"AI is good at tricking us into thinking there is intelligence there"
Harry Holmwood reveals how Magicave is leveraging AI for its new turn-based RPG while navigating ethical concerns.
Welcome to today’s edition of AI Gamechangers. Every week, we interview a leader in the games industry to find out the role artificial intelligence is playing in their work. Forward this to your colleagues and encourage them to register for free. We already have Q&As with award-winners Kinetix, Alison.ai, Quicksave Interactive and more ready to go. This week’s interview is with Harry Holmwood of games studio Magicave Scroll down for more news and links.
Harry Holmwood, Magicave
In today’s AI|G feature, meet games and technology specialist Harry Holmwood, co-founder and CEO at Magicave in the UK. Previously, he was managing director of The Secret Police, responsible for the mobile version of Stardew Valley and games like Dragon's Watch. He joins us to discuss Magicave's innovative approach to integrating AI into game development, focusing on their upcoming projects that procedurally generate daily content while emphasizing the importance of maintaining a human touch in the creative process.
Here are the top takeaways from our Q&A:
AI is best used as an enhancer rather than a replacement. You can sprinkle AI on top of a game (generating daily content and, in Magicave’s case, “meta-stories”) while maintaining human oversight and creative direction.
Current AI has major limitations and inconsistencies. Despite impressive demos, AI struggles with longer-form content and true understanding. It can be good at specific tasks like summarisation but falls apart with complex world-building.
Ethical implementation is critical. We talk about Magicave’s thoughtful approach to AI voice acting, demonstrating how companies can innovate while being sensitive to workforce disruption and rights issues.
AI Gamechangers: What is Magicave working on at the moment?
Harry Holmwood: We’re working on two projects, one of which has been announced, and one hasn’t. Bizarrely, the one that isn’t announced will come out pretty soon, and the one that is announced won’t because it’s a bigger project!
One of the products is a mobile puzzle game, which will come out in some form in the next couple of months, which we’re excited about. It’s by Ste Curran – it’s his baby. It’s what started us working together. It’s a concept that he a few years ago, and I worked with him, made a prototype of this game, and then founded Magicave. We’ve had a tiny team working on that for the last year or so. We’re really excited about it. It’s a whole new type of puzzle game; no one’s done it before, which is why we’re not announcing it. Remember, Threes? It was a great game, and then 2048 came out, and no one mentioned Threes again! We’re just trying to keep it under wraps until we’re ready to launch it.
And the other project: the purpose of Magicave was to create a games studio that is interested in, but not drinking the Kool-Aid of, the new tech that’s open to us now. We’ve looked a lot at web3 stuff, and we think there’s potentially a role for some of that for a mass audience, maybe not right now, but perhaps in the future.
AI falls into a similar category. The games industry is quite challenging. It’s hyper-competitive. Like many people, we’re asking, “Can AI improve our processes? Can we do things faster? Can we do them better?” But also, we’re asking, “What can we do with this piece of technology that we couldn’t do before from a gameplay point of view?”
We have these two games, and the only thing they have in common is that they both create new content for players to play every day. We’ve started to miss that a little bit in the world. We all used to watch the same TV shows at the same time, and we don’t do that any more, everything’s on demand. Occasionally, you’ll get something like Wordle, which exploded a couple of years ago. It’s a good example of a simple game, but everyone was doing “today’s Wordle”. And I think there was something in that that just made it interesting to talk about. It’s the only video game my mum has ever played! These kinds of watercooler moments you can create are fascinating – and if you want to do that, you need something for players to do every day.
And you’re using AI to help keep generating that content?
Big service games are quite a challenging business model in that these things can cost an enormous amount of money, not just to develop but to run. I’m interested in the efficiencies that new technology can bring. Every game is a huge risk, but what can we do that means it’s got a better shot at being a sustainable success?
Think about games people play for a long time, particularly on mobile – you release a game, and the chances of that game generating good numbers for retention or monetisation are infinitesimally small on day one. You launch the thing, and then you have to bridge the gap between where you are and where you want to be. “Have we got the money to get there? Or is the team so big that we’re burning huge amounts of cash to get there, and we can’t?”
We’re interested in where you can use technology to create some of that content, which means you don’t have to have huge numbers of people churning it out. It’s not necessarily about using AI to do work that otherwise would be done by someone else. It has more to do with how we can generate something that someone’s going to enjoy today.
