ChatGPT Advanced Properties

The “Advanced Settings” for ChatGPT are just numbers we can tweak to try and get preferred responses back from OpenAI.

These inputs come with subtext to describe them already, so I’ll just regurgitate them here.

• Frequency Penalty: This controls the likelihood that the model repeats a previously generated line.
  • Positive values decrease the likelihood of lines being repeated. The higher the more random sentences become.
  • [-2.0, 2.0] is the range, however, it is recommended to be between [0.1, 1]
• Presence Penalty: This controls the likelihood to talk about new topics.
  • Positive values increase the likelihood of the topic changing. The higher the more sporadic topic changing becomes.
  • [-2.0, 2.0] is the range, no recommended values as far as I know. I use 1.
• Temperature: This controls of “focused” the model becomes when generating responses.
  • Higher values like 0.8 will make the output more random.
  • Lower values like 0.2 will make the output more focused and deterministic.
• Max Tokens: This is sorta like a “character limit” and controls how large the generated response is.
  • “On average” 1 token = 4 characters. This isn’t set in stone anywhere.
  • If the max is too low, it can sometimes generate a response that gets cut off.
  • I don’t recommend going over 100 personally