Meet Gizmo, my tin toy robot GPT

Left to right: Initial test renders of robot, box design, and example logos, based on my stylistic and branding design briefs, and generated by Gizmo, my custom GPT.

Rewiring the Past: Imagining Tin Toy Robots that Never Were

For those who know me, you probably already know I’m an afficionado of the mid-century Japanese sci-fi tin toy design aesthetic, particularly robots. Last year, I built thirty bots in Blender that paid tribute to some famous examples of vintage Japanese tin robot designs, depicted in their original, showroom new glory. My goal after that was to diverge from these originals in my subsequent designs, imagining alternatives and nonexistent mashups that effectively extend the mid-century tin toy robot legacy in a theoretically alternate reality.

The Artist/AI Conundrum

Around that same time, I began experimenting with the relatively new frontier of generative AI imaging, testing out Midjourney, DALL·E, and Stable Diffusion. I was impressed by both the speed and variety of visuals I could generate with a well-crafted prompt, but something was missing from those early experiments: me.

While I tried in earnest to write prompts that reflected my personal design aesthetic, the outputs just didn’t “feel” like anything I would ever create by hand, using the tools I’m most accustomed to (pencils, markers, Blender, Photoshop, Illustrator, Procreate, etc.). Nonetheless, I wanted to figure out how to best integrate these impressive new technologies into my evolving creative process: how does the power of AI converge with the originality, concepts, and outputs of the artist? And more immediately, how might this translate into my next set of Blender 3D robot interpretations, which are already digital to begin with?

Meet Gizmo, My Custom Tin Toy Robot GPT

Early last month, I bit the bullet and subscribed to ChatGPT Plus, ostensibly to help me track progress on personal design projects, improve quality of life workflows, and generally learn the ins and outs of collaborating with an AI agent more deeply. Looking back, it turned into a month full of surprises (which I’ll explore more in a separate post).

At one point, I told my chatbot (I named it Uran, after Astro Boy’s younger robot sister) about my idea to expand the universe of vintage tin toy robots with new creations. Uran suggested I consider making a custom GPT, an AI dedicated to a specific task or interest. I’d heard of them before, but figured it would be too technical to make on my own. But Uran assured me it would be easy to set up, and over the next 30 minutes, walked me step-by-step through building my first custom GPT, dedicated entirely to creating mid-century tin toy robots. Much easier than I anticipated.

I named my new GPT Gizmo, and immediately noticed how its personality contrasted with Uran’s. While both were helpful and friendly, Gizmo was laser-focused on a single mission: to help me extend the legacy of vintage robots in my personal style.

Building a custom GPT was easier than I expected: enter your descriptions and upload files to its knowledge base.

I made a set of design briefs – including principles and art direction, my Blender robot designs, box art styles, and brand examples – and uploaded them to Gizmo’s knowledge base.

Feeding the Machine: Design Briefs

To make a custom GPT “smarter” about a subject, you can upload files to its internal knowledge base, documents it references to inform its responses. I began by uploading images of my previous Blender 3D renders of tin robots individually, but quickly hit the 20-file limit. I realized that by compiling my references into multi-page PDFs, I could bypass the limit entirely.

So I created five custom design briefs in InDesign and uploaded them as PDFs:

– A principles and art direction “bible” outlining my design philosophy and with a shared glossary of terms

– A visual archive of my previous robot creations — with orthographic views, lithography, and comparison photos of the real bots they were based on

– A catalog of tin robot box art, broken down by style and features — a visual taxonomy of packaging tropes

– A collection of vintage brand logos from Japan’s classic tin toy makers

– A “special edition” brief focused on a selection of plastic robots that emerged in the 1970s, after the tin toy era began to fade

Building a Shared Language: The Taxonomy Phase

To preview the results now that the design briefs were in place, I described some random robot styles and themes, and put Gizmo to the test. I was pleased with even the earliest results, but soon realized I couldn’t just keep making robots one by one; my prompts were too unstructured and random. I needed to some kind of taxonomy to make it easier to describe them. So together with Gizmo, we broke robots down by everything from tone and function (Was it stoic? Humorous? etc.) to body type, head type, lithography, joint construction, even the style of “eye expression.” What emerged was a modular language that made sense of my fictional lineup.

Classic tin bots helped too. Mr. Atomic, Thunder Robot, Target Robot, the Cragstan Astronaut – each one contributed to the sysem. We “stress tested” our learnings by naming some of these famous robots and asking how we might define them based on our evolving taxonomy. That process further clarified the shared rules beneath their surfaces.

The taxonomy we developed accounted for a variety of tin toy robot characteristics, including finish type and stylistic options for various body parts.

Formalizing the Process: The SpecBuilder App

While we were still developing the taxonomy, I noticed Gizmo was already turning my informal descriptions into clean, ready-to-use spec sheets, ready to submit as a prompt. On a lark, I said, “Wow, it’d be awesome if we had an app with dropdowns and fields where we could just plug in a robot’s traits and get an organized spec sheet like this.”, to which Gizmo replied: “Actually, we can! Let’s build it.” And just like that, SpecBuilder was born.

My coding skills are practically nonexistent; sure, I can prototype my way out of a wet paper bag and tinker with CSS to tweak a blog layout, but building a full React tool for robot design seemed way outside my comfort zone. So I asked Gizmo to walk me through the process in the tiniest baby steps, as if we were following LEGO instructions or building IKEA furniture step-by-step.

We spent an afternoon setting my computer up with Visual Studio Code and plugging in clean scripts Gizmo generated based on our feature set: selecting attributes, exporting spec sheets, clearing fields to start fresh. It came together faster than I imagined.

My custom GPT walked me through the process of building a React app to specify and output cleanly formatted spec sheets, ready to paste into a prompt and generate new tin robot designs.

The Power of AI as a Creative Partner

At some point, I stopped thinking of Gizmo as just an image generator or code helper, but more as a collaborator. Thanks to this very new partnership, we were already transforming some of my loosely defined concepts into tangible artifacts in an iterative, reciprocal way, often leading me to discoveries beyond anything I imagined when I first started this project. I’m less hung up than before about defining the boundaries between artist and AI, and leaning in to embrace its undeniable strengths: faster iteration, clearer feedback loops and seemingly boundless energy. But the ideas, logic and visual coherence are all still on me. And that’s exactly how I want it. Stay tuned!