Hours and Cost for gpt-4-0314

[HRHLALALA]

Hi, I am trying to run ChatGPT 4.0 for zero-shot trajectory prediction. However, I found that using gpt-4-0314 requires unreasonable running hours and costs. May I ask what your consumption looks like on your end to run this model on ETH/UCY, especially on UNIV dataset? Do you have any suggestion to solve this? Thanks.

[InhwanBae]

Hi @HRHLALALA,

Thank you for your interest in our work. In my memory, the cost was around $5k to $10k. The best way to reduce costs is (1) by using the new OpenAI batch API, which can cut the price in half and deliver responses much faster than what is stated on the website. Other methods to reduce costs include: (2) reducing precision from two decimal places to one, (3) using a cheaper model like GPT-4o, or (4) modifying the prompt to generate 20 outputs at once instead of predicting 5 paths recurrently.