BLOOM Story Writer
An interactive storytelling experience powered by BLOOM, the last of the largest pure text continuation models. Unlike ChatGPT, Claude, or Google Gemini, BLOOM is pure textual analysis—no guardrails, no weighted responses. It's fun because you never know what you're going to get!
Visit nlp.henzi.org to start a new story or browse existing ones
How It Works
Choose Your Own Fate uses the BLOOM transformer model to generate interactive stories with unique flavor and creativity. Each story begins with a random prompt from a curated collection or from today's top news headlines. Three continuation options are generated at each step, each with its own randomized parameters (temperature, top_k, typical_p, etc.), so every option is unique and different each time you press "Generate." Select the path that resonates with you and watch your story evolve—your progress is saved automatically so you can browse, continue, or share it later.
Stories generated from news headlines are particularly interesting, as they have a chance of containing named entities (proper nouns, places, people), and it's fun to see the model generate text in the context of the latest events it actually has no knowledge of. If one of the three options isn't suitable, simply generate more until you find the right one to fit your story.
Ahead of Its Time: A Different Paradigm of Human-AI Collaboration
When this project was conceived in 2022, most of the world treated language models as oracle textboxes: type a prompt, get a single final answer. This app took the same technology and built a co-creative instrument instead—and the industry is only now catching up to that thesis.
Human-in-the-Loop Beam Search
Inside a language model, "beam search" evaluates multiple possible token paths and silently picks the most statistically probable one. The Story Writer's Creativity Trail turns that hidden process into a user interface: instead of letting the model blindly choose its own path behind a black box, the app surfaces distinct branching realities and hands the human the steering wheel. The real value of an LLM isn't its ability to finish a thought—it's the ability to offer divergent forks in the road that trigger human inspiration.
A Spectrum of Predictability, Not a Single Slider
In 2022, tuning an LLM usually meant sliding one global "temperature" bar in a playground. Here, generation parameters (temperature, top_k, typical_p) vary progressively across the three options themselves—one slot tightly constrained and probable, another wild and improbable—mapping a literal spectrum of predictability into a scannable, human-readable layout, complete with a "Creativity Match" score and an aggregate creativity trajectory tracked over the life of the story.
Preserving Creative Variance
Modern instruction-tuned models have been heavily aligned to be polite, safe, and homogenous—and in the process much of their creative variance was flattened away. By capturing raw text continuations across varied sampling weights, this interface preserves exactly what creators are starving for today: genuine unpredictability. Chat bubbles are a poor way to write a book; branching, multi-alternative generation is how creative platforms are now trying to work—and it was running here in production years earlier.
About BLOOM: The Last Pure Text Continuation Model
Built by the BigScience team (at Hugging Face), BLOOM was once the world's largest language model. Today, it stands as the last of the largest pure text continuation models—the same size as GPT-3, and the last vestige of models before they became instruction and chat-focused.
Modern models like ChatGPT, Claude, and Gemini are trained with guardrails and weighted responses, optimized to be predictable and "safe." BLOOM is different: it's simply a large language model doing pure textual analysis, without the constraints of instruction tuning or safety filters. That means more varied, unfiltered, unpredictable output—true text continuation rather than chat completion. It's fun to try because you never know what you're going to get! 🎲
nlp.henzi.org also holds a unique distinction: it's the world's first "vibe coded" app, originally written in 2022—well before the term entered the popular lexicon. The creator built a functional React app in 24 hours with no formal JavaScript or React background, using ChatGPT to write the code and BLOOM (via Hugging Face Transformers and a FastAPI backend) to generate story continuations, with video evidence on YouTube documenting the process.
A Project by The Henzi Foundation
This project is a technology initiative offered by The Henzi Foundation, a nonprofit committed to covering child funerals in Cincinnati and surrounding areas, providing compassionate support and financial assistance to families during their most difficult moments.
This BLOOM Story Writer project represents one of the foundation's technology initiatives, exploring the creative potential of open-source AI models while contributing to the broader community of developers and storytellers.