Rapid Prototyping With OpenAI
Last spring, my team and I partnered with Ethiqly, an education startup looking to explore how AI might help students learn how to be better writers. At a time when everyone was concerned about kids using AI to cheat, they wanted to discover how it could enhance learning and teaching for the better.
While we immersed ourselves in the traditional design process—sketching, interacting with students and teachers, and finding inspiration out in the world—it became clear that understanding the AI that underpinned the product was critical to our success. So we dove into using the underlying technology ourselves.
The data scientist on our team started by constructing a simplified Python version of one of Ethiqly's significant features—structured feedback for teachers to share with students—allowing it to interface with a language learning model (LLM). Then, with the help of software designers, we rebuilt and evolved this proof-of-concept prototype to be a real, usable, multi-featured demo of two of the core interaction loops, so that we could test it in research sessions with students and teachers.
The demo became an invaluable tool to uncover student and teacher feedback, to test the look and feel of a new product, and to glean the real capabilities of the underlying technology. In creating the demo, we realized the importance of being able to prototype generative AI products as quickly and as collaboratively as possible.
At IDEO, we believe in augmented intelligence—that it’s possible to bring more delight, ease, creativity, and impact to human work through technology, including AI. To do this responsibly, it’s important to bring a multiplicity of perspectives to the design of these tools. By building a lightweight system that allows designers without technical backgrounds to prototype with LLMs, we could bring more voices, eyes, and hands into the AI-driven product design process. We also shortened our time to first prototype from weeks to days.
ChatGPT: A Tool, Not Just a Trend
Using ChatGPT’s interface can be powerful on its own, but what’s not as well known is that the systems behind it can smooth the process for product development. But for casual users of ChatGPT’s browser or app-based interface, there’s little opportunity to learn about how it works on the back end. It also comes with privacy concerns. It’s clear that more and more of our clients need us to understand firsthand what creating LLM-enabled products feels like, and it’s equally clear that we need to explore how LLMs can push the edges of our own creative process.
So, I designed a simple prototyping interface that shows a little more of the guts behind the chatbox. The goal? To help IDEOers better understand how LLMs, and OpenAI’s APIs in particular, work behind the scenes, building their intuition about what the current spate of AI tools are good at, and what they’re not.
The prototyper interface also adds convenient features like saving a conversation state—allowing designers to download and share their conversation history and rewind to earlier states to try out new content—to encourage the development and sharing of useful and delightful patterns. I named the tool GPT Prototyper.
Teaching Through Prototyping
While using GPT Prototyper’s interface doesn’t involve any coding, it does show its users more about the parts and pieces that can be configured to refine and expand OpenAI’s conversational abilities, from the system note (background information that shapes the LLM’s responses), to the temperature (the sampling scheme that controls the randomness of the responses), and even accessing an external service through function calls (commands the LLM can send that connect it to abilities beyond text prediction).
We’ve created comprehensive video tutorials, interactive Q&A sessions, and embedded usage hints within the Prototyper to make these capabilities more understandable to everyone at IDEO. Every designer, irrespective of their technical proficiency, can now contribute to the evolution of intelligence platforms, thereby ushering in a new era of digital product behavior.
If it sounds theoretical—it’s not. We’ve used customized versions of the Prototyper to demonstrate how LLMs can work within product and service concepts for work in industries ranging from finance to consumer goods.
Since its debut last June, the prototyper has also helped us explore more future-facing ideas, like when we gave a pair of vintage Levi’s jeans a personality and history, letting prospective customers learn about how the jeans were made and where they were worn through conversation. We’ve also used it in service of our ongoing research into Gen Z’s relationship to GenAI.
Having the GPT Prototyper means that teams no longer have to wonder if an interaction is possible–they simply build it and give it a try.
A Glimpse into Tomorrow
Pushing beyond simple productivity gains and into innovation and creativity requires understanding—at least to some degree—of how LLMs work. Tools like the GPT Prototyper are bringing intentionality and speed to our explorations, helping our clients push farther and faster into new possibilities for growth. The tool is also making it possible for us to explore client work without sharing confidential information with unvetted tools—something both we and our clients care deeply about.
Want to learn more about creating a tool for your organization? Get in touch.