I’ll give you an example: Ste Curran is a game designer, not a programmer. But he used ChatGPT to teach him how to make games in Unity. He asked how to do something, and ChatGPT would give him an answer. It may or may not have been right! But in that process, he quite quickly created a few simple quiz games. He created PubQuiz.ai, a game with AI-generated celebrity lookalikes which leaned into some of the weaknesses of AI at the time. It was all automated. We didn’t do this as a commercial thing, it was an experiment, and it turned out to be really fun.
It was interesting to use AI to generate content every day. We used AI as a sort of junior programmer/mentor for someone. Ste’s now written another game that is kind of his own personal project, which is another mobile puzzle game. It’s not our game. It’s just a side project of his. He’s written that using AI as a tutor, and I think that’s fascinating.
How does generative AI feed into the Beneath The Six work?
Beneath The Six is a roguelike multiplayer, turn-based game. And when we say roguelike, we literally mean ‘like Rogue’, so it borrows a lot from the old turn-based, dungeon-crawling, ASCII-character thing! It’s 3D and pretty, but at its heart, it’s about going through dungeons, finding loot, bashing monsters.
We’re still experimenting, but we wanted to include elements of lore and stories that put context into why you’re doing things. First of all, we started using ChatGPT to see if it could tell us a story about all this. As soon as you start playing around AI, you realise, first of all, that it’s incredible… And very quickly after that, you realise it really isn’t!
There are so many problems with it. Whether it’s image generation, video generation, or text generation, you very quickly realise it’s nowhere near as clever as you thought. You start asking, “What can we do with this that’s useful?!” We discovered very quickly that you can ask ChatGPT to write you a short story, and it will kind of do it. As it gets longer, it becomes more and more internally inconsistent. If you’ve got a big dungeon you’re exploring, and you’re looking for why there are goblins there and who they worship, you quickly realise the whole thing falls apart. If you pick up a book or a scroll and read it here, it will contradict something over there. You can’t currently use AI to generate big, consistent worlds.
But we did realise that GPT and similar models are really good at summarising. You give them an academic report and say, “Give me the bullet points.” It will do that okay! So rather than using generative AI to generate the whole story, we use procedural generation to create a story based on the data we’ve procedurally put into today’s game world. We came up with the term “meta stories” to describe this.
We’ve already publicly put out a tiny version of the game to test out the combat mechanics, which we called the Arena. Rather than it being about exploring and fighting and getting loot, it was just about fighting. You’ve got a single player dropped into a very small environment with some monsters to fight. Who gets the furthest today? At the same time, we wanted to dip our toes in the water to see what new things we could do with AI. Fortuitously, at the same time, one of our pre-seed investors, Concept Ventures, introduced us to ElevenLabs. At the time, there were a handful of people who had this really cool voice technology. They’re now a billion-dollar company a year later! But I think we were among the first people in gaming to be working with them.
We wanted to be able to describe to players, as a prototype of what we want to be doing more broadly with lore, what’s in today’s dungeon. “You’re going to find these monsters”, for instance, and add a bit of personality to it. We were pulling that data out of the game, saying, “This is today’s date”, and adding something relevant to today’s date. That’s the “meta story” – it’s data [from the game] that we then feed into an LLM which has been fine-tuned with a style guide using our own writers. That data is turned into a piece of text; we then take that text and turn it into speech.
We’ve trained a speech model with ElevenLabs on the voice of an actor called Tom Canton, who was in the TV show The Witcher. We were quite early into this area, and were keen not to be evil with voice actors and AI. We were starting to hear stories of voice actors finding that their voices had suddenly been cloned and used against their will! We pre-empted that and came up with a business model that said we’re not going to own this digital voice. We’re actually going to give it to the voice actor. He now owns his digital voice, and we license that back. When you’ve got these new technologies, there’s always a lack of clarity. There’s a lack of clarity around IP; there’s a lack of clarity around some of the ethics of it. Any new technology changes the dynamics of job markets. I think you need to be sensitive to that.
It sounds like you’ve been very thoughtful about how some of these things could go. Video game voice artists were striking over this sort of thing!
Yes, exactly. We did the deal with Tom last year. We were early in trying to think through the repercussions of that, and you don’t want to be the bad guys in that scenario. But you don’t want to stifle all innovation either. It’s trying to work out, if I was in that position, how would I want to handle that?
So, we generate this voice automatically every day. A dungeon is generated by our procedural tech, and this meta story is created. It gets turned into text by ChatGPT, then gets turned into Tom’s voice. It then goes through another AI to turn it back into subtitles so that we can simultaneously show subtitles that are timed with the text.
We’re not going to own this digital voice. We’re actually going to give it to the voice actor. He now owns his digital voice, and we license that back. Any new technology changes the dynamics of job markets. I think you need to be sensitive to that.
