On a bright, clear morning in February 2017, a motley group of high-powered individuals from the entertainment, policy and security industries gathered on the Warner Brothers lot in Burbank, California. In the middle of the 25-hectare plot, surrounded by large warehouse sets, lies a ghost town: broad streets and pavements, a library, a bank, salon, shopfronts. The buildings are sets for Hollywood films (most recently, 2017’s La La Land). It’s New York, at some point in the recent past, on the outskirts of Los Angeles. The subway stairs descend into nothing.
The group – which included the president of a major animation studio, the director of a world-famous superhero franchise, British MPs and senior military staff from both the US and UK – had gathered for a two-day private event held by British startup Improbable. The set was a nod to the theme, “Virtual Worlds” – not virtual reality, with its cumbersome headsets, but rather alternate realities: simulations.
Inside one artificial edifice, Herman Narula, Improbable’s co-founder and CEO, addressed the group. “We’re in a place today where it is actually possible to create artificial realities,” he said. “Not in some abstract sense, but genuine, living, breathing recreations of this one, powered by technology, that allow people to have totally new experiences.” Narula has a boyish face, scruffily cropped hair and stubble; he has the energy levels of a small nuclear reactor, and speaks faster than most people think.
“AI gets all the press,” he said, but, “this idea of recreating reality is going to become something in the public consciousness that’s as important, as significant, as artificial intelligence.”
Improbable has yet to reach the profile of British startups like the AI company DeepMind (now a division of Google), but its ambitions are just as lofty. It has raised more than $540 million (£421m) in funding from venture-capital firms including SoftBank, Horizons Ventures and Andreessen Horowitz. It has partnered with Google worldwide, and counts the UK and US defence departments among its clients. In May 2017, the company joined the select few British tech startups valued at over $1billion.
Improbable’s platform, SpatialOS, is designed to let anyone build massive agent-based simulations, running in the cloud: imagine Minecraft with thousands of players in the same space, or researchers creating simulated cities to model the behaviour of millions. Its ultimate goal: to create totally immersive, persistent virtual worlds, and in doing so, change how we make decisions.
Or more simply, as Narula often jokes, “Basically, we want to build the Matrix.”
Data is old news. By its nature, it tells us about things that have already happened. So to make decisions about the future, industries from engineering to biosescience still turn to simulation.
Broadly speaking, simulations are digital recreations – approximations, really – of reality. All video games are, to an extent, simulations, whether explicitly (Euro Truck Simulator 2) or implicitly (the parody Los Angeles of Grand Theft Auto V).
Large, detailed simulations require vast amounts of computing power. So rather than, for example, simulating all the traffic lights in a city, transport researchers might simulate a single junction or route. (While you can A/B test your app design, if you build a motorway in the wrong place, it’s tricky to move it later.) But that itself is still expensive and limited, particularly when studying, for example,the emergent, messy effects that occur with enough moving parts.
Imagine a London street during rush hour. There’s a collision in the road; with the lanes blocked, traffic backs up. As drivers divert, other routes start to become clogged. A few kilometres back, a queue starts to form on the motorway. Drivers, cursing under their breath, call in to work to cancel their morning meetings, overloading the nearby cellular towers. Deliveries are delayed; in the hospital, an ambulance is redirected, meaning another patient faces a longer wait for medical attention.
Right now, simulating that kind of emergent complexity is impossible, but it’s exactly what Improbable wants to be able to recreate. As Narula explains: “The really interesting things happen at scale.”
It’s a few weeks before the LA event, and Narula and Rob Whitehead, the company’s co-founder and chief technology officer, are sitting in Improbable’s London office. On the ground floor of a Farringdon mid-rise, cheap MDF desks compete to hold the most monitors. The company is growing quickly – in February, it opened a San Francisco office – and, each time I visit, the space has sprouted a new meeting pod.
On a whiteboard in one conference room, Narula and Whitehead sketch out SpatialOS. In a video game, our imaginary London street might be made up of “Entities”. “An entity represents a thing, like a noun. In a city simulation, your entities would be every traffic signal, every segment of road, every building, every pedestrian, every car,” Whitehead says. Entities have states, called “Components” (a car may have fuel; it may be On Fire, or Not). Components are governed by “Systems”, such as physics, which interact with them. (The car crashes, now it is On Fire.)
