What separates AI from other tech waves is the broadened accessibility it creates out of the gate. Think of other waves, like cloud or mobile, which required specialized knowledge to take advantage of. Thanks to generative AI, swaths of people around the world are creating in ways they never have before. This is what true democratization looks like.
Ultimately, there’s no better way to appreciate what AI offers than to build something with it, so we wanted to put our money where our mouth was.
We built a game called Real Fake, in which you have to guess which companies are real and which are AI-generated. It’s best experienced on a mobile browser, as you swipe between six cards generated fresh daily. The real company data is provided by Crunchbase (a Felicis portfolio company), one of the world's largest sources of company data.
Below, we will share more about how we built this game, what we learned, and why it makes us even more excited about AI’s future.
How we built it
After we came up with the initial concept, we partnered with Chance Agency (makers of the indie game Neo Cab) to build the game. The first part was choosing our stack for building.
Crunchbase
We used the Crunchbase API to obtain best-in-class private company data to incorporate into our game. We plugged this data into our AI model so that it could create fake companies that resembled real ones in terms of their funding, product offering, and more.
OpenAI
We used OpenAI GPT-4o to generate the fake companies, and also the logo for the game.
Supabase
Our entire backend is hosted on Supabase (another Felicis portfolio company), which allows us to easily set up a real-time database and schedule the edge functions that generate each day’s game. Some of the most cutting-edge AI applications use Supabase, so we wanted to use the infrastructure that many founders are using.
Vercel
The web frontend is hosted on Vercel, which pairs great with Supabase.
Learnings
The building process was pretty straightforward; we had phases of 1) Design, 2) Build, 3) Test, and 4) Refine. Here’s what we learned.
Data quality is the bedrock of any AI experience - There are millions of companies within Crunchbase, along with their funding data, valuation, location, and more. Gathering this data from the Crunchbase API, we filtered it based on the kinds of companies we thought players of our game would be most familiar with.
Prompt engineering requires some clever workarounds - When we used minimal filtering on the companies early on with simple prompting, the game was too easy. Once we added filtering, though, the game became too hard. OpenAI was generating companies that sounded too real (and sometimes some that were real). Once we instructed ChatGPT to behave more like an ambitious MBA student (with a bit more direction we’re keeping private to keep the game fun), the generated companies started feeling right. Sometimes, the AI will even devise a good idea that gets us talking!
Our serverless architecture is lean and mean - At first, we wrote a function to generate both the questions and the game simultaneously, but that took too long, so we broke it into two functions—one for the game and one for the company cards. The first function generates the game and the category. The second function calls out to Crunchbase data and ChatGPT to bring in the real and fake companies. There’s no process running all the time. Instead, a cronjob in Supabase triggers a Deno edge function in the wee hours each morning, picking the day’s category, querying for real companies, and dreaming up fakes to fit. The beauty of this is that it’s allowed the game to be pretty hands-off, so we get to “set it and forget it” and enjoy the game we built every day!
Playing with AI is all about context and the interaction layer - When we initially thought of this game, we thought about it as a series of multiple-choice questions, but that proved to be the wrong mode of gameplay for many reasons. Once we switched to the swiping format, the pace of the game felt perfect and matched the time expectation someone would have for something like Wordle or Connections. The format is key. If the game were just an interaction between a person and a chatbot, that would be unexciting. Adding extra layers of context (like company valuation, status, and country of origin) provided enough clues to make players feel more sleuthy in their deductions.
Entertainment comes before transformation
We’re still largely in the “entertainment” section of AI’s early history, but Real Fake has made us think about a world where games are as easy to make as videos, and the cohort of game creators is 100X bigger.
While copilots were helpful in building this game, we still can’t help envisioning a future where you can prompt an AI to create a bespoke game and have it update every day. Some of the signals that we’re getting closer to that reality will be as AI takes on increasingly more significant parts of the build process like data cleanup, writing functions, requesting and setting up the right APIs, fixing permissions, etc. Yet it wouldn’t surprise us to see the reality of “write prompt, get game” come sooner than we think (from companies like Astrocade).
This entire process made us realize how gameplay might evolve thanks to AI—from how rules are defined and how NPCs interact with people (and each other) to how entire levels and quests are created. Of course, some amazing companies are already building in these areas, like Altera. Clementine, and others. These founders are building next-generation infrastructure for games that makes it easier to model and scale entire worlds.
Building Real Fake has given us a distinct appreciation for how to construct novel experiences with AI. Reach out to us at game@felicis.com to tell us what you think about the game or if you’re building something crucial for this dynamic era of gaming.
