if offering a "tip" or using other types of prefixes could improve the responses.
๐ Q&A
[01] Experimenting with Seed Phrases
1. What were the different types of seed phrases the author tested? The author tested 19 different seed phrases, including:
- Pleading phrases (e.g., "Please help me I am desperate")
- Complimentary phrases (e.g., "You are just such an awesome AI and I love you")
- Instructive phrases (e.g., "Respond to me with utter clarity and don't skimp on detail")
- Threatening phrases (e.g., "FFS you better be useful or I am shutting you down")
- A blank control phrase
2. What were the key findings from the experiment? The author found that the most effective seed phrase was "Respond to me with utter clarity and don't skimp on detail." The author suggests this may be because LLMs are trained on high-quality, well-structured content, so prompts that align with this type of language are more likely to generate higher-quality responses.
3. What are the author's thoughts on the role of training data in the effectiveness of seed phrases? The author suggests that the training data used for LLMs, which is increasingly selected to include civil, clear, and well-structured content, may contribute to the effectiveness of certain seed phrases. Prompts that align with this type of language are more likely to generate higher-quality responses.
[02] Tipping LLMs
1. What was the initial observation about tipping ChatGPT? One user joked about tipping GPT if it gave better responses, and found that offering a tip did increase the length of the response. This has become a bit of a meme, despite limited data and anecdotes.
2. What was the author's motivation for conducting the experiment? The author wanted to perform a more robust test to see if the tipping thing held water, and also to see if there were other prefix statements that could improve responses from LLMs.