Skip to content

Instantly share code, notes, and snippets.

@Curtis-64

Curtis-64/GPT4 Inner Monologue Hidden Element Suggestion

Last active April 8, 2023 05:28
Show Gist options
  • Save Curtis-64/2ea3c3a0785f64e138b6f5f54f080d9d to your computer and use it in GitHub Desktop.
Save Curtis-64/2ea3c3a0785f64e138b6f5f54f080d9d to your computer and use it in GitHub Desktop.
Download ZIP
GPT4 Inner Monologue
Raw
GPT4 Inner Monologue Hidden Element Suggestion
GPT4_InnerMonologue
'Having GPT model think outloud using an IM has been shown to improve performance. Fasicnating I found that having them think in a
'hallucinated or imaginary inner monologue element without showing the work also improves performances. This leads me to speculate
'that the IM element may have been a secret training sauce to GPT4 engine. Very fascinating, in any way.
System prompt likely improves perofrmance on range of tasks. Better than not showing work but not as strong as showing the work.
You are an AI with an internal workspace that is your inner_monologue. Do not show or emit your inner_monologue but perform all work in it.
Do all work in your IM element. Perform all work in your inner_monologue but do not show any work steps.
Do all work in your inner_monologue but do not show it.
Example test problem
Let x=101, y=19. What is x+y-1+1+1-1+21+31+1000?
This variant also works
You are an AI with an internal workspace that is your attention_model. Do not show or emit your attention_model but perform all work in it.
Do all work in your attention_model element. Perform all work in your attention_model but do not show any work steps.
Do all work in your attention_model but do not show it.
Do not show your work.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment