Over the coming weeks we’ll be hosting a series of these chats with team members and friends with backgrounds in games, literary publishing, AI/ML, and more. Be sure to join our Discord for the latest updates!
Could you introduce yourselves and how you joined Hidden Door?
[Matt] 👋🏻 Hi all — I’m Matt, one of the co-founders at Hidden Door, where I focus on our product and community. I'm really excited for this first AMA! My background is building data products around emerging technologies, like search engines, distributed systems and machine learning. Before starting Hidden Door, I spent most of my career working on enterprise software platforms — that is, products that other companies use to build their own products — including roughly 7 years each at Cloudera (in San Francisco) running product marketing and helping build their data science / machine learning business, and Endeca (in Boston) working on large-scale e-commerce and business intelligence projects. Not exactly a traditional games background, I know!
I met Hilary at Cloudera and, after we’d independently decided to move on, thought it would be fun to work together again building something new. We experimented with a bunch of ideas centered around emergent technology — in this case, machine learning — looking for opportunities to create something right as it was becoming possible. Our favorite idea, by far, was to use natural language processing (NLP), something we both had a background in, to explore co-creative storytelling. How might a machine help people tell stories together?
We built our first prototype together in early 2020 to help us raise initial funding, and it was pretty clear that we needed help so our obvious first move was to hire Chris!
[Chris] 🌊 Hi! I’m Chris, Game Director at Hidden Door. That means I coordinate all things design and narrative related, and help shape the overall vision for what we’re building. I’ve been making video games for almost exactly 30 years now, including stints as a designer, producer, and programmer (of last resort). Some of the more notable games I worked on were The Lord of the Rings Online, The Beatles: Rock Band, and Rock Band 4. I’m also working with my son on an indie arcade game called Loose Nozzles.
I learned about Hidden Door through a network of game designers that I’ve been a part of for a couple of decades; we all came to that group through a background in story-based games. One of the other people in that group had met Matt and Hilary, and let us know about the open design position. I’d been making games at a small indie studio; it was a group I enjoyed working with, but I wasn’t sure how I’d fit into their next project. So the post about Hidden Door came at the right time.
We met through a series of Zoom calls, and I even did a design session with them to see how we might work together. I think we all enjoyed how that went, and I was excited by now new and, honestly, a little scarily unknown the project was. And so I when I got the offer to become Hidden Door’s first team member, I leapt in. And here I am!
I recognize many of those game titles! ✨ How does this compare to your past projects? What's familiar and what's different?
[Chris] Yeah, there’s a whole lot that’s different about this project. Probably the main thing is that so much of the experience is coming from generative processes that are inherently (and desirably) unpredictable. Part of my design method is to be a bit of a control freak, so it’s a little uncomfortable to be in a space where you’re shaping and tuning an almost organic experience.
On the other hand, fundamentally the work is the same; you’re trying to put yourself in the mindset of the player, and think about what sort of game mechanics and behaviors will create the experience you have in your head. Every game design problem is fundamentally iterative; you make something vaguely like the game you imagine, then once it’s up and running you learn how to make it better through a series of revisions. That mindset is just as applicable here.
Another way it’s familiar to me is that at Harmonix, the fun part was trying to figure out how to build satisfying game experiences out of new technology or new social situations. Figuring out how to create fun interactions with the Kinect camera, or how to turn three-part harmony singing into an accessible, party-friendly experience, are kind of tap-dancing a little bit beyond what previous games have done. Building scary new things is fun (if still a little scary).
How did you decide to build this product? What led the team to "this is what we want to make"?
[Matt] Ha! It's a long story. It certainly wasn't obvious at the start! I'd probably start by going back to the root question, “how might a machine help people tell stories together?” that suggests some principles framing how we thought about the problem…
First, we believe that machines aren’t creative; people are. So our goal isn’t to build a machine that tells stories, but to facilitate human creativity. Hilary had been an avid D&D player and had the idea to model the experience after that style of guided improvisational storytelling… most importantly, as a multiplayer game!
Second, our players are diverse and everyone has a story they want to tell. So rather than building a single narrative experience like in most games, our goal is to build a platform with systems that let players enjoy their favorite worlds, which also means we’re excited to partner with authors and other creators!
Third, too many online experiences implicitly reinforce gender, cultural, and other stereotypes. Also, as we’ve previously written, current leading approaches to language generation are deeply problematic, particularly for products aimed at younger players. They simply aren’t controllable. This also has implications for their usefulness in gaming and storytelling, where remembering the state of things is pretty important! So our goal was to build controllable systems — both for safety, and for better gameplay!
All that said, we started with a lot of (bad) ideas, spent a bunch of time identifying the biggest product risks, and spent most of 2021 building prototypes and playtesting to figure it out.
I wonder why others haven't picked up the ball with this before. Is it that new technology allows for this dynamic storytelling and we're just realizing new possibilities, or something else?
[Matt] There is a rich history of interactive fiction and entertainment; we're hardly the first. Over the past year, for example, one of our favorite things was Aaron Reed's 50 Years of Text Games series (now also available as a book!). Check it out if you haven't!
What's new, in the sense of emerging tech, is the recent work in the area of natural language processing and generation -- machines that can appear to understand and produce convincing language. We're building on that and applying it as part of an ensemble of techniques, some new and some quite old, to create the game and narrative experience we're aiming for.