Harry Holmwood
It’s really neat! We’ve done a really good job, especially in training the voice AI. If you listen to a lot of voice AI demos, they sound very robotic, whereas we got Tom to really act. He was method-acting for hours on end, and those highs and lows really made a difference. If you train a voice on a really talented actor, you get much more humanity back.
Tom summed it up, saying, “It’s not as good an actor as me, but it’s hard to tell the difference”.
Do you feel there are a lot of companies that are yet to see the potential of AI and embrace the tools that are already available?
At the moment, you see the holes in it as soon as you start using it! One of the first things people do with ChatGPT is ask it about yourself. I remember being amazed that it obviously soaked up my LinkedIn profile at some point… but it gave me two “facts” which were complete lies. AI is good at tricking us into thinking there is intelligence there, and there isn’t intelligence there. So how do you control it?
I’ve been playing around a lot recently with video generation. It can be interesting if you want to make meme-type stuff that you just want to throw onto TikTok and see what sticks. What immediately becomes obvious is the impressive demos that you see… you have so little control over the output that they can only be getting those videos by making 1000 and choosing the one that isn’t shit! There is some skill in building these prompts, but broadly there is not enough control.
It seems in certain areas, like music, it can work better. I say this as someone who used to write music for video games for a living. It’s much better than I ever was at it now, and I don’t think that applies to art or text yet! It’s because if you want to put a bit of music over a scene in a game, or you want to put a bit of music over a video, there are 1000 different pieces of music that you can stick on it that would do.
Whereas, if you want to put a voiceover on something or you want to write a story for something, inconsistencies become very apparent. So you can say, “Write a piece of music that starts mellow and gets aggressive afterwards”, and AI can do a great job. Or I’ll write a bit of music, and I’ll run it through an AI, and it will come up with permutations that I wouldn’t have thought of. It’s nice when you can use it as a tool, and you have some input. I think that the risk is that you can probably delude yourself into thinking you’re a bigger part of the creative process than you really are.
It sounds like you’re very conscious of the role of the human in the process. Is that on your mind a lot?
Yes, and it’s the challenging bit. What we want for Beneath The Six is no human in the daily generation process. Phase one is there’s a new dungeon every day to play. Phase two is EVERY player has their OWN dungeon. And regardless of how much money you’re making, there’s no way you can have an artist building your dungeon and my dungeon every day. You can’t have a voice actor doing individual voiceovers for every player, so clearly, you need technical solutions to that. That generation process has to be human-free.
But if we were creating video at the moment, there’s no way you could create useful video without a human curation process. Most of the video coming out of AI is rubbish. So it’s about us getting that human element in there to define the style. Even then, currently, it can only do so good a job at consistently matching a style. I think we’re a long way off even concept art, and 3D is even further. If you’ve seen any of the 3D asset generation tools, you’ll know they are unbelievably useless at the moment.
Guy Gadney said that one challenge of using the existing models is that they can be quite boring – ChatGPT could never come up with a villain like Darth Vader…
I think that’s very true. Related to that is not just that AI is boring, but once you know something is AI, it becomes even more boring. I always draw a parallel with playing chess. I’m not good at chess, but I like chess, and I quite enjoy playing chess online against someone else, but if I’m playing chess against the computer, somehow, even though really it’s exactly the same game, the enjoyment is gone from it, just because it’s a machine opponent. There is some kind of need for human interaction, even if that human interaction is completely hidden from you.
There’s always going to be a role for the creative genius, for the person that creates a new genre of music, that has an art style that is incredibly distinctive, the best writer of their generation.
Harry Holmwood
For us, the AI isn’t creating the stories. We’re using AI as a little seasoning on top. Having it in the context of a multiplayer game adds something – just playing or conversing with an AI can be quite a joyless process.
If you could get in your time machine and look to the future, do you think AI will have a disruptive impact on games that everyone is predicting? Where are things going?
That’s a very good question. It’s the question.
The person whose opinion on this resonates the most is Demis Hassabis. He’s at Google DeepMind, but he’s an ex-games guy (Elixir Studios, and before that, he worked on Theme Park).
If you read what Elon Musk or a handful of similar AI proponents say, they will say AGI [artificial general intelligence] is going to be here this year. The reality is that LLMs are a long, long way from intelligence! We don’t yet know if this whole transformers, LLMs, thing is a step towards AGI or not. It’s certainly an amazing parlour trick. But what I read Demis and DeepMind say is that there need to be a couple more big leaps to happen in order for us to get to AGI. Maybe there’s a 50/50 chance we will get those in the next five years; maybe we won’t. If the smartest people in the world are saying that, that’s what I go along with.
