Hilary Mason: Deterministic vs. Probabilistic, and How AI is Changing Storytelling

29:31I don't know. It's it's very fuzzy. Yeah. What about artificial intelligence? So that one, I would say that term came back when deep learning started to be a useful set of techniques. um and i believe again it is primarily a marketing term used to differentiate from what has come before and and if we go back into the history of artificial intelligence as a field like you know there's been many rebrandings um machine learning itself was a rebranding of ai coming out of an ai winter um where the whole field essentially stole statistics and like started using it a lot and then rebranded a bit. But again, same people, same kinds of work.

30:19And so, you know, if you asked me in 2014 or 2016, I would have said like AI is a term of use that is largely driven by people who are now able to approach problems with deep learning. And that has given us a step function forward in terms of certain capabilities in certain domains, comes with a ton of trade-offs. Um, but it's not, I would still say that most things that we call AI systems are machine learning systems. Like they, you could draw the line in terms of like, is it like, what kind of black box is it? But like, that seems to be getting into pedantism rather than a useful, a useful difference.

31:01Meaning like, we're going to use this approach. We're going to use one approach versus we're going to use transformer models or something like that. Like, is that kind of what you're getting at? A little bit in the sense that like, if somebody comes up to me with a regression model and they're like, look at my data science. I'm like, cool. If they come up to me with a regression model and they're like, look at my machine learning. I'm like, maybe not. And then they're like, look at my AI. I'm like, oh, no, not OK. Right. Like, that's not it. But also, what is like, where do we draw those lines?

31:29And right now, like I'm doing it on Vibes. Like there's probably more more clear ways to do it. So we shouldn't think about it as such, hey, we have a problem to solve. Should we use AI? Should we use ML? Should we use data science? Because it almost feels like that's how people use these terms now. And it sounds like that's the wrong way to think about it. I don't find that to be a useful conversation. I find it more useful to be saying like, can we clearly state the problem we're trying to solve? Do we have a quantitative objective function we're trying to optimize for? If not, how would we even know what good looks like?

32:04And almost trying to get into like, is it meaningful to solve this problem? How is it meaningful? How meaningful is it? Is it even worth the investment of time and energy to do it? And if so, what are we going to do with that? because there are obviously, I mean, I'm the sort of person who can ask a million questions and get really deep down rabbit holes and then realize none of that work has actually changed my decision. So I should probably not have done it. But I do want to bring up that generative AI, which I believe is the name of this podcast. Generative now, yes. Something else. That's total marketing I heard.

32:40I heard our generative is like pure marketing. Oh, you mean broadly? I have people say generative AI is a pure marketing term. Is that fair? I don't think so because generative functions, again, are a thing. And they are a thing that says like, it is the kind of function that can take many examples and then produce things that sample from the space of those examples to resemble but not equal. And that's itself a research point. The underlying examples. I actually quite like that we're using the phrase generative AI because it is wrapping a little bow around that particular set of capabilities.

33:18And now I do think it gets overused to mean chatbots, which are terrible in many ways. So it definitely is an overloaded marketing term also. But I think there's a fundamental grain of truth in it, that it is worth mentioning the huge step forward function and capabilities around a lot of this stuff and sort of calling it out as an area of capabilities, whatever techniques you might use to actually produce that. I think we could definitely do a podcast episode that debates this because I've had people tell me the exact opposite. And both of you sound right. Yeah. We should do that at some point.

33:56Let's fight. It'll be fun. Um, cool. So talk to us about, I mean, I mean, you sold your business to Cloudera, which must have been fascinating and an amazing experience. I'm sure acquisitions are always like huge, huge learning experiences, whether they're good or bad, you learn a ton throughout them. Talk to us about that experience. And then maybe let's, let's hear the transition to hidden door. Cause I have so many questions about hidden door and I want to make sure we have time to get to them. Awesome. Um, I mean, the one thing I'll say is that anytime I was a solo founder and bootstrapped.

