Artificial Intelligence-an evaluation of the hype, reality and potential
The Wall Street Journal’s Future of Everything Festival May 8th, 2018 in NYC provided a meaningful overview of key issues and controversies around Artificial Intelligence, Health, Learning, Food and other topics. The AI track featured several different perspectives from experts and practitioners including Garry Kasparov, futurist Amy Webb, Nicole Eagan the CEO of Darktrace and actor Jeffrey Wright, who plays the role of Arnold and his android successor Bernard in the HBO series Westworld.
Some key points:
- Concerns over AI’s potential destructive impact on jobs and on the economy are exaggerated, but capabilities are progressing at an impressive rate.
- AI and Machine Learning are the next evolution of tools that we need to learn to master. While the technologies are not new, the capabilities of self-learning systems are still nascent, and we need to be cognizant of the risks and pitfalls of how we use these technologies.
- Ethics remain a core concern, both in terms of algorithmic bias, and the use of AI in making decisions of life or death.
- Garry Kasparov sees a new opportunity for human-machine collaboration as most promising in the future, but the inertia of human nature, and resistance to change remain the most potent obstacles to adoption.
- One of the notable concerns raised by David Sigal of Two Sigma is that network effects from large business can have a chilling effect on the ability for small firms to innovate and compete
- Security is an arms race, where AI capabilities are proving increasingly helpful in discovering exploits of IoT connected devices (such as Dark Trace’s discovery that hackers had compromised a casino fish tank thermometer to gain access to high roller data).
- Actor Jeffrey Wright observed that machines lack judgement or good taste and noted how a jazz musician’s ability to turn mistakes into coherent musical ideas will always elude the capability of machines.
From left to right: Ed Maguire and Jeffrey Wright
Overall, businesses should continue to adopt Machine Learning and AI technologies with an eye toward delivering “Augmented Intelligence” to solve more and more problems.
“Deep Learning is greedy, brittle and shallow”
Gary Marcus is a professor at NYU who opened up the event. In his talk, he highlighted how the potential of AI in fiction is way ahead of where the technological progress actually is. Currently we take big data in and make statistical approximations. AI is good at object, face, speech recognition – perceptual classification problems. Deep Learning – A Critical Appraisal is his latest paper, which Wired summarized as saying “Deep Learning is greedy, brittle and shallow’. When problems get complex or unusual situations, Deep Learning classifications are inaccurate.
Marcus is skeptical of the extent to which tasks can be automated, as even straightforward tasks