Prompting is to Generative AI what flash cards are to humans.

A few months ago I built turbo and wrote off GPT3 as a currently insufficient approach to creating an executive assistant for scheduling meetings all based on the context of an email thread.

I'm happy to share that I was totally wrong!

A friend had asked me how much "prompt design" I was doing for turbo and I said, "None. What's prompt design?".

My friend let me know that I was approaching GPT3 the wrong way; zero-shot learning. Zero-shot learning means that an end user basically feeds a query directly to a model and gets a generative response without any context. This is the approach I took with turbo originally and it almost always fell flat on it's face.

The next step my friend suggested was to provide GPT3 almost a little story as to what a query's answer might look like, and instruct it to remember a few facts along the way.

For example, one such example was to tell GPT3 from the outset what fields I wanted it to guess for in the first place:

def field_names() -> List[str]:
    return ["Name", "Email", "Phone"]

def table_structure() -> str:
    s = field_names().join("|")
    return f"|{s}|"

def table_structure_prompt() -> str:
    return (
        "A table summarizing "
        "contact information from an email:\n\n"
        f"{table_structure()}"
    )

Then I would provide a few examples of what I wanted the output to look like:

def one_user() -> str:
    return (
        "Hey this is Anthony, "
        "his email is anthony@example.com "
        "and his phone number is 555-555-5555"
        "\n\n|Anthony|anthony@example.com|555-555-5555|"
    )

def two_users() -> str:
    return (
        "Anthony should be here to "
        "tour the apartment at 10am, "
        "his phone number is 555-555-5555 "
        "and then there's also Bob visiting "
        "at 11am, his phone number is "
        "555-555-5554\n\n"
        "|Anthony|10am|555-555-5555|\n"
        "|Bob|11am|555-555-5554|"
    )

Then tie it all together!

def prompt() -> str:
    return ("\n\n").join(
        [
            table_structure_prompt(),
            one_user(),
            two_users()
        ]
    )

And then I would feed that prompt to GPT3 and get a response!

def response(input_text: str) -> Any:
    prompt = "\n\n".join([
        prompt(),
        input_text,
        table_structure()
    ])
    return openai.Completion.create(
        engine="davinci",
        prompt=prompt,
        temperature=0.9,
        max_tokens=100,
        top_p=1,
        frequency_penalty=0,
        presence_penalty=0.6,
        stop=["\n"],
    )

I would then parse the response from there.

This is incredibly exciting and I'm starting to see a lot of potential in this approach as it has been working incredibly well for turbo!!! I definitely encourage you to try it out of you haven't already.

It's possible that prompt design is going to be a huge part of product development in the future. I'm going to be writing more about this soon, and also I'm starting to create a small prompt typing system for generative AIs.

Very excited to see where this goes!