34:31And like, anytime you decide to sell a business, it is a very emotional decision. I think most people know that, but you also can't, this is the same thing for raising money. Like you can't just like the business is not more valuable tomorrow than it is today. Like businesses hit certain inflection points where they have, you know, like a certain amount of potential energy in them. And you can decide at that point, like what do you want to do with this energy um and fast forward was a business that was growing nicely like slowly and linearly like um had a really again great team really great culture and it was in a moment again it seems like it was like the little baby ai hype before the the big one but it was in a moment of a lot of people starting to build out ai product research um That happened to be something my team and I were very good at.

35:22And I had a couple of inbound offers that were incredibly interesting, like joining, you know, One Unicorn as their sort of AI product research team or joining another company, like to sort of accelerate some larger scaled businesses in an area that was very interesting to us. And I ended up joining Cloudera because it seemed like the best fit for my vision for I wanted to go, the scale I wanted to get to, the kinds of work we wanted to be doing. And in many ways that did indeed work out. It was really interesting to join a large then public company to also honestly, for the first time, be an executive in that kind of context.

36:01As you can tell, that's not my natural habitat, but one that it was fun to play in for a few years. And also to think about like, how do we now take the scale of this business and have the kind of impact we want to have. And we built a lot of great stuff. So when I say fast forward, did our own research, like we had very wide research interests and anyone who's interested can still go to blog.fastforwardlabs.com and you can see a lot of the stuff there. But, um, and you'll see a lot of like our very first research report was in natural language generation in 2014. We built an app that wrote real estate advertisements for New York city apartments.

36:36And we wrote a technical report with a bunch of code as to like how you too can have this fun, this fun capability. And again, did a bunch of work in summarization, both extractive and abstractive using embeddings and then using what would now be called generative AI to our earlier question. But we also worked in things like causality and in sort of multitask learning systems, which was one of those research things that absolutely blew, like it broke my brain for a week. the idea that training one deep learning model on multiple unrelated tasks could actually give you better results than training a single model.

37:14Like, I love those moments where just your, your fundamental understanding of the tech, like just gets shaken and re re scrambled around. Um, we had a bunch of those at fast forward. Um, and then what we would do is actually take a lot of that stuff and help people build products. One of my favorites was, um, with state street bank, We helped them build this app that sat on a news feed of 15 ,000 news articles a day and then clustered and summarized them for their commodities traders. And again, the philosophy of this was very much like machine learning in this sense and embeddings everywhere is really great at organizing a volume of information that is outside the scale of the human mind.

37:55And I bring this up as an example because it is the kind of technical and product thinking we're doing a lot of because it is really it is not about coming in and being like, you know, we're going to replace people with our algorithms. It is really about thinking about what is the capability of this thing? What is it great at? How do I use it in a way that helps this person do their job more effectively, you know, or even have insight into any of that stuff? we worked with another bank on like all the emails their successful salespeople sent all the emails their unsuccessful salespeople sent and like how do you build essentially a email coach for to get people up to mediocre who are sort of struggling using that data and again this was all similar tech but like all before this generative ai rebranding um it's about making people smarter, more efficient, more effective.

38:49It's not about replacing people. Yes. Talk to us about how you ended up leaving Cloudera and when you decided to start Hidden Door and maybe what inspired Hidden Door. Yeah. I mean, I'm really, I had a great time at Cloudera, but also it became very hard to lead a, you know,$70 million a year business inside a billion dollar business where, you know, the job was not to build new things with new people. It was to, you know, manage internal politics. So it was time to move on. I'd had, again, great time. I met my co-founder there. And, you know, let's come around to a hidden door. So I'd already had this absolute fascination with the technology of summarization, embeddings, even generation.

39:40Can you clarify embeddings? All right. So if we're talking about something like embeddings, the thing that changed was that before embeddings, we could not do this. After embeddings, we could do it. was that it made language computable and comparable. And it did that by essentially mathematically representing the relationships between words or sentences and collections of words. It was just this ability to take language, make a vector out of it that is a compression of that language based on the relationships between things in that language and then compare it to another vector and say, essentially conceptually, like, do these two bits of text have the same meaning-ish?

