AlphaGo背後的人腦 Demis Hassabis master of the new machine age

Lee Se-Dol is the world champion of Go, the ancient Chinese board game that is considered the world’s most complex. This week, the South Korean took on an artificially intelligent computer program called AlphaGo created by DeepMind, a British company owned by Google.

李世石(Lee Se-Dol)是圍棋世界冠軍。圍棋這種古老的中國棋盤遊戲，被認爲是世界上最複雜的棋類遊戲。上週，這位韓國棋手與谷歌(Google)旗下的英國公司DeepMind研發的人工智能計算機程序AlphaGo展開對弈。

In the series of five matches in Seoul, the machine is winning, taking a 2-0 lead in the contest.

在這場於首爾舉行的五局對弈中，AlphaGo目前以3：1領先。

The victories have a human mastermind in Demis Hassabis, co-founder and chief executive of DeepMind. He describes Mr Lee as the “Roger Federer of Go”, and for some the computer program’s achievement is akin to a robot taking to the lawns of Wimbledon and beating the legendary tennis champion.

AlphaGo背後的人類策劃者是DeepMind的共同創始人兼首席執行官傑米斯•哈薩比斯(Demis Hassabis)。他把李世石形容爲“圍棋界的羅傑•費德勒(Roger Federer)”，因此對一些人來說，AlphaGo的成就類似於一臺機器人站上了溫布爾登的草坪並戰勝了這位網球傳奇冠軍。

“I think it is pretty huge but, ultimately, it will be for history to judge,” says Mr Hassabis, speaking to the Financial Times from Seoul, where the matches are taking place. “Many people predicted it was at least a decade away so we’re thrilled to have achieved this milestone.” The 39-year-old has long dreamt about the victory. But his ambitions stretch beyond the Go board. His aim is to make “machines smart”.

我認爲這是一個大事件，但是，最終還是要留給歷史來評判，”哈薩比斯在首爾接受英國《金融時報》(Financial Times)採訪時稱，“很多人之前預言至少還需要10年才能實現這一成就，所以我們對於達到這一里程碑感到很興奮。”39歲的哈薩比斯長久以來一直夢想着這場勝利。但是他的雄心已經不僅限於圍棋的棋盤。他的目標是“讓機器變聰明”。

The London-born son of a Chinese-Singaporean mother and a father of Greek-Cypriot descent, Mr Hassabis is a modern polymath whose career path has seen him become a chess prodigy, master computer programmer, video games designer and neuroscientist.

哈薩比斯出生於倫敦，母親是新加坡華人，父親有希臘裔塞浦路斯人血統。他是現代版的通才博學家，既是國際象棋神童、大師級的計算機程序員，還是視頻遊戲設計師和神經學家。

These experiences led him to create DeepMind in 2010, alongside Mustafa Suleyman, a technologist and childhood friend of Mr Hassabis, and Shane Legg, whom he met when they were postgraduates studying neuroscience at University College London. The artificial intelligence group was acquired by Google for ￡400m in 2014.

這些經歷使得他在2010年與穆斯塔法•蘇萊曼(Mustafa Suleyman)和謝恩•列格(Shane Legg)一起創建了DeepMind。蘇萊曼是一位技術專家、哈薩比斯的童年好友，而列格是哈薩比斯在倫敦大學學院(University College London)讀神經學研究生時的同學。2014年，谷歌以4億英鎊的價格收購了這家人工智能公司。

“What is even more unusual about Demis is people that gifted can be difficult to mix with,” says Hermann Hauser, the computer scientist and entrepreneur. “But he’s very open, generous and humble. There is no arrogance on display.”

“讓傑米斯更加與衆不同的是，天才往往很難相處，”計算機科學家和企業家赫爾曼•豪澤(Hermann Hauser)稱，“但是他很開朗、大度又謙遜，一點都不傲慢。”

Mr Hassabis was introduced to artificial intelligence while studying computer science as an undergraduate at Cambridge university. Lecturers insisted on teaching “narrow” AI, where programmers attach “labels” to data for a computer to make sense of information.

哈薩比斯本科在劍橋大學(Cambridge University)學習計算機科學時接觸到了人工智能。當時大學講師堅持傳授有關“弱”人工智能的知識，即程序員爲數據添加“標籤”讓計算機理解信息。

Mr Hassabis was unsatisfied by this approach. He wanted to create “general” AI systems that use “unstructured” information from their surroundings to make independent decisions and predictions.

哈薩比斯對這種方式並不滿意。他希望打造“強”人工智能系統，後者能夠利用來自周圍環境的“非結構化”信息獨立決策並作出預判。

At DeepMind, engineers have created programs based on neural networks, modelled on the human brain. These systems make mistakes, but learn and improve over time. They can be set to play other games and solve other tasks, so the intelligence is general, not specific. This AI “thinks” like humans do.

