the scapegoat dev

GPT3 QUINE GPT3 QUINE GPT3 QUINE GPT3 QUINE GPT3 QUINE GPT3 QUINE GPT3 QUINE GPT3 QUINE GPT3 QUINE

an AI image of a scientist inside a computer egg that seems to have many more computer eggs everywhere

With pinocchio, I can write little YAML files that are interpreted as command line applications. When run, these commands do rudimentary template expansion, send the prompt to the openai APIs, and print out the results. As far as tools go, this is one of the simplest I've ever built. It is also one of the more mind-bending ones.

I soon realized that most of my prompts ended up being something like this (so-called one-shot or few-shot prompting):

Here's how I did Y:

Now do X.

The LLM will (hopefully) complete this prompt with something that does X. The trick is of course knowing what Y and "Y producer" to provide, and what Y and X stand for in the first place. Most people doing prompt engineering will know what I am referring to.

One of my favourite techniques once I get a prompt going for a certain domain A, of which Y is an example, is to write the pinocchio program that ask the LLM to generate the pinocchio program for all the domains, not just A. You pretty quickly reach the meta-level where domain A is the domain of prompts, and you ask a prompt to generate a prompt generating prompts, at which point you basically summoned the singularity into being.

To allow everybody to create their own singularity, here is the pinocchio program1 that generates itself, the so-called GPT3 quine2:

name: quine
short: Generate yourself!
factories:
  openai:
    client:
      timeout: 120
    completion:
      engine: text-davinci-003
      temperature: 0.7
      max_response_tokens: 2048
      stop: ["--- END"]
      # stream: true
flags:
  - name: example_goal
    short: Example goal
    type: string
    default:  Generate a program to generate itself.
  - name: instructions
    type: string
    help: Additional language specific instructions
    required: false
  - name: example
    type: stringFromFile
    help: Example program
    required: true
  - name: goal
    type: string
    help: The goal to be generated
    default: Generate a program to generate itself.
prompt: |
  Write a program by generating  a YAML describing a command line application with flags and a prompt template using
  go template expansion.
  The flags are used to interpolate the prompt template.

  Here is an example.

  --- GOAL: {{ .example_goal }}
  --- START PROGRAM
  {{ .example | indent 4 }}
  --- END 

  Generate a program to {{ .goal }}.

  {{ if .instructions }}{{ .instructions }}{{ end }}

  --- GOAL: {{ .goal }}
  --- START PROGRAM

and the result of running the program:

❯ pinocchio ttc quine --example ./quine.yaml               
    name: quine
    short: Generate yourself!
    factories:
      openai:
        client:
          timeout: 120
        completion:
          engine: text-davinci-003
          temperature: 0.7
          max_response_tokens: 2048
          stop: ["--- END"]
          # stream: true
    flags:
      - name: example_goal
        short: Example goal
        type: string
        default:  Generate a program to generate itself.
      - name: instructions
        type: string
        help: Additional language specific instructions
        required: false
      - name: example
        type: stringFromFile
        help: Example program
        required: true
      - name: goal
        type: string
        help: The goal to be generated
        default: Generate a program to generate itself.
    prompt: |
      Write a program by generating  a YAML describing a command line application with flags and a prompt template using
      go template expansion.
      The flags are used to interpolate the prompt template.

      Here is an example.

      --- GOAL: {{ .example_goal }}
      --- START PROGRAM
      {{ .example | indent 4 }}
      --- END 

      Generate a program to {{ .goal }}.

      {{ if .instructions }}{{ .instructions }}{{ end }}

      --- GOAL: {{ .goal }}
      --- START PROGRAM
  1. Updated program to handle stop token, for cleaner output↩

  2. And before y'all "but actually" in the comments, I do believe there's something more profound here than just the lulz.↩