40:22Are they similar? And so embeddings are a way to sort of generalize anything that does that. Is that a way to think about it? Yeah, it is a way to take a whole bunch and like in this case, a whole bunch of language. So like a lot of it. And then from that, say like, OK, these tokens, which are generally words is an easy way to think about it, are clustered in a space together. And then for any new document, which can be a sentence or even just one word or a whole document or 10 ,000 pages, like here's the representation of it projected into that space. And I can compare that representation to others.

41:02And if you want to do what? What does that teach you? What does that teach you? Or what can you learn from that? It gives you a sense of the probabilistic overlap between those things and also can often be computable. So the classic word to back example is like, you can do the computation of like king plus woman and you get the word queen out the other side. And you can do analogies mathematically in the sense of like, you know, United States, Washington, D.C. and then like Canada. And you'd expect, you know, Ottawa to come out the other end, like analogies in the sense of like the SATs. and I want to give you a hot take though because I think that embeddings as a class of capability are going to end up being more impactful for the products that can be built than like text generation is okay so you'd been studying this for a while building stuff with it building stuff with it and this this leads to hidden door yes though not directly I think anyone who leaves a high pressure job should take some time off.

42:12So I did that. Definitely. Very kind to myself. I had a lot of fun with that. And then started thinking in this area of tech, like, again, I like this area. I still felt like, and I think this is actually even more true today, which is strange to say, but I felt like we had not yet explored the products that had become possible because of these generative AI capabilities. And there was a wide open space. And so my co-founder, Matt Brandwein, and I started exploring that space from a product and business lens. And we were looking for something that, you know, was essentially something people would want, but that hadn't been economically feasible or really possible without the tech.

42:58something where the business model was generally additive. Like going back to my social media experience, like I have some bad reactions to ad-driven business models. I don't really like selling people's data when they're not opting into that. I think it is quite possible to build businesses that are where everyone understands the economics and are content with it, where people are actually paying for a thing they want and people are getting paid for work they do. We're looking for that. And then also thinking frankly deeply about the side effects of whatever products we're building in the sense of, I'll give you one example.

43:31We explored this idea of like a problem I have every day is like, I'm communicating with people on email, on Slack, on WhatsApp, on Signal, on texts, on wherever I'm reading a bunch of stuff on the web. And if we have a message, I have to be like, oh, was that an email or was it a text or was it, where was it? I have to go find it. I have to search five things. Why could I not just pull that data into one place? And then, you know, as I'm working on something, it can be, it can be summarized for me like again this is going back to like the star trek sort of like two button presses on that computer can do anything in the universe like how do we get to that um but we realized that uh frankly the business dynamics the costs of running something like that are very high the business dynamics are not great like you're selling to professionals this fairly like saturated market um so okay you're gonna do an enterprise sale oh shit you've now built a surveillance product, like where companies now have full access to everyone's text messages.

44:26I don't want to do that. Right. Right. So, so it was a process of going through the space and then realizing that actually a lot of the challenges of the tech and, and like, we can get into this because generative models hallucinate. It is by design. It is what they are and what they do fundamentally and mathematically. What do you mean by hallucinate? I mean, they spout stuff that is not true. So you can ask, and we all have egos, right? Like instead, the new Googling yourself is to go to chat GPT and say, who is Hillary Mason or who are you? And first even see if you were lucky enough to be tokenized in the model.

45:07And then if you are, what is it going to say about you? And it's almost certainly going to get it wrong. Something wrong, like maybe directionally correct. I actually had this happen where somebody wrote this beautiful LinkedIn essay and they quoted me in it. And it was a quote I'd never said. And I reached out to them and I was like, oh, this is great. But where did this come from? And she was like, oh, ChatGPT did it. I'm like, oh, you know, well, it was something really lovely about like data should be democratized and empower everyone. I'm like, I totally say things like that, but I am sure I never said those exact words.

