Skip to main contentSkip to navigationSkip to navigation
Artificial intelligence robot
DeepMind’s “generalised artificial intelligence” was able to figure out how to succeed at nearly 50 Atari computer games without any foreknowledge of how to play them. Photograph: Blutgruppe/Blutgruppe/Corbis
DeepMind’s “generalised artificial intelligence” was able to figure out how to succeed at nearly 50 Atari computer games without any foreknowledge of how to play them. Photograph: Blutgruppe/Blutgruppe/Corbis

DeepMind: 'Artificial intelligence is a tool that humans can control and direct'

This article is more than 8 years old

Co-founder of technology company insists AI is not a danger to humanity, but will help tackle lack of clean water, financial inequality and stock market risks

Fears that artificial intelligence will wipe out human beings are completely overblown, according to the co-founder of Britain’s DeepMind, who has insisted that the technology will help tackle some of the world’s biggest problems including accessing clean water, financial inequality and stock market risks.

Mustafa Suleyman, who with Demis Hassabis and Shane Legg set up the London-based machine learning company that was bought by Google in January 2014 for £400m, mounted a spirited defence of the company’s successes. He told a conference on machine learning that “artificial intelligence, AI, has arrived. This isn’t just some brief summer for this technology, and it’s not about to go away again. These are production breakthroughs.”

High-profile figures including Elon Musk, Stephen Hawking and Bill Gates have all warned that the rise of AI poses a threat to humanity – a threat that has been echoed in recent Hollywood films such as Ex Machina, The Terminator and Transcendence. Yet Suleyman insisted that AI is, and will remain, a tool that humans can control and direct, rather than a threat.

The best use for AI would be to help decisions about how to tackle some of the world’s biggest problems such as lack of access to clean water, inequality of access to food and finance, and stock market risks, he suggested.

DeepMind’s systems use neural networks and “deep learning” methods that deploy low-level transistor networks to produce high-level effects so that they can, for instance, distinguish a cat’s face from a human one – a trivial task for a human, but hard for a machine. That has been developed into “artificial general intelligence” (AGI) that can learn to solve tasks without prior programming, and have already been used to replace 60 hand-crafted systems across Google. The AGI system’s deployment into speech recognition, now used in Android phones and Google Translate, had led to the biggest overall improvement in speech recognition in 20 years, Suleyman said, with a 30% reduction in transcription error rates. Yet training the program for the task took less than five days.

Speaking to a conference on machine intelligence in London on Friday, Suleyman said that he was dismayed by the negative attitudes being shown towards AI. “It’s sad how quickly we’ve adopted to the reality and don’t acknowledge the magic and the good that these systems can bring. The narrative has gone straight from ‘isn’t it terrible that AI has been such a flop’ to ‘isn’t it terrible that AI has been such a success’.”

He said that the technology was going to be “a hugely powerful tool that we control and direct within its limits – like any tool that we have ever built … Artificial generalised intelligence is a form of intellectual horsepower – a cheap and abundant resource to solve our toughest global problems.”

Suleyman observed: “We have global information overload from overwhelming systems complexity – they’re so complex and interlinked it’s possible that the US financial crash in 2008-9 caused the Egyptian revolution [which was sparked when bread prices rose in line with wheat prices].

“But everything we have built is a product of intelligent human activity. AGI is a tool to massively amplify our ability to control the world.”

DeepMind, based by Kings Cross station in London, has developed a “generalised artificial intelligence” which was able to figure out how to succeed at nearly 50 Atari computer games without any foreknowledge of how to play them. Given inputs of just the score and the pixels on the screen, and control of the games buttons – again without any knowledge of their relevance – it was able to play as well as a human after a few hundred games. In Breakout, it played competently after 300 games – then figured out after 200 more games that the best strategy was to knock out the side bricks and let the ball bounce behind the wall: “that surprised us,” said Suleyman.

The company’s systems are now used on the Google+ photo categorisation systems, and apparently on Google’s new Photos service, to categorise and label pictures by their contents. The company is also seeking to expand that categorisation so that when there are multiple recognisable objects in a picture it can describe them all in a single coherent sentence.

But Suleyman said the idea that a machine-based artificial intelligence could take over decision and pose a threat to humans was “preposterous”.

“Any talk of a superintelligent machine vacuuming up all the knowledge in the world and then going about making its own decisions are absurd. There are engineers in this room who know how difficult it is to get any input into these systems,” Suleyman said to applause from the audience of machine intelligence specialists.

“If we fear that we won’t control them, then we should slow down their use and implementation, just like with nuclear weapons and genetic engineering” [which saw a moratorium in the 1970s].

Suleyman said he wants to make public the names of the people who sit on the company’s ethics board, which was set up at the insistence of himself and Hassabis when Google bought it. “We will [publicise the names], but that isn’t the be-all and end-all. It’s one component of the whole apparatus,” he said.

Asked what gave Google the right to choose the ethics board members without any public oversight, Suleyman replied: “That’s just what I said to Larry [Page, Google’s chief executive]. We will make more public.”

He said it had been a bold move for the 100-strong company to suggest to the much bigger buyer that there should be an ethics board at all. “Being able to put something like this on the table is a first step to being more open and helping to steward this,” he said. The company is seeking to recruit more people to its ethics board, as well as to its policy and legal teams.

Most viewed

Most viewed