DeepMind founder James Demis spoke at the Ai Summit today, the second day of the Go Summit. Demis Hassabis gave a talk on the development of AlphaGo and what it means.

Giiso Information, founded in 2013, is a leading technology provider in the field of “artificial intelligence + information” in China, with top technologies in big data mining, intelligent semantics, knowledge mapping and other fields. At the same time, its research and development products include information robot, editing robot, writing robot and other artificial intelligence products! With its strong technical strength, the company has received angel round investment at the beginning of its establishment, and received pre-A round investment of $5 million from GSR Venture Capital in August 2015.

AlphaGo is the artificial intelligence system built by DeepMind, which was founded in London in 2010 and joined Google in 2014. According to Mr Hassabis, DeepMind was involved in Google’s Artificial intelligence “Apollo programme”, as well as forging new ways of doing science.



“DeepMind’s vision is to study what AI is and then use intelligence to solve all the problems, how we make effective suggestions to solve the problems. Ultimately, we want to build general purpose AI.” Hassabis said.

According to Hassabis, the universal learning machine built by AlphaGo has two features. One is “learning”, that is, learning raw material independently without being programmed to do so. The other is versatility, which means the same system can perform multiple tasks. “A series of algorithms and systems capable of doing a series of tasks that may never have been seen before.” , sabe si said that gm’s strong ai, and now is not the same as the weak ai and the weak ai is preset, actually in the 1990 s IBM design of chess program is also the default default of artificial intelligence, “it is through a brute force search, machine passive to accept the application, not self learning.” Hassabis said.

DeepMind wants to build reinforcement learning frameworks, which Mr Hassabis says involve “intelligent agents” in a real/virtual environment to achieve a specific goal. To accomplish this task, an agent can observe the environment, including visual, auditory, tactile, and so on. After that, set an idea in your head and then take action to implement the plan. You can do it in real time. “If we can solve this problem, we will already have strong ARTIFICIAL intelligence, in the same way that human beings observe and learn.” Hassabis said.

Mr Hassabis points to Atari Agent, an AI program DeepMind developed three years before AlphaGo, which tested hundreds of eight-bit Atari games from the 1980s on a 2600 test board. At the time, the Atari agent input only raw pixels (~30) and the goal was to use deep reinforcement learning to play Atari games well.

AlphaGo is now DeepMind’s latest artificial intelligence system. Mr Hassabis said that although AlphaGo played Go, it could also do other things.

Why is it very difficult for a computer to play Go? This is because it is so complex that even exhaustive search is hard to solve, both because it is “impossible” to write an evaluation program to determine the winner and because the search space is so large.

More difficult, in Hassabis’s view, is that go is not a game of calculation like chess, but intuition. “There is no concept of rank in Weiqi. All the pieces are the same. Weiqi is a game of defense, so you need to think about the future. When you play chess, the board is in your mind and you have to predict the future. A small chess piece can shake the overall situation, lead a launch the whole body. The ‘magic’ of Go is inspired by the apocalypse.” Hassabis explained intuition in Go.

Technically, AlphaGo uses two networks, a strategic network and a valuation network, which were published in Nature last year and have inspired many researchers to design their own AI systems.

We then tested it, hassabis recalled, and in 2016 we played AlphaGo against Lee, and AlphaGo beat Lee 4-1. “This is a moment we’ve been waiting for 10 years,” hassabis recalled. Lee Sedol sighed.

AlphaGo has attracted 280m viewers, 35,000 reports and a tenfold increase in board sales in the West. Mr Hassabis recalls the highlight of the 37th move in the second game, which made us think: you humans have underestimated the value of the fifth down shot for thousands of years. And then there was a fantastic fourth game 78, which he won. “I think this has brought new ideas to go,” Mr. Lee said. “I feel like I have found a new reason to keep playing.”

‘A lot of art is subjective,’ Mr. Hassabis said. ‘AlphaGo treated Go as an objective art, analyzing the impact of each step.’ “My definition of intuition, therefore, is a primal perception acquired through experience, unexpressed, and confirmed by action, right or wrong.” Mr Hassabis said AlphaGo could already mimic human intuition and be creative, with the ability to combine existing knowledge or unique ideas. So AlphaGo already has intuition and creativity, but those abilities are currently limited to Go.

After that, DeepMind wanted to fill in the gaps in AlphaGo’s knowledge, and then released a new version of the game, “Master,” to play online. It also won a big victory. Ke jie lamented after playing chess with Master that after thousands of years of human evolution, computers tell us that human beings are all wrong. “Wu brought revolutionary power to go in the 1930s and 1940s,” Hassabis said. I believe AlphaGo can also open a new era of Go. Chess programs are about tactics. AlphaGo is about strategy.”

“How far are we from optimal, what is the perfect game? 3000 years of playing is not enough to find the best game. AlphaGo allows us to explore these mysteries.” Hassabis said.

Beyond Go, Hassabis hopes to apply AI to a variety of fields. “Human-machine collaboration can be 1+1 > 2,” Hassabis said. “Human intelligence will be amplified by AI. Ai and AlphaGo are tools, like the Hubble telescope, that can advance human civilization.” “Countless other fields will also be bombarded with combinations,” Hassabis said. “Strong AI is also our best tool to explore, such as materials design, new drug development, and real life applications such as healthcare, smartphones, education and more.”

Giiso information, founded in 2013, is the first domestic high-tech enterprise focusing on the research and development of intelligent information processing technology and the development and operation of core software for writing robots. At the beginning of its establishment, the company received angel round investment, and in August 2015, GSR Venture Capital received $5 million pre-A round of investment.

DeepMind has already used AlphaGo successfully in data centre optimisation, with results showing a 50% reduction in power consumption.

In the end, Hassabis concludes that information overload and system redundancy are big challenges, and we want to use AI to find meta-solutions. “Our goal is to achieve AI science, or AI-assisted science, but of course AI is bound by ethics and responsibility. All in all, artificial intelligence technology can help us better explore the mysteries of the human brain.