45:39It was just really funny. And so hallucinations are because the model is compressing the underlying data and making things that look like it, but not exactly it. And it has no sense of fact or no sense of reality. And that is the, like, it is what it is. It is not something you can fundamentally change without changing the architectures pretty dramatically. So we were in this space of thinking about like, where does this actually become an asset? Where is this actually fun and playful? And it's in fiction. It's in telling stories. It's an expanding. It's when you don't want or care about factual accuracy.

46:19Um, but it opens up a whole host of other problems. So at the time we were working on this, initially, GPT-2 was out. You could, it was, frankly, like super sexist and racist. Like you, we were asking it, um, to generate kid stories to show our kids and it was coming up with things like women hate technology. I'm like, you know, I'm personally offended by some of this. And so all of this was input into thinking about what would an architecture be that allowed us to create safe and controllable stories? And by the way, and this goes back to your earlier point, when we read something linear, it has a beginning and a middle and an end that has been designed.

47:05And there are structures in that that are valuable in communication. how can we create that value while still allowing for the incredible openness of a probabilistic story experience and how do we by the way give the power of that to the reader of the story and that's where we had the insight that oh this isn't about stories it's actually a game Like it's a video game. Because games are about interacting with the world, exploring your creativity, and the affordance is there to be able to say, okay, we're not generating what you see until you interact with it. It's sort of like we call it our just-in-time generation.

47:51Gives us that ability to both apply structure, but also explore the space and the context and allows us to use models for what they're great at. and not use certain models for the things they aren't great at. Super interesting. So tell us exactly what the product is right now. So what we're doing at Hidden Door is working directly with authors and creators and other IP holders to take the worlds that they have already created through their books or movies or TV shows and make them playable social role-playing games with the dynamics of something like a Dungeons & Dragons game with an AI narrator.

48:29player so from the player experience you are a fan of something you read the book you can't stop thinking about it you saw the show you have your own ideas for what stories you might want you can then go someday we want you to be able to like read the book listen to the book play it on hidden door and it is a social game because a lot of the fun of this is not just like hey i made the computer make a funny thing, but rather we together, each with our own characters, collided in this world we already both know and love in this fun way. And so what we're building is a platform for having these fan experiences inside the world people already know and love.

49:12Did you, I know you were focused on kids, you mentioned that earlier, is the fact that these models are not deterministic, did that factor in to your to move away from building for kids? No. In fact, we started because that was personally, you know, of interest to have kids, want to make stuff for kids. And like realized that in order to build a safety, a safe architecture where like I would be comfortable putting a generative thing in front of a kid, we had to build controllability. Controllability gives us the ability. And I think that is a fairly unique thing in the market even now because it is, it was a large technical investment.

49:55But it gives us the ability to sign contracts with people who have contractual requirements about how their characters need to behave because we have that level of programmatic controllability in our system. And honestly, we moved away from just kids as our audience because we were doing all of the Zoom play testing with kids in the pandemic. And the kids would play and they were into it and it was fine and like we got good feedback we got weird feedback like it was fun we had one one girl who was like this is my favorite thing ever and we were like great what does it remind you of she's like english homework english is my favorite class we're like oh this is not a good game we're gonna go back and work on that some more um but like uh actually it was because the parents would would then like they would always be hovering in the background because like when someone's lending you their kid and they don't really know you that well like of course they're gonna supervise and the kids would be done we'd be like cool anything else the parent would be like can i play?

50:46I want to see. And then, you know, and so that was more of a product business insight for us rather than a like data or technology constrained one. And we still built in like all the safety and compliance to have kids as young as nine in the game. And eventually like we will, but our emphasis on launch is actually 18 to 34. Cause that's where we're seeing that, frankly that fan energy like the people who would um like maybe comment on reddit about a movie they saw or like read a fanfic or you know talk about it it's that but it sounds like this controllability the word you used you know i i applied it in the sense of controlling the experience for for kids but it sounds like that's actually important to the the demographic you are building for as well Like you said, IP holders, authors, creators.

