Remember to tip your ChatGPT

A GPT enthusiast tested if financial incentives, among others, make ChatGPT work better. The answer is that LLMs are weird.

Let me explain.

@MaxWoolf did an exciting experiment using financial gain, rewards, and punishments to coerce ChatGPT into providing better-quality output.

The setup is simple: generate a story that's 200 characters long.

With basic instructions, it could have performed better.

However, when telling the model, it will receive $500/$1000/$10000 if the constraints are followed correctly (200 char. limit).

The results, although not perfect, were more aligned with the criteria. After these tests, he tried other reward and punishment systems, which yielded interesting results.

Most notably, the best results for rewards came from promises of world peace and Taylor Swift tickets. Yeah, LLMs are weird.

Prompting with punishments had similar results, with some providing better results than others.

Interestingly enough, when he started combining multiple punishments and rewards, the two worst-performing punishments individually gave way to the best output.

Personally, when generating images, providing positive encouragement when it creates an image you want to build on, especially involving text, got me further than basic prompts.

Nonetheless, there's something about coercing LLMs for answers, and I can't wait to see further developments.

I highly suggest you check out the full blog post -> https://minimaxir.com/2024/02/chatgpt-tips-analysis/