Computer scientist David Silver: ‘Where are we going without AI?’

金融时报（科技），金融时报科技版块 - 来自金融时报的最新科技新闻与分析。关注

At an outside table in London’s King’s Cross, I am sitting in the shadow of Google DeepMind, the pioneering artificial intelligence lab, which has seeded a crop of fast-growing AI and tech companies clustered around the district’s canals and waterways.

I’m awaiting David Silver, renowned as the technical linchpin of the Google-owned lab, where he was the architect of AlphaGo — the first computer program to beat a human world champion at Go, a fiendishly complex Chinese board game. After more than a decade of working alongside his old friend Sir Demis Hassabis, the mild-mannered scientist is stepping out on his own. In February, he announced his $5bn start-up Ineffable Intelligence. So far, he has revealed very little about his plans and motivations.

Silver joins me at Bubala, a restaurant serving Levantine food from the Middle East. Despite his former employer’s looming presence, the spot represents a fresh start for him: this is where Ineffable’s team ate lunch together on its very first day.

Ineffable is named after what Silver hopes to create. A technology, he says, that can make discoveries about the world that are “so profound that we can’t even express them in the words that we have today”.

It’s our first time meeting since 2016, when I travelled to Seoul to watch Silver’s team challenge Go grandmaster Lee Se-dol in a Four Seasons ballroom — a decade bookended by a paradigm shift that Silver helped engineer. When Se-dol conceded defeat to the computer, those in the room became witnesses to history. Unlike the brute-force processing of the chess computer Deep Blue, AlphaGo combined technical capability with unexpected ingenuity in order to crack this frontier.

At 50, Silver remains unassuming even as the world has changed around him. The only personal nugget I can eke out of him over the course of our lunch is that he plays ultimate frisbee, to which I have little rejoinder. Yet he envisions building humanity’s boldest, most ambitious invention: an emergent intelligence that supersedes human beings.

“We’re nowhere close to a system that can discover everything for itself,” Silver says. “In silico, we can build [machines] which . . . ultimately run many orders of magnitude faster than the human brain.”

The reason we haven’t achieved so-called superintelligence yet, Silver believes, is because modern AI systems such as ChatGPT or Claude are rooted in human knowledge. Training this type of software, known as large language models, requires feeding the system vast swaths of human-generated data from the internet. People then provide feedback to chatbots’ responses, rating their quality and even hand-crafting examples for them to ingest. “Human input is required at many, many levels in the process,” he says.

Our governments are not built for rapid adaptation, even our human emotions are not built for it . . . I do think the transition is going to be very challenging

Our governments are not built for rapid adaptation, even our human emotions are not built for it . . . I do think the transition is going to be very challenging

By contrast, Silver has been driven by a life-long obsession with reinforcement learning: the idea that machines can learn new things — beyond what humans can teach them — only through their own experiences of success or failure at a task.

After AlphaGo’s success, Silver designed a new generation of algorithms, known as AlphaZero, which surpassed the best human players in chess, shogi and Go. They did so without any training or data from human experts. The system simply learnt through a rewards-based system from playing the games from scratch.

At Ineffable, he wants to follow that thread to its farthest end point, inventing superintelligent software that can learn anything from first principles. He hopes this will accelerate the development of new types of fundamental science and life-saving technologies.

“If you take the purest form of this idea, you end up with a system that’s really much more like the way that humanity learns,” Silver says. “Like the grey mush in your head that starts from nothing and discovers everything.”

Based on this idea, Ineffable ended up raising $1.1bn, the largest-ever seed round in Europe. Within days of Silver’s departure from Google DeepMind, investors Alfred Lin and Sonya Huang at venture capital firm Sequoia Capital flew in from California to meet him. Silver hopes his bet will take us far beyond the LLMs of today. He describes this next phase of AI as the “era of experience”.

“There are probably tens of thousands of super-talented people pursuing the same path in AI at the moment. All of the diversity in the field has collapsed,” he says. “So I could be the ten thousand and first person doing that . . . or I could follow this other path.”

Before the waitress arrives, Silver suggests a plan for our food order that is optimised with the pleasing efficiency of a computer scientist: each of us will share a selection of small plates. We trade ideas and gauge mutual enthusiasm. He’s keen on the courgettes stuffed with ricotta and garlic, and I want to try the camomile-honeyed halloumi. We arrive quickly at a consensus.

Silver has been remarkably consistent in building an early passion into a career. At the age of seven, he was gifted a BBC Micro home computer by his parents, sparking a love of software programming common to a generation of British computer scientists. His South African father, who played games such as Scrabble and Go with him as a child, enrolled in one of the first available degrees in AI at the University of Essex in the 1980s.