51:39That sounds really hard, like a hard problem to solve, given hallucination of models, the probabilistic nature of them. Talk to us about building in that way. I mean, that just that seems like a huge task. Is that is that true? It's all relative, right? It's a fun, hard problem. And it's the thing we started with. And again, like we were living in a world of GPT-2. GPT-3 came out when we were starting to work, when we were building this. And there were a few things that were very clear, which is that when you have unstructured input in, so text in, you do some stuff, you have some big knobs to tune, like how much repetition do I want?

52:21How far out on the distribution should you go in giving me an answer? And then the answer comes out. You cannot put that raw in front of a person. Or at least you cannot do it and expect any... You cannot have the levers of controllability that we wanted. And those are in many dimensions. And so what we've done instead at a very conceptual level is essentially that we accept unstructured input. Like you still put in words and we generate sentences for you. And then you pick the sentence you want. You can write in a sentence if you want to. Unstructured input comes in. But at every turn of the game, we do a bunch of work to update what we call our game state, which is basically a game engine representation of everything in the world, its relationship to each other, where it's located, its properties.

53:07Some of those properties are represented in language, like your character might be energized because you, you know, drank the coffee, and you would have a condition energized, which would be a language-based token. But we have pre-built a dictionary of tens of thousands of words and phrases with the metadata around what they mean in terms of materials and properties and conditions and what that means. And so at every turn of the game, everything maps into that game engine database. And then from that, we do the computational work of controllability and say basically like, what is allowed? What isn't allowed?

53:42What is happening? We update that game state and we use that data structure to generate the text and the art you see in the game. So it is not this like words in, words out with very crude levers of controllability. It's almost a different thing. And we're doing something we're doing right now at 16 different machine learning tasks in every turn of the game. Many of them are done with completely homegrown models. Some of them are LLMs, but the LLMs are used for what they're great at, which is more or less, we might say our game. We have what we call a story governor. That is our sort of like GM process, which will be like, hey, at this point in the story, this is the rail I'm enforcing.

54:21I'm going to introduce this trope to like make things fun and colorful. we have a library of thousands of tropes that are pre-generated hand edited sometimes handwritten which are largely in text but also with metadata and then what we do is use the llms to color them we say like great we're in a scene in you know this like we've adopted the wizard of oz for playtesting we're doing pride and prejudice next like all this stuff we're in one of these worlds what does a barb roll feel like right now for this character in this situation in this moment like who are you going to say is the antagonist?

54:56Who's going to play the ominous music? Who's going to make the found weapon? And so I'd say that like, yes, it's a hard technical problem, but also we've decomposed it. And then we've built, we're trying to use each system and each capability and each bit of AI for like what it's great at. We use embeddings all over the place. We use, you know, a few different LMs. We've built a bunch of our own stuff. We also use a little bit of procedural generation. We do a lot of pre-calculation, which also means our thing actually runs at scale at a reasonable cost, which is another big issue for AI startups.

55:35So yeah, it's fun. Who's telling the story? Is it hidden door? Is it the AI? It sounds like you've built some guardrails, you've built some controls, but it also strikes me that there are ways to tell good stories. Yes. So who's doing the storytelling? It is a collaboration between the original world author who has set out what is the starting state of the world? What is the moment you enter it? Who exists in that world? What is their condition? What needs do they already have? What story arcs are in progress? They set out the laws of physics, like what kind, to map it into our sort of trope space, like what kinds of stories do I want to be generated?

56:22what kinds of characters what other rules are there is there a lot of death no deaths how far down the sexy times rabbit hole are we going um like are there specific things like this character can like in the wizard of oz in our adaptation dorothy can only be hinted at you never meet her but she is in the world at the same time as you because that's the moment that most people are most familiar with in the wizard of oz um so you set out all those rules and then the system is the thing that puts the guardrails on. And honestly, it is a game. It is not a book. It is not a writing tool. So you have an objective.

56:59And then you as the player are saying like, cool, I can use language to meet this objective or avoid it however I want to. And so that objective might be something as simple as like, there is a, you know, a robot character, you know, missing a limb blocking the road and you can choose to like befriend it, to help it, to, you know, smush it, like, you know, whatever you want to do that the system decides meets that objective. And that, by the way, is where a lot of those embeddings come in. It's like, does this thing meet this generic form of the thing in language? Who's using the platform? What's the status of it?