With one street and a few cars, that works fine. But add hundreds of cars, each containing a driver with their own priorities, and suddenly things get too much for an individual machine to handle. Typically, online games tackle this by limiting the number of players on each server (for example, Battlefield 1 has a limit of 64 players per game). In games with very large number of players – known as massively multiplayer online, or MMO – developers chop the play space up into different pieces and recreate each piece many times on different servers. (This is called “sharding”, an obscure reference to the early MMO Ultima Online, because: game nerds.) That’s why even though World of Warcraft might have millions of subscribers, you’ll rarely encounter more than a few dozen at any given time.
Read more: DeepMind: inside Google's super-brain
SpatialOS introduces an additional computational abstraction, called “Workers”. Simplified, rather than chopping up the simulation by physical space, Workers distribute the simulation by the type of task happening. “You might have the vehicle Worker, which would simulate the vehicles; then you would have maybe a crowd Worker, which simulates how people move around. Then you would have, say, a power Worker working out how power is distributed across the entire city,” explains Peter Lipka, Improbable’s chief operating officer. Read more: Theory claims to offer the first 'evidence' our Universe is a hologram
Crucially, if the computation is too much for one Worker to handle, or a server crashes unexpectedly, SpatialOS will dynamically bring up additional servers and distribute the load across them in real time. Narula describes this as “a giant game of musical chairs”. “As a whole, what you’ve got now is this set of different Workers all collaborating to simulate a single living, breathing city,” says Lipka. “SpatialOS is the core fabric that binds them together.”
It’s an elegant hack, based on a simple underlying notion: that even in such simulations, each part doesn’t need to know about every other part, only the things in proximity to it. Think of a murmuration of starlings, in which each bird is only aware of the small group around it, yet in large enough numbers can produce behaviours of beautiful complexity.
Before it wanted to build the Matrix, Improbable wanted to make a game. Herman Narula was born in Delhi and grew up wealthy: his father, Harpinder Singh Narula, is a billionaire construction magnate, whose company builds large infrastructure projects across India. Herman received his first computer, an Intel 486, at the age of seven. “I was wandering around the house with a DOS book,” he recalls. The family split his time between India, the US and the UK. “It would be long months with different relatives and family, and little to do, just a computer and some books.” He took naturally to programming and passed his IT GCSE four years early, later landing a place at the University of Cambridge to study computer science.
It was on a Thursday in 2012, two months before his finals, that Narula met Whitehead in a dissertation review. Growing up in Liverpool, Whitehead was an avid fan of the virtual world Second Life, where he made money as a virtual arms dealer. “That’s where I developed the entrepreneurial streak,” Whitehead says. In school, he’d build ioses games and put them up on the app store for fun.
The pair started to discuss the limitations of online worlds. “He felt the same way I did about the flaw, the big flaw,” Narula says. “Why are online games so cack?”
They started talking about what the perfect game would look like: a first-person shooter, with an enormous, endlessly complex virtual world. “Games compromise,” Whitehead says. “We wanted to have so many simulated objects in the world, tens of thousands of people in the same environment.” In short: “We didn’t want to compromise.”
The next day, the pair left Cambridge and went to Hyver Hall, the 19th-century Hertfordshire mansion owned by Narula’s family, and started coding. By the end of the year they’d formed a company and were hiring engineers. The team worked out of a barn on the property. One of the earliest hires was Lipka, who joined from Goldman Sachs, where he worked on the bank’s back-end system. A jovial former rugby player with a masters in computer science from Imperial College London, Lipka was intrigued by the daunting technical challenge. “I’ll be honest, they were a little batshit crazy,” he says. “I loved it.” Lipka’s finance background came in handy. “High-frequency trading is basically ‘speed is money,’” Lipka says. “So taking those techniques and applying them to what we were building was incredibly valuable.”
