Playing improv
...with Python?

My first improv?

Reading about Improv

Impro: Improvisation and the Theatre
By Keith Johnstone

My GoodReads rating from 2015:

So Many Improv Classes

Endgames Improv: 2965 Mission Street

Leela SF: 901 Mission Street

Studio A.C.T: 1127 Market Street

Berkeley Rep: 2071 Addison St, Berkeley

Running an improv club

Bringing improv to work

Our afternoon "shake-off" at Woebot:

Laser cutting for improv

Documenting improv

Because apparently I document everything...

Flavors of improv games

• Verbal
  • Storytelling
  • Word games
  • Generative games
• Physical
  • Space objects
  • Working with props
• Character building
• Scene building
• Status games
• Musical improv

Let's play: Story Spine

• Once there was…
• And every day…
• Until one day…
• And because of that…
• And because of that…
• And because of that…
• Until finally…
• And so…
• The moral of the story is:

Full rules: Group-Story-Spine.md

Can we play improv with a computer?

We need the computer to generate a creative human-like response.

Possible tools:

• Hidden markov models
• Custom trained ML models
• Large/Small Language Models

Language Models

Let's try these models...

Ollama

Ollama is a tool for easily running local LLMs on your computer.

You can also run it from GitHub Codespaces: 🔗 github.com/pamelafox/ollama-python-playground

We can call Ollama programmatically using:
HTTP, ollama package, or openai package.

GitHub Models

GitHub models are freely available models with very low rate limits. 😢 Currently waitlist only, sign up at: github.com/marketplace/models

We can call GitHub models programmatically using:
HTTP, azure-ai-inference package, or openai package.

Using the OpenAI SDK

Since the openai package can be used for both Ollama and GitHub models, we'll use it for most examples.

                response = client.chat.completions.create(
                  model=model_name,
                  messages=[
                    {"role":"system",
                     "content":"You're playing improv games."},
                    {"role":"user",
                     "content":"Suggest a random first line."}],
                )
                

Full code: example.py

Game: Story Spine

Let's start easy!

Rules: Game: Group Story Spine

Code: storyspine.py

Game: Sentence-at-a-time Story

AKA: Will the language model ever end the story?

Rules: Sentence-at-a-time-Story.md

Code: sentenceatatime.py

Game: Word-at-a-time Story

AKA: Why can't language models count?

Rules: Tweet-Generator.md

Code: tweetgenerator.py

Game: Yes, And Product Factory

AKA: Can we use props??

Rules: Yes,And-Product-Factory.md

Code: yesandproduct.py

Yes, with multimodal models!

Three Things!

AKA: Can we make lists??

Rules: 3-Things!.md

Code: threethings_simple.py

Lessons learned

• Larger models are better at following the rules
• Doing *anything* with language models teaches you a lot about them
• Improv with computers is fun!
• Improv with humans is more fun!