57:42Yeah. Can you share any of the IP? holders using it or or not yet not yet we are launching in november um wow with a let's say a few different genres to show the to show the full range of um of what it's capable of and uh the kinds of stories that it's great at in those different genres and where will the games exist do they exist in hidden door do they exist in you know rights holders website. They're on hidden doors. So you will come to, to hidden door to play. That's awesome. Yeah. So the rights holders will send, we'll say, Hey, you want to, I don't know. I don't want to, I don't want to use any examples in case I'm, in case I'm right or wrong.

58:27But imagine I'm some big media property and I've created some fan experience using, using hidden door. I will send my customers for my website or my experience to hidden door. Right. They get to interact. Said another way, we're doing a lot of cross promotion. A lot of our go to market is reaching the fans where they already are. And sometimes we do that. Sometimes they do that. With the authors we're working with, they usually have pretty great social media connections with their fans already. So that's really fun. And we also, if any authors are listening, love when people want to give their fans that extra way to engage with the world while they're waiting for the next book.

59:04But then also they can create like sort of special content for the hidden door story world that would maybe foreshadow some plot points or some new characters you might encounter. So we're experimenting with ways to provide that kind of access as well. And is this something that they'll monetize? So they send people to indoor and then you'll, I don't know what you'll sell or how you'll monetize if they'll get to participate in that revenue. Yes, absolutely. Our monetizations, we're an alpha, like we're going to be an alpha. We're not charging anyone yet. But the monetization strategy is largely around that is that let's pay the creators of these worlds or it's a marketplace, right?

59:47And the creators certainly get a fair bit of that revenue. The other thing we're looking at is our players sometimes want to have access to more of it, like more characters in a particular world, more friends joining the adventure. So we're exploring, inspired more by Discord's model, some opportunities there eventually. Again, that's not where we're putting all our energy right now. Just an incredible experience. Of course, that is the most important thing. Going back to building these games, building storytelling through this way, it feels like programming, software engineering, it's so rooted in logic, right?

1:00:33Do this, you know, create this function, get this output. But it sounds like building with AI is not that way, right? You don't know what you're going to get out on the other end. How does a software engineer kind of shift the mindset when it's no longer perfectly logical? It's no longer perfectly deterministic. It's a little bit of a surprise every time. I think one of the biggest mindset shifts in doing data science or machine learning or AI work versus doing software engineering work is that you are moving from that deterministic world where we probably know the thing we're building is very possible.

1:01:15at the beginning to one in which it is probabilistic and we likely don't know how possible or at what level of quality we can address the original problem and there are a couple of things you have to learn one of which is that um and this is also why i always say that like agile is where data science work goes to die it's because it was a process built to support this deterministic construction of software and not one to support a sort of research driven data machine learning process where honestly like the dirty secret of great product machine learning work is that you start out with one problem you're trying to solve you realize you can't quite do that at a level of quality or speed or utility that's useful to you and you reframe the problem driven by what your product needs are to solve something else and then you try to solve that and then you probably do that three or four different times.

1:02:10And so it requires sort of thinking about the problem space in a different way than a lot of software engineering does, because there are certain things you could take as, you know, you just take them for granted as assumptions about what you're doing. Like you're making an app, but you know apps are possible to build. You do not know that your model is possible to build with the underlying data unless you've already built that model. The other thing I'll say is that in this particularly weird world of generative AI stuff, of a lot of the work of building a great experience is actually in getting a really strong intuition for what is being encoded in the specific model itself and how you might access it and where you might put it and what you need to build around it to make sure that you're getting what you need out of it.

1:02:57And by that, I mean, I'll give you another hot take. I think prompt engineering will not actually be a job forever. and what prompt engineering is, is just getting an intuition for the underlying data encoded in the model and figuring out the language to like pull it out the other side. Yeah, that makes sense. It's like a moment in time thing. Yes. For these models or at least the way we're interacting with these models. Yes. Which could change. Yes. And hopefully, hopefully will in a bunch of ways. But then the other thing I'll say is I talked to a lot of founders and like people engineering products on top of generative AI stuff.