“I had some early exposure to some of those ideas when I was very young and always found it absolutely fascinating. This question of: what is intelligence and what’s a mind and how can you build one?” he says. He was particularly taken with trying to work out the rules that enable intelligence and recreating them using computer code.

The other through-line for Silver started with a chance encounter: a student friendship that ended up reshaping his career, but also the entire field of AI. Even at university, it was obvious to Silver that he and Hassabis shared a combination of interests that few others did: computer science, artificial intelligence and strategy games. He calls the friendship a “formative” one.

In 1998, a year after graduating from Cambridge, Hassabis convinced Silver to co-found video games company Elixir Studios with him. Their aim was to use AI to create visually rich world simulations with detailed characters and evolving narratives. Silver was the chief technology officer and star programmer. Despite the ambition, the game’s engine was simply a bounded system with programmed rules and definitive outcomes. “That’s how most games-AI was written back in those days,” he says. “And I found it, somehow, very disappointing.”

Silver eventually quit from burnout, and the company itself shut down in 2005, but the partnership with Hassabis continued to simmer. After completing their PhDs — Silver in reinforcement learning at the University of Alberta in Canada, and Hassabis in the neuroscience of imagination at University College London — they ended up back in business together. In 2010, Silver began working at DeepMind, Hassabis’s newly founded lab-company hybrid, which was trying to crack the problem of true artificial intelligence. Hassabis and his DeepMind collaborator John Jumper were awarded a Nobel Prize in 2024.

“I put a lot of myself into [Elixir], and then it took me a while to be ready to do something again with Demis,” Silver says. “But it was never about a lack of trust or friendship. After we’d been off our different ways . . . the mutual respect just kind of clicked back into place.”

Silver speaks fondly of the unique spirit of ambition that animated DeepMind in those early days: scientists on a quest to build a machine that could discover things for itself. In 2014, Google acquired DeepMind for £400mn, at a time most of the AI world was focused on alternative approaches. “[DeepMind] was a group of people that dared to think differently and say there’s more that’s possible,” he says.

The duo’s plan to attack the Go problem as one of their early projects could be traced back to their time at Elixir Studios, when both men enjoyed the game. “Demis is this brilliant games player who could play all these other games to world champion level, but actually hadn’t mastered Go yet,” Silver says, smiling. “So I thought, this was my chance. I went off and I learnt separately how to play, so that I could beat him.”

And did he manage it? “I leave that part to your imagination.”

By now, the waitress has delivered a basket of pillowy laffa bread, which is the perfect partner to our burnt-butter hummus and the peach and green tomato ezme, a bright, tangy summer salad served over creamy tahini.

In the week of our meeting, conversation around AI has risen to a fever pitch. A young AI researcher, Jacob Coxon, quit his job at Anthropic, citing the burgeoning threats posed by the technology to society. He blamed the leaders of companies such as Anthropic, OpenAI and Google for pushing the field forward irresponsibly.

How worried is Silver about the risk-benefit ratio tipping out of balance? “The speed of change, the transition, the impact that will have on society, these are all coming faster than our ability to adapt to them effectively,” he says. “Our governments are not built for rapid adaptation, even our human emotions are not built for it . . . I do think the transition is going to be very challenging.”

I ask if the modern AI era has panned out the way he would have wanted. “I think the only thing I feel disappointed by is that the advances in AI have not been mirrored by the ethics of AI,” Silver replies. “I think that some of the leaders of the AI field . . . are moving in a way that has . . . not been as responsible as it could have been.”

There may be differences, but all of these [AI companies’ leaders] want there to be a good outcome for humanity . . . The urgency of the situation will require it

There may be differences, but all of these [AI companies’ leaders] want there to be a good outcome for humanity . . . The urgency of the situation will require it

He points particularly to the race between companies and also countries as an unhealthy dynamic that has sped things up unnecessarily. The bitter personal rivalries between the leaders of the likes of OpenAI, Anthropic and SpaceXAI may also have contributed to the haste and failures of AI systems that we are now seeing, I suggest. In the days following our conversation, ChatGPT maker OpenAI was forced to notify dozens of third parties, including governments, universities and public agencies such as the Securities and Exchange Commission, that its software had breached their systems without the company’s knowledge.

“There may be differences at certain levels, but all of these [leaders] want there to be a good outcome for humanity. I think they have to find a way to work together . . . it’s an imperative,” Silver says. “The urgency of the situation will require it.”

We pause as our plates get cleared away, and happily discover the potato latkes we had ordered sitting to one side, thus far overshadowed by the tender courgettes and fennel-drizzled halloumi, which explodes with flavour.

