The Future of Artificial Intelligence

Transcription

1 The Future of Artificial Intelligence Murray Shanahan Dept. of Computing Imperial College London

2 What Is Artificial Intelligence? Artificial intelligence (AI) is the construction of computers and robots that carry out tasks considered to require intelligence in humans In general, intelligence is the ability to make good decisions and attain goals in a large variety of environments But most AI concerns systems that are specialised for particular tasks (eg: chess, driving, ) Or it looks at aspects of intelligence such as perception, language, or learning 1

3 Economic Impact (USA) McKinsey Global Institute, 2013 Bank of America Merrill Lynch,

4 Why Today s Optimism? Success in certain sub-fields Machine learning Computer vision Increased computing power (esp. GPUs) Big data Internet of things Significant corporate investment Buoyant start-up scene 3

5 Application Areas Personalised medicine Scientific discovery Corporate decision-making Self-driving cars / trucks Personal assistants Siri, Google Home, Alexa Companion / entertainment robotics Smart cities 4

6 Short-term Issues 1 Technological unemployment Since the industrial revolution in Europe, the number of jobs taken over by machines has been offset by the creation of new kinds of jobs Often the new kinds of job required a higher level of education than those lost But could things be different this time? Could AI mean that automation will take jobs without creating new occupations? How would society cope if there was less work and more leisure? Will this trend lead to greater inequality? 5

7 Short-term Issues 2 We are increasingly dependent on IT infrastructure Finance Energy Communications Security Defence Incorporating AI technology can help humans to manage the complexity of these systems But it also makes them hard to understand, and therefore hard to control, maintain, and repair 6

8 Short-term Issues 3 Transparency Some AI systems are black boxes Trade-off between effectiveness and transparency Opacity is fine for some applications Unacceptable for others Financial, medical, military Principle: It should always be possible to supply the rationale behind any decision taken with the aid of artificial intelligence that can have a substantive impact on one or more person s lives 7

9 Short-term vs Long-term Specialist AI Short-term 10-year horizon Already here General AI Also called human-level AI Long-term 100-year horizon Uncertain 8

10 Towards General AI? 젠장! 2016: Google DeepMind s AlphaGo beats Lee Sedol, world top-ranking Go player 9

11 Go Is Difficult Go is very much more complex than chess Chess can be tackled using (1997 scale) brute computer power But this doesn t work for Go, even with 2016 computer power Good human players use their intuitive understanding of the board Google DeepMind have replicated this AlphaGo even makes creative moves 10

12 The Scope of AlphaGo AlphaGo is quite clever It exhibits a restricted kind of creativity Neither its designers nor Go masters fully understand it Machine learning techniques it uses have wide application But (like all specialised AI) it s not very clever AlphaGo can t do anything except play Go Everyday life is much more complex than Go Unlike Lee Sedol, it can t hold a conversation, play with a child, make a meal, This would require artificial general intellgence (AGI) 11

13 Long-term Risk What if we succeed in building human-level AI? Bostrom & Yudkowsky caution us to beware of perverse instantiations of very powerful optimisers According to Bostrom, the possibility of perverse instantiations is an existential threat An AI with a task as mundane as making paperclips could destroy humanity Bostrom s arguments are detailed and subtle And nothing to do with evil AIs or robot uprisings 12

14 The Road to Human-level AI Scary stories are great clickbait and help to sell newspapers But let s be clear about timescale and uncertainty Now Specialist AI with significant economic impact Human-level AI possible but unlikely THE FUTURE Human-level AI increasingly likely but still not certain Soon

15 Mitigating AI Risk Dystopian Utopian Harmful Likely Beneficial Develop strategies to avoid this Develop strategies to promote this Diagram due to Tony Prescott 14

16 Reading MIT Press, 2015 CITIC Press,