Narula’s family had expected him to join the construction business like his brothers, and weren’t happy about his decision to start a tech company. “In some ways, not going into the family business is very much an act of rebellion. It would be silly to argue that it wasn’t as a result of conflict,” says Narula. “I think a lot of it was about demonstrating that there’s a difference between your wealth and their wealth, and there’s a distinction between your achievements and their achievements. But I greatly admire my brothers and my father continues to do such amazing work.”
“Very few people, given his background, would have made the life choices that he has made,” Lipka says.
Narula’s family wealth had another advantage: not only could the company operate out of the family home, but when the company needed early funding, Narula raised £1.2 million from family and friends. “I will be honest with you: I couldn’t have done this if it wasn’t for being able to get some money from family and relatives,” Narula says. “We talk about how VCs [venture capitalists] fund innovation, but actually if you have the desire to solve an incredibly hard problem and you don’t know how you are going to solve it, you don’t have anything.”
The deeper they got into the problem, the more Improbable realised that the underlying technology could be used for more than gaming. “We started talking to epidemiologists and civil engineers,” Whitehead says. “This kind of simulation infrastructure didn’t exist.” The idea of creating a game was abandoned. Instead, they set out to create a platform so any developer could build the kind of large-scale simulations they had envisioned.
They hired Sam Kalnins, who had developed Hangouts at Google, and Eric Molitor, now Improbable’s VP of engineering, from Amazon. They moved into the London office and funding soon followed: first, from a group of influential British tech investors [Disclaimer: including WIRED editor-at-large David Rowan, who was not involved with this story]. Then, in March 2015, it secured a $20 million investment from Andreessen Horowitz, only its second European investment. “When I first met them, it was a clear that they had something special,” says Vijay Pande, a former Stanford professor who worked on computational biology and distributed simulations prior to joining Andreessen.
“[Herman] is a force of nature,” says Chris Dixon, the managing partner at Andreessen Horowitz who led the investment. “The first time I met him I was like, ‘Woah – what just happened?’ And Rob is brilliant, too.”
With investment secured, Improbable began rapidly expanding. In March 2016, it debuted a working simulation of the internet’s entire underlying infrastructure. The simulation, built with an unnamed department of the UK government, was designed to test what would happen if the web’s routing infrastructure was attacked.
At Google’s Cloud Next conference in March 2017, Narula unveiled an even bigger project: a working mock-up of the city of Cambridge, with 130,000 virtual inhabitants. It included simulations of the traffic and public-transport networks, utilities, power lines and mobiles-phones and internet systems. Narula claims the Cambridge simulation is “the largest of its kind ever created”.
“They solved a really difficult technical problem,” says Nan Boden, head of global partnerships at Google Cloud. “The first time I saw it, I said, ‘Wow, that’s some really nice computer science work there.’”
In May 2017, as WIRED went to press, the company announced it had raised $502million in funding from Softbank, valuing it at over a billion dollars.
We come in peace.”
“He’s trying to get away.”
“Prepare to die!”
It’s late January, a few weeks before the LA event, and Narula and a few Improbable employees are hanging out in the London offices of Bossa Studioses, playing Worlds Adrift. An MMO set in a world of flying pirate ships and floating islands, it is the first game to be built on SpatialOS. As such, the game is also a test bed for the technology. Most weekends, Narula drops by to test the latest build and discuss new features.
In the game, Narula, Bossa’s co-founder Henrique Olifiers and I are being hunted. Our play session is being live-streamed on Twitch; Bossa has declared a reward for whomever can find us. In the corner, there is a display with live analytics from SpatialOS. Worlds Adrift is still in Alpha, accessible to only a limited number of testers, but more than 600 players are sharing the same physical space, stretching over 1,000 square kilometres.
Worlds Adrift is unlike other MMOs. “We wanted to make a physical world, because in MMOs it has never been done,” says Olifiers. “MMOs are very much turn-based, smoothed out to look like something which is action, but is not. The server is ticking at a couple of times per second.” Chop down a tree in a game like Minecraft, and timber neatly appears in one’s inventory. In Worlds Adrift, the pieces crash realistically to the ground. Load a ship with too much weight, and it’ll start sinking. The world is also persistent, meaning if I drop my axe and come back in a year’s time, it’ll still be there. Likewise, the tree doesn’t just appear for the next player, you actually have to wait for it to grow back.