GPT4 has no understanding of what it’s saying. But imagine if it did. It’s already better at many tasks than we are, and it has no understanding. Imagine if it had understanding! All of a sudden, it could eclipse us in a terrifying way.
Even though it has no understanding of many things, it can appear like it does. Going back to music, I stopped composing music for video games because I realised that I wasn’t very good at it! I was better than the other guys in the office, but I wasn’t world-class and would never be.
That’s what we might start to see with AI, and I think this is where people are, understandably, concerned. There’s always going to be a role for the creative genius, for the person that creates a new genre of music, that has an art style that is incredibly distinctive, the best writer of their generation… But the reality is, most of us aren’t that. Most of us do workman-like work. Many of us are like the artists in the 50s who drew the frames in between the keyframes in Disney animations. I’m certainly not the genius that would create Mickey Mouse. It’s challenging to think technology always undermines those things. Power tools took away jobs from people that were good with hand tools. It’s being a good, solid artisan, whatever it is you do, whether it’s coding or drawing or writing or audio, which feels like it’s at risk.
Humans historically strive to automate things. Whether it’s the plough, washing machines, or the printing press, we create opportunities to do things faster, don’t we? Even if it means older trades go out of business.
Historically, there have always been new opportunities that fill those gaps. It worries me that if the machines really are smarter than us, then it’s hard to see what those [opportunities] would be.
When is it really going to be useful and mass market? It’s quite easy to look ten years in the future and see what the world looks like; it’s really hard to look at next year and see what that looks like!
Harry Holmwood
It’s another thing that concerns me about the games industry: is making games becoming more and more of a hobby, and less and less of a living for people? In the same way that if you used to want music, you’d have to go to listen to musicians playing it live; then recording came along, and that changed things, and there are more people making music now than there have ever been, but not professionally. Fender, for example: most of their guitars don’t get sold to professional musicians. It probably used to be the case that anyone you knew who was musical was a professional musician. The number of people who are making content, it’s always expanding. It’s fascinating.
Your projects, dNo and Beneath The Six, sound fascinating. What’s the roadmap for you, and what’s your priority and focus?
We’re really lucky that we are a VC-funded company, and we’re a very small team, so we’re not in a rush, financially, to get products out. The next thing you’re going to see is this mobile puzzle game in a few months. Last year, we put something out from Beneath The Sixth, which was a system test. You’ll see the next iteration of that in the coming months too. It will be a multiplayer game that will have dungeons and generate daily content. At the same time, we do want to get as much feedback as possible! The first test we did, you can only get it on the Epic Games Store, with a code you can get on our Discord.
You have some interesting backers including the likes of Sir Ian Livingstone.
I’ve known Ian for a long time, and he has been an investor in every company I’ve ever founded. He was one of the first investors. We did a small pre-seed round when we first started. Games industry investment comes and goes; we were lucky that we founded at a point when money was flowing really well. With my previous company, The Secret Police, we made Stardew Valley for mobile, and that had done really well, so we were in a credible position to raise money. The timing was good, and people liked our message, which we described as “sceptical optimism” about these technologies.
There’s a lot of exciting stuff there, but it’s finding the path to where is it going to resonate with customers – when is it really going to be useful and mass market? It’s quite easy to look ten years in the future and see what the world looks like; it’s really hard to look at next year and see what that looks like! There’s all the sci-fi stuff if you look really far ahead, but a lot of things take longer to actually come to fruition than you’d think when you first see the hype.
Further down the rabbit hole
Some useful news, views and links to keep you going until next time…
You can see Harry Holmwood talk live on the Practical AI track at the Big Screen Gaming Summit in Helsinki on Wednesday, 2 October. Tickets are available now.
Published by Pan Macmillan this week: Supremacy. This new book by award-winning journalist Parmy Olson is about the competition between Microsoft and Google to develop mass-market AI and the ethical dilemmas involved.
OpenAI announced a new AI model this week. With enhanced reasoning skills, it was previously codenamed Strawberry but was previewed yesterday as OpenAI o1. It’s slower than GPT-4o and cannot search the web or parse images, but it is better at problem-solving, reports Wired.
Kinetix, the AI emote creator, has launched a $1 million fund to support game developers in integrating generative AI technologies.
You’ll be hearing from the Kinetix CEO in next week’s AI|G newsletter! Image from the Kinetix grant homepage.
