The genesis and secrets of Lionhead's latest curio
If there was one thing that generated a lot of buzz at E3, it was Microsoft’s Project Natal – and, more specifically, Lionhead’s latest curio: Milo, the virtual boy.
Similarly, nothing from E3 has been so misunderstood. Look at your average forum or blog and you’ll find people polarised into two camps: those that believe it was an elaborate set-up, and those that believe it to be the biggest development in artificial intelligence and the beginning of a new era for human-computer interaction.
When we catch up with Peter Molyneux at E3, he isn’t entirely surprised by the reaction. “People want to believe this so much,” he tells us, laughing. “Journalists have already come in, having done research on the Turing Test, and stand there saying the classic question: ‘Do you remember what we talked about yesterday?’ But, of course, he’s not really intelligent – what we’re showing is a tentative first step towards making something that feels real. And by real, I mean that he’s aware of you.”
The truth, of course, is that Milo is equal parts smoke and mirrors as he is an intelligent agent; it’s less cheating and more exploitation. “An awful lot of Milo is like a Derren Brown mind trick,” says Molyneux. “Derren Brown can’t really read your mind, what he does is collect together a number of little tricks that tell him things. You’ve got to remember that we’re not creating a piece of academic research; Milo can’t actually think – we’re just making the illusion that he can.”
What Milo typifies is a long-running ambition and interest that has permeated through each of Lionhead’s titles: what Molyneux calls emotional AI.
“At Lionhead, we’ve always been fascinated by AI and using AI in gameplay. A lot of AI is academic research; funnelled down neural nets and learning – it’s very antiseptic, there’s no emotion behind it. It goes back further than Lionhead, back to Bullfrog, where we were playing around with these simulations of little people and their little minds. Emotional AI is not real AI; you couldn’t write a paper on it. It’s more about how you use weak learning to make people think there’s something going on there.”
It stretches all the way back to Populous, Molyneux recalls, with people convinced that the villagers had intelligence and desires. “They swore that these little people exhibited fear, and curiosity, and ambition, and all of that stuff, when none of that was really there – it was just a random number saying ‘walk left, walk right, walk on this path’. People think it’s far deeper than it really is, but I guess that’s still AI – it’s this emotional AI.
“There was this stupid thing in Theme Park that got us a huge amount of press, but was really just a tiny thing: you could put a drinks stand next to a stall that sold chips, and then if you increased the salt content of the chips you’d sell more drinks. That wasn’t real AI, but people thought it was great – that these little people were getting thirsty and everything, but it was just a single line of code. What’s interesting is that, again, it’s the belief that there’s more going on behind the scenes than there really is. A lot of the time it’s exploiting the numbers in a way that makes it look interesting.”
But there was one title in particular that also cemented Lionhead’s reputation when it came to AI, and something a little different than the emergence of behaviour in simulated communities. It was the company’s first title, Black and White, that introduced the Creature – something that would shape almost everything that Lionhead has done since, from the dog in Fable II to Milo himself.
“The creature’s mind was kind of a breakthrough. ‘Mind’ is an evocative word, but we had this system that was based upon desires – it had a desire to eat, a desire to be happy or angry. It had ways of satisfying those desires, say by eating or sleeping or being aggressive. And then there were actions that could satisfy those. And because those actions weren’t fixed, the player could decide through interaction which desires were more important, and how they should be satisfied. That really felt like this was your creature – some people could say ‘my creature likes eating crops from a field’ whereas others could say ‘my creature likes to eat people with blonde hair’. It went down to that resolution. This was much more like a mind; it wasn’t a sterile neural net.”
It was from this, the experience of having created something complex that people felt attached to, that another project was born – the fabled and long-teased Dmitri, which eventually evolved into the Milo we see today. Having tackled a creature, the team found itself asking whether they would be able to do the same thing for humans. Could they create a virtual person that people would become attached to?
The first key decision – and one that still illicits sniggers from all over the industry – was that it would have to be a child.
“We knew we couldn’t do an adult – adults are tough; they are complex and disturbed creatures,” Molyneux explains. “As adults, we’re the products of our childhoods and our experiences. Look into the eyes of an adult and there’s all this history there. If you try to make a real one of those, it’s very very tough to do. But children are semi-mad; they’re unpredictable – and that means we can get away with a lot more.
“The big problem with Dmitri, the one we couldn’t solve, was the controller – how can you give the illusion of something being real when the only way to interact with it is to press a button? It doesn’t work. Even when I talk to you over e-mail or Messenger, you don’t feel real to me, it’s only my experiences of you that make that like a real conversation. We were never able to solve that, it never really worked.”
Until, of course, Redmond showed them what they were working on. But while the technology in Natal certainly brings a whole new range of possibilities to the table, it was what Lionhead discovered within the labs of Microsoft Research that really took that interface problem away. Given its wide remit as a modern technology company, Microsoft had already been doing work on voice recognition for Windows 7, as well as a whole raft of computer vision projects covering object, body, face and handwriting recognition – so Lionhead simply put it all together, alongside almost a decade of research into developing a empathic virtual personality, to create Milo.
“When all of this came along, it was like, ‘Wow, our biggest problem with Dmitri can be thrown away’. And that’s why we were so far ahead of the curve – we’d done all of this backwork, so it was amazing for the people over at Redmond to see that we had Milo come to life after just a couple of weeks.”
So, we ask Molyneux: does emotional AI, the most current iteration of which we’ve just been exposed to, take as much inspiration from psychology as it does computer science and academic artificial intelligence?
“Absolutely,” he asserts. “The other people to really get inspiration from are people like directors and actors. We’ve got a full-time director and a full-time scriptwriter working on Milo – they’re here all of the time. So we ask them, ‘Is the way Milo is acting – is his distance from the camera – taking emotion away or adding it?’ I mean, ask yourself: why are there good actors and bad actors? Why does Meryl Streep consistently provide good performances while Madonna doesn’t? There’s a definable skill there: it’s poise, it’s body movement, it’s blink rate. All of things communicate emotion.”
But there is a downside to being hung up on the details; on the little things that stand out – and Molyneux is not unaware of how this obsession with making a believable world can come at the expense of the game. “People love the idea of emergence, and that is an important game mechanic to exploit. My criticism of us is that, at times, we’ve become too obsessed with it and have forgotten about the core of the game. I think you can probably see that in Fable 1, that we got a bit too obsessed with the simulation and forgot that it was a game that was fundamentally about being a hero.”
Nevertheless, the exploitation of emotional AI as gameplay remains one of Lionhead’s core ambitions – and, as such, we shouldn’t really be that surprised that Milo is the latest thing to emerge from its secret corridors. And, just as every one of Lionhead’s previous titles has shaped what comes next, the lessons learnt in making Milo a more believable character will have implications for making NPCs less robotic throughout the whole industry.