Bossa and Improbable are betting that this permanence will add a layer of immersion unseen in other games. To illustrate the point, Narula tells the story of the Saint Beyoncé. “I don’t see my brother very often. He came over and we spent four days together building this awesome ship,” he recalls. “And I blew it up. I was deeply sad that I destroyed that ship.”
“We wanted to make games where everything matters, but apparently that also makes you cry.” Persistence also adds unforeseen levels of complexity. Worlds Adrift, for example, has its own ecology. Early on in its development, the game featured two kinds of wildlife, flying manta rays and beetles. Except the beetles wouldn’t have sex. “The levels of libido were too low,” Olifiers says “All the females were in heat, but the male beetles were not copulating, so they all died.” (Its developers have since fixed the problem.)
MetaWorld, a SpatialOS app being developed by San Francisco-based HelloVR, lets two people play chess in virtual reality together. MetaWorld is also persistent – which seemed like a great idea, until players started throwing virtual chess pieces into the grass, and the Improbable team had to crawl around on their virtual hands and knees to find them.
This type of behaviour – complex results emerging from simple systems – is what gets Narula excited. One inspiration for Improbable was the space-set MMO EVE Online, players of which have developed their own political systems and stories; battles often take up thousands of combatants. (EVE Online doesn’t run on Improbable technology, but its own solution, built up over decades.)
Because SpatialOS handles most of the back end itself, Improbable says the platform will allow even small indie developers to build games previously reserved for the kind of AAA publishers with huge budgets and server infrastructures. Currently, more than half a dozen developers are building games on SpatialOS – including Klang, a new studio founded by two veterans of EVE Online. If they succeed, Narula expects that number to expand quickly. As part of Improbable’s partnership with Google, developers will be given free use of Google Cloud to test their titles in development.
But for Improbable to truly succeed, its biggest test will be outside of gaming. There, too, it’s made impressive strides. In fact, the company has already spawned its first startup. Immense Simulations was spun out of work that Improbable had been collaborating on with the UK government’s Transport Systems Catapult, to simulate fleets of self-driving cars. Using SpatialOS, Immense built a full-scale simulation of Manchester on a match day, using licensed and open-source data to accurately model traffic, population density and other granular details. “Being able to predict demand and where people are going to want to be at different times and how that affects the mobility system – from a business point of view that’s really valuable,” says Immense CEO Robin North.
Of course, identifying the technology underpinning any simulation is only half the battle. Whether it’s actually of any use comes down to designing the model of the world. If your simulated tumour or motorway system doesn’t match real-world results, it’s useless.
To help solve that problem, Improbable’s plan is for SpatialOS to feature an app-store-like platform for developers to share their models. “Now, if somebody wants to think about how Ebola might spread over a city, and there’s already a simulation of the transport networks that are being used by Transport for London for example, they can say, ‘Well, hey, let’s just plug in our Ebola model of how that might spread on top of that,’” says Narula.
It’s that vision – an app store of simulation models – Narula says, that will open the platform up not just to academics, but to any developer in their bedroom who might want to model a tumour spreading in the body. “You’ve not only made it possible to build things that couldn’t be made before, but also you’ve made it ridiculously easy and cheap for a whole bunch of people out there to attack new applications." Improbable has already discussed potential medical use cases: simulating biological systems. “This could be very interesting to build a model of a cell,” says Pande. “You can imagine jumping to a much larger scale, like an individual human body.”
“We want to start building virtual worlds for social science research,” says Ed Castronova, a professor at Indiana University who has studied video game economies. He is planning to use SpatialOS to build behavioural simulations to study, for example, the emergence of violent extremism in populations, or economic models. Researchers at Oxford Martin School are exploring using Improbable to model the UK housing market.
Simulations are not only useful for decision-making: as representations of reality, virtual worlds make useful training grounds for AI. DeepMind is using StarCraft; Microsoft is using Minecraft. The think tank OpenAI, part funded by Elon Musk, is building a library of AI environments. “Everyone talks about compute [power],” says Andrej Karpathy, a researcher at OpenAI. “These environments are the second puzzle piece.”