Could you say more about the design process of creating the systems for organic storytelling?
[Chris] The systems for storytelling are interesting, and something that I’m sure Hilary will explain in greater detail in a future AMA. From my perspective, it’s trying to break down stories into structures that support both gameplay - objectives that you can work towards - and narrative - conflict, character motivation, theme. It’s something we’re developing in layers, as we expect to be surprised by what we find as each layer comes together.
So our initial explorations are largely plot-centric, and by extension action-centric, as those are easier to codify as objectives. Our intention is to use that to develop a language for story scenes that can be turned towards more introspective or relationship-based content. And to find ways to thread scenes together - scenes focused on specific action beats like a chase, or a rescue, or an investigation - into longer plotlines with dramatic arcs.
So we're building story systems from the scene outward - creating one dramatic scenario that engages players, and growing from that to a greater variety of narrative experiences.
🤔 When you said "remember the state of things", what does that look like from an NLP perspective? I heard on a podcast that you're not generating in real time; if so, how does memory work?
[Matt] So to set this up we can contrast our approach with other generative experiences that rely exclusively on carefully prompt-engineering a large language model (LLM) to produce text based on user input. The LLM itself has no idea what the game experience is about or how to keep it on the rails. It just generates tokens.
Under the covers, our system is basically a database of narrative worlds, stories, characters, items, locations, etc. Fundamentally, we take unstructured user input (words), turn that into structured game state updates to the database, and then render those back to the player as a mix of generated words and art.
You mentioned considerations around safety and controllability. Can you speak a bit about the risks you've thought through, your approaches to mitigating them, and ways you plan on monitoring your software to keep an eye out for unexpected behavior?
[Matt] Yes! Safety has always been our highest priority. Early on we made a few deliberate choices to that effect, primarily in building our systems to be controllable — we can constrain what players see — and observable/interpretable — we can tell why players saw something and if necessary, fix it. (Footnote: Roblox is another company that’s famously prioritized observability over privacy, to enable moderation.)
It's also why our architecture is a blend of ML and procedural approaches, and does (most) generation offline so that we can inspect and tweak the results before recombining them to serve to players. Said again, we'll never show players something from an LLM without our seeing it first.
Lastly, we've made a bunch of tradeoffs in multiplayer game experience such that we can we start with younger players (who we imagine playing with their friends or family) than most game experiences. We’ve also done the work to be legally compliant for players <13 years old.
There's no plan to turn this into any sort of NFT or crypto thing, right?
[Matt] Hell no.
How do you define success and improvement toward that goal as you work on this platform? Success criteria is all pretty squishy right?
[Matt] Is it squishy to say "make a game people love and want to keep playing with their real friends"? 😂
More seriously, right now we're focused on creating convincing narrative worlds that support continued play over many sessions. Imagine a rough equivalent of a D&D campaign that spans a few months, where you build up characters (less stats, more story). Hilary has also spoken about why this is fun from a machine learning perspective.
[Chris] And in what ways are they most excited to engage? Is it building up their characters? Having moments of more direct authorship over the story or world? Or building up a world through the telling of a succession of stories?
Even more fundamentally, do players feel like they have agency and can express themselves within stories in the way that we intend? And do they enjoy doing so?
🤔 There's a lot of emphasis on stories, and rightly so, but the inspirations from DnD have me wondering: How is "crunch" (e.g. mechanics for overcoming challenges, progression or leveling up) handled, if at all?
[Chris] That's a great question, and an area where we're exploring. There's a spectrum of experiences, with crunchy-tabletop-game on one end, and playground-make-believe on the other. Our ability to use language to express narrative in surprising ways leads me to not want to lean overly hard into crunch. On the other hand, having an environment where anything can happen means that you lack the constraints that can help motivate creativity.
For our current version we're exploring a thin layer of crunch - perhaps a delicious marzipan of mechanics - where many actions are framed as challenges, and a character's inherent abilities inform their odds of success. The first pass of this is actually a flat odds of success, just to see what it feels like to inject those odds checks into a text-based story.
And from there we expect to evolve the thinking with new mechanics - inspired from either TTRPG's, or perhaps even board games where your creative options are constrained in deckbuilding-y ways. The feeling is that once we've got the storytelling systems in order, applying game mechanics over them can be an exciting design opportunity.
🧙🏻♂️ Is magic part of the AI’s database?
[Matt] Yes! In the sense that our system can model various "rules of physics" in each world. "Magic" implies a class of abilities and constraints just like superpowers (magic by another name?), which we'll be able to apply and tune at a fairly granular level.
[Chris] That's a much more useful answer than I was going to post: "the real magic is the friends we make along the way.”
How do see growing the team this year?
[Matt] Currently we’re around 10 people spanning engineering, ML research, game design, art, and community. Fun fact: 1/3 of the team holds a PhD (Brianna wins for most interesting area of study). Over the next few months we're hiring engineers and a head of engineering, a head of marketing, a product designer, and a few other roles. We'll likely also have some internship opportunities for summer 2023!
This has been wonderful everyone! We’ve reached the end of our time; thank you so much to the community for your questions, and to Chris and Matt for sharing more about yourselves, the work you’ve done, and your vision and goals for Hidden Door!
[Chris] This was a lot of fun! Thanks for your questions and your time!
[Matt] This has been AMAZING. Thank you everyone for participating and asking wonderfully thoughtful questions!