1:03:36And like every single one is building a system on top of the model to make sure that what comes out is actually what they want to put in front of their people, their users or their players. And so it's making really smart decisions about all of that, which is a completely different mindset, honestly, than where software engineering starts. It feels like we haven't yet wrapped our heads around all the different ways in which it will be different. product development, that is, right? You know, you think about going from web to mobile. We went from clicking to tapping. We went into mobile thinking it would be the same.

1:04:11It would be skeuomorphic. And then we just tried to sort of like wrap mobile around that framework. Yes. What are going to be the big differences in generative, you know, to use the word generative, generative product building or AI first products? I mean, you just made a great point that I will call out explicitly, which is that frankly, like we don't know what the design metaphors of these products are and the design skill is as different as the software engineering skill is in its own way and we're kind of stuck on like chat as metaphor but i think it's a terrible interface and it's also not fun for anything right it's not efficient and unless you actually are chatting for the sake of chatting like that is the activity, it is messy.

1:05:00It's just a lot of friction, right? To get a thought into input into the system, you have to, especially on mobile, right? You have to tap a bunch of buttons with your thumbs over and over and over. It's not that efficient. No. So what will AI native products look like? We don't know yet. You just made the point. We don't know. But if you had to guess. So I think there are, the way I like to approach a question like that is basically to tease out the assumptions I'm making because each of those assumptions is highly debatable. And then they sort of drive to a conclusion. And the assumptions I'm making are that we will actually use the ability to generate what someone sees like at the moment they see it, which allows us to have a lot of contextualization, personalization, whatever you want to call it, that is currently not something we see that will find efficient interface metaphors.

1:05:51And we will, as a community, sort of grow up around them ways that we can understand what's going on or like direct it. I don't think we found those yet. I think there are things we can look to and the way we interact with each other as people as inspiration, but that's the stage we're at. And then it's thinking about like, okay, we have a bunch of constraints at doing this stuff right now. Like, frankly, text output is great. Image output is variable quality and controllability. Like we have a whole different visual art generation system at Hidden Door where we compose things dynamically, but we generate them offline or hand draw them because we want a level of consistency and quality and style that is not yet there.

1:06:35Maybe a year from now it will be. And then there's sort of the multimodality of the world, which we're starting to see a lot of energy into multimodal interactions, but again, no interfaces around them that are sort of standard. So there's a huge opportunity space there. So given all of those assumptions, I mean, it's, I really think about like, it becomes a set of tooling for navigating information in ways that are interesting and fun and contextual to our world and to ourselves. Oh, and the last thing I'll say is that right now, most of these models require GPUs for inference or other specialized hardware.

1:07:11We are starting to see stuff that runs on CPU. And I want to imagine what it looks like when we can run it in the browser, run it in the mobile app, right? Because that changes the design a lot and what you can do pretty dramatically. You earlier in the conversation mentioned that AI hype is dangerous, which I actually tend to agree with. Talk to me about why that is. Why is AI hype dangerous? It's a hard thing to address concisely, but I think it is dangerous in a bunch of ways because a lot of people rightfully react But what they're reacting to is the hype idea of AI. And we have several people out there pushing, you know, sort of AGI dumerism, which is as far as I can tell, you know, a religious affliction.

1:08:01And we have other dumerism is a religious affliction. The absolute focus on an existential threat, while largely ignoring real scaled harms that are occurring today, shifts our attention to what could happen in a way that is largely setting up business dynamics to be useful and profitable for a few large companies and not for not helping anyone who is actually today impacted by the tech or anyone who is actually today. who wants to, you know, be a small company creating something new in this space. And it is not coming from like an objective evaluation of harms because that does exist and that work is being done and that's largely being swept under the like, but what if it came to life and wanted to kill us?