Virtual worlds built on SpatialOS could be particularly useful for AI research. “Multi-agent environments are much more interesting, because there are other agents in the environments. You have to learn how to co-operate, how to react. Things become much more complex very quickly.”
In Narula’s vision, each of those models could be integrated into the platform: simulations built on simulations, layer by layer. The end goal: a one-to-one virtual representation of the real world that researchers can use to run experiments – what Narula calls reality as a decision-making platform; “a ‘what if’ machine”.
Of course, Improbable’s technology is still at an early stage. And while it’s simple to model a motorway network, nobody has come close to accurately modelling human behaviour, with our illogical quirks and tempers.
But Castronova is bullish. “We’ve learned a lot about human cognition by looking at rats running around mazes,” he says. “You create the experimental environment and then you make sure you only explore questions that are appropriate to that environment.”
In recent years, Silicon Valley has become gripped by the idea that we might be living in a computer simulation. The simulation argument was popularised by a 2003 paper by the Oxford academic and AI doomsayer Nick Bostrom. Simplified, Bostrom argues that given the continued progression of computing power and technologies such as virtual reality, it’s possible that our reality is not the true reality, and instead we live in a simulation created by our own descendants sometime in the distant future. Proponents for this argument include Elon Musk; according to The New Yorker, two Valley billionaires are so convinced that they have employed researchers to explore how we could break out of such a simulation.
In LA, at the end of a day of deep, off-the-record discussions about the ethics of creating such virtual worlds, I asked Narula whether he believes that we are indeed living in such a simulation. Read more: Elon Musk: 'It's likely we're living in a simulation and Pong is proof'
“Almost definitely, right?” he said – laughing, but only half-joking. “Nick Bostrom’s argument I find quite plausible: the notion that we’re either in one, or it’s impossible to create one. I would argue it’s possible to create a virtual world – I don’t really see why it wouldn’t be. It seems like a very, very challenging engineering problem, but, at the same time, it doesn’t seem like a hard science problem.” Narula was pacing – he rarely sits still – inside a faux New York shop front where we’d retired to talk.
“The other thing to think about is if we’re not, maybe we’d like to be,” he said. “Simulations are great. We have mind-body duality with simulation, right? So you can die and it’s OK. There could be an afterlife. So I think one of the interesting things about the simulation argument is, separate from the scientific plausibility of it, it is arguably a return to discussion among atheistic circles about: is there an afterlife? What happens when I die? It’s sort of a cyberpunk version of ‘Is there a God?’”
Whether or not we would want to live inside such simulations is debatable, but Narula is aware that Matrix-style worlds are a far-off possibility. Gaming, in particular, has seen startups promise persistent virtual worlds before and failed, hard. “A lot of companies are founded around this problem,” says Hilmar Pétursson, CEO of CCP Games, the creator of EVE Online. “Usually, I have the response of crossing my fingers and saying, ‘I hope it works out.’”
Others are more optimistic. “There’s a wide-open space here for doing simulations you couldn’t even conceive of before,” says Boden.
Dixon agrees. “This will become a real field. I think we will see other companies in this space.”
“I believe that the underlying technology and the methods they are using are sound,” says Mike Sellers, who teaches game design at Indiana University. Sellers has tasked his students with building worlds on Spatial, and is impressed with the results so far. “I’ve moved past cynical and sceptical into hopeful.”
Narula recognises the sheer scale of what he’s suggesting. “The company’s called Improbable,” he grins. “It’s not called ‘extremely certain technology.’”
Together, we kicked around the implications of what living in a simulation might entail. In a simulated reality, he suggested, there could be a multiverse; by playing the simulation faster, we could time travel, perhaps see the future. “The other issue is you may not need to replicate the Universe in enormous detail to fool the human mind, so it could be that there’s no one outside there right now,” he says, and gestures to the door. Evening shadows crept along the sidewalk; the ghost town on the Warner Bros lot was indeed suspiciously quiet. “It’s kind of beautiful, in its own strange way.”
This article was originally published by WIRED UK