在DeepMind，計算機工程師在模擬人類大腦的神經網絡的基礎上創建程序。這類系統會犯錯誤，但是會隨着時間的推移學習和提高。可以對它們進行設定，讓它們玩其他遊戲和完成其他任務，因此這種人工智能是通用而非專用的，會像人類一樣“思考”。

Games are an ideal way to test such AI programs, allowing researchers to measure performance against set goals. And Mr Hassabis is ideally placed to train the computer. A chess master by age 13 and a competitor at the Mind Sports Olympiad, he is remembered for dashing between matches to battle various competitors at once. Organisers have described him as “probably the best games player in history”.

遊戲是測試此類人工智能程序的理想方式，讓研究人員能夠將程序在遊戲中的表現與設定目標相比較。而哈薩比斯非常適合訓練計算機。作爲一名13歲就獲得國際象棋大師稱號並參加了智力奧運會(Mind Sports Olympiad)的選手，他因在賽場間奔跑、同時與不同選手對弈而被人銘記。組織者認爲他“或許是史上最佳選手”。

Mr Hassabis enjoys games filled with human randomness. He has won poker tournaments and says he enjoys the game because players can make all the right moves and still lose. He likes Diplomacy, a fraught game with loose rules, where players need to negotiate deals, forge alliances and backstab each other to secure world domination.

哈薩比斯喜歡玩人性隨機性強的遊戲。他贏過撲克錦標賽，並表示他喜歡這種遊戲是因爲選手們可能每步都正確，但仍會輸掉比賽。他喜歡玩《外交》(Diplomacy)這款有着鬆散規則卻精彩紛呈的遊戲，在這款遊戲中，選手們爲了稱霸世界，需要討價還價、締結聯盟、互相背後捅刀子。

Go is the “holy grail” for AI. The game originated 2,500 years ago in China, is played by 40m people worldwide and has 1,000 professional players.

圍棋是人工智能的“聖盃”。圍棋在2500年前起源於中國，如今全世界有4000萬人下圍棋，有1000名專業選手。

“I know how to play Go well enough to be able to appreciate its beauty,” Mr Hassabis says. “But it is not one of the games I’m strong at, so I’ve not actually played AlphaGo myself as it surpassed my ability almost from the beginning.”

哈薩比斯表示：“我的圍棋水平足以讓我欣賞它的美。但圍棋不是我的強項，因此我沒有親自與AlphaGo對弈過，因爲幾乎從一開始我就不是對手。”

Computers have long “solved” other games like backgammon and draughts. In 1997, IBM’s Deep Blue supercomputer beat Garry Kasparov, the then world chess champion. With Deep Blue, programmers built a system that tried to analyse every outcome of every possible move. But Go is far more complex than chess. There are more possible configurations on a Go board than atoms in the universe. This is too much information for even the most powerful supercomputer to process. Beating the best human player required an unprecedented technological breakthrough.

計算機早就“解決了”諸如步步高和跳棋之類的其他遊戲。1997年，IBM的“深藍”(Deep Blue)超級計算機擊敗了當時的國際象棋世界冠軍加里•卡斯帕羅夫(Garry Kasparov)。程序員用深藍打造了一個試圖分析每一種可能走法的所有結果的系統。但圍棋要遠比國際象棋複雜得多。圍棋的棋局變數比宇宙中的原子數量還要多。即便是最強大的超級計算機也無法處理這麼多的信息。擊敗最強的人類選手需要史無前例的技術突破。

That moment came on Wednesday when, after three-and-a-half hours play, Mr Lee conceded to AlphaGo. The human champion was in “shock” after the loss. The next day the computer won again. The third match begins this weekend. Though marvelling at this achievement, Mr Hauser warns that progress in other fields, such as robotics, is some way off.

突破的時刻在上週三來臨——在3個半小時的對弈之後，李世石向AlphaGo認輸。這位人類冠軍棋手在輸棋之後感到“震驚”。第二天AlphaGo再次獲勝。儘管對這一成績感到驚歎，但哈薩比斯警告稱，機器人技術等其他領域還有很長的路要走。

“One of the curiosities of the phenomenal progress we’re making with AI is that it looks as though we have a world champion at Go, but we don’t have a computer that can physically move the Go pieces,” he says. Mr Federer will not face a similar challenge just yet.

他說：“我們在人工智能領域取得的重大進展的一個古怪之處在於，看起來我們有了一個名叫AlphaGo的世界冠軍，但我們還沒有一臺能夠在實體棋盤上落子的計算機。”費德勒目前還不會面臨類似的挑戰。

For Mr Hassabis, creating machines that beat humans in games is just a testing ground before unleashing DeepMind’s technology on “real world challenges like making smartphone assistants smarter, and further in the future, using it to help scientists solve some of society’s most pressing problems in healthcare and other areas”.

對哈薩比斯來說，創造在遊戲中擊敗人類的機器只是個試驗，是爲了以後利用DeepMind的技術，“解決讓智能手機助手更智能等真實世界的挑戰，並在將來，利用這種技術幫助科學家們在醫療和其他領域解決一些最爲緊迫的社會問題”。