1:08:53And also, you know, that is in many ways, like that is dangerous. We've had that in many other areas of tech too, but it's something in this community, I think we do have to speak to. I also think that hype is dangerous because it convinces people that hard problems are easy. And, you know, the conversation five years ago used to be a lot around like, well, that's not possible. Or, you know, like, how does that even work? And now it's more like, but that's, you just plug in chat GPT and it's done, right? Like, that's all you're doing. Like, and I think that leads to a lot of sloppiness and a lot of things that are built.

1:09:37And again, one of the issues with machine learning, AI things in general is that they scale, they work at scale. And that means they can scale harms. They can scale them very easily. They can scale and magnify underlying biases in data. That is kind of what they do by design. And it is a large, it is a huge distraction from the real issues we already have with all this stuff. No, it's interesting. I mean, the big theme between both of those things, I think is that the hype will overshadow the really important and frankly, like the very, very valuable outputs of what is being built right now, which could potentially transform everything, everything, right?

1:10:20About how we work and how we live, frankly. So maybe, okay, now let's flip it. Okay. What is, what are we optimistic about? Like, what is, what is the great potential of everything that is happening right now? And what gets you really excited? Yeah, I'm, I'm an optimist and like very much a pragmatist, but also deeply optimistic. And it is, you know, it is the ability to interact with computers using interfaces we use to interact with each other. It is the ability to build computational systems that can help us model and understand information that we as like our poor human brains cannot possibly consume alone.

1:11:00It is the ability to think about taking a lot of the drudgery out of the, not the creativity and not the joy and the love and the, the like emotion of it, but the drudgery out of a lot of our tasks, like finding a time for a calendar appointment, you know, or things that are, you know, like, oh, I, you know, I need to remember to pick up coffee beans on my walk home, like, you know, put it that somewhere where I can have it remind me when I get close to the place where I always do that, right? It is using these systems in a way that create more space for us to be human that I'm really excited about.

1:11:40And you see this in our, like the work we're doing in Hidden Doors is like, we're all fans of stuff. We all have been creative together. The game is really about being a scaffold for us to have these like, really creative, fun, happy experiences with our friends and with our family. And to explore things we already love. That's the stuff that I'm emotionally excited about. And then very practically, it comes down to giving us the ability to control our space and our devices and having input into the way things are parsed for us. Like if I'm reading a news article, why do I have to see exactly the same article you see?

1:12:18Like maybe you're an expert on the topic, but I know nothing and I need five extra pages of background and you need only a paragraph to be like, here's what's changed. Why don't we have that? Like, help me get there faster and better and remove the extra work of navigating this world of information we all live in. That's a great way to start to wind down our conversation. Real quick, Hidden Door, what should we be on the lookout for? It sounds like you're launching soon. Where can we learn more? Yeah, Hidden Door.co. or I'm HBson on all the socials and hiddendoor.co on all the socials as well.

1:12:51And we are launching in November in Alpha, bunch of exciting worlds, bunch of exciting stuff happening. And if you love the energy of collaborative storytelling and role-playing and doing that in worlds you already know, like come sign up. That sounds awesome. And are you hiring? We are. We are hiring a few team members in engineering right now and expanding that next year. And where can we learn about those roles? Also on hiddendoor.co. And are these roles in New York City? Are they anywhere in the world? Oh, we're a fully distributed team. We're from Vancouver to Amsterdam and try to stay within that time zone range.

1:13:30Awesome. Hilary, thank you so much. This was a fascinating conversation. I really appreciate you giving us all this time. Thank you. Thank you all so much for listening to this episode of Generative Now. I really hope you enjoyed my conversation with Hilary Mason, co-founder and CEO of Hidden Door. If you like what you heard, please do us a huge favor and rate and review the podcast on Spotify and Apple Podcasts. That stuff really, really helps. I am Michael Magnano. If you wanna follow me, you can find me at Magnano on all the different platforms. And if you wanna follow Lightspeed, you can follow us at LightspeedVP on all those same platforms.

1:14:10and like always we'll be back next week with another awesome conversation of generative now generative now is a production of light speed and pod people special thanks to everyone on our production team make it all possible see you next time