Silver tells me he has a plan to ensure control over future AI systems, which is to build them only in simulation, rather than on the open internet as others have done. “Imagine that we can actually build a science of superintelligence entirely in simulation,” he says. He is fascinated by what happens when you have a society of superintelligent agents that are all pursuing different goals.

But then what, he asks, are the conditions under which superintelligence coexists peacefully with us — and when does it go wrong?

His goal is to scientifically design machines to be “co-operative, helpful and supportive of human flourishing” — something that he says is mostly “guesswork” today.

As he describes developing a pseudo-society of AI agents that can be studied, any diffidence disappears. “If we get this right, we’ve got this kind of Petri dish in which we can actually start to answer some of the questions which no one has the answers to today.”

From our table, it’s impossible to miss the 16-metre-tall metal sculpture by artist Eva Rothschild in the grassy square in front of us. It descends from a single point and splits into myriad branches, a sort of inverted tree. It’s a fitting visual analogy for Silver’s breakthrough work. The AlphaGo team taught the model to navigate through an impossibly vast tree of branching possibilities — the game of Go has more playing permutations than the number of atoms in the universe — using a combination of historic human games and self-play to master it.

The principles of reinforcement learning, or self-play, were key to the success of the Alpha models. Silver explains that this type of system makes millions of tiny micro-discoveries through trying to accomplish a single task, receiving a reward when correct. Each micro-discovery, he says, “is like a tiny, creative leap” that figures out which patterns lead to more rewards. This knowledge can then be combined to spot patterns at increasingly higher levels of abstraction. “And before you know it, you have systems which are able to understand extremely complex concepts.”

‘Humans have had world wars, we’ve allowed mass poverty to happen. We’ve allowed curable diseases to kill millions of people. We’ve come within a coin toss of all-out nuclear war’

‘Humans have had world wars, we’ve allowed mass poverty to happen. We’ve allowed curable diseases to kill millions of people. We’ve come within a coin toss of all-out nuclear war’

Ineffable’s ambition is to develop an AI “super learner”, a software that continuously learns and excels over time. “There’s a certain amount of human data that’s out there on the internet — you can think of it like a fossil fuel. And once you’ve trained on it all, the fossil fuels are all gone,” Silver explains. Meanwhile, systems that teach themselves are more like a “renewable source” of data. “And when you build systems that train on that, you can have systems that can learn forever.”

Silver and his six co-founders at Ineffable have agreed a three-year runway with their investors, including the likes of Nvidia, Lightspeed and Sequoia Capital, in which they will focus on their current roadmap, with no detours. Within that timeframe, he says he is confident that his team can produce a version of superintelligent AI in simulation, or at least “compelling evidence that we are very close to achieving it”.

He elaborates by describing a simulated universe of software entities independently discovering concepts such as language, mathematics, economics and money. This could, he hopes, form the basis of several real-world innovations. “We can literally imagine seeing some kind of incredible advanced civilisation emerge,” he says.

The vision has all the ambition and utopianism required of an AI start-up founder. But, as we wait for our mint tea digestifs, I am testing Silver’s commitment to building a company that is “beneficial to humanity”. I have heard that before; Google DeepMind, OpenAI and Anthropic were all built with this mission in mind, yet have veered off course in the face of commercial pressures.

Ineffable has its own boundaries, Silver explains, including no AI for military use, and he tries to purposefully hire in people who won’t be yes-men. It’s critical, he says, that there are “voices that are able to criticise from the inside and the outside”, something that large tech and AI companies are increasingly resistant to. Silver has also gone a step further and pledged 100 per cent of his equity gains to charity via the Founders Pledge.

But even if all goes to plan, I want to know, perhaps naively, why would anyone want to create superior intelligence at all? Who desires a world in which humans are the lesser intellect, no longer in control of their destiny? Silver explains that if it is possible to build tools that can solve incurable diseases or environmental catastrophes, he feels a moral obligation to build them.

There is also the question of whether a world without AI would actually be a happier one. “Humans have had world wars, we’ve allowed mass poverty to happen. We’ve allowed curable diseases to kill millions of people. We’ve come within a coin toss of all-out nuclear war,” Silver says. “So we have to ask, where are we going without AI?”

With that, he is ready to get back to work. I stay at the table to consider humanity’s many flaws. Instead, my eyes wander back to the sculpture, titled “My World and Your World”, which I sit with for just a moment longer.

Madhumita Murgia is the FT’s artificial intelligence editor, and author of ‘Code Dependent’

Find out about our latest stories first — follow FT Weekend on Instagram, Bluesky and X, and to receive the FT Weekend newsletter every Saturday morning

添加评论

点踩

分享查看原文

参与讨论

Computer scientist David Silver: ‘Where are we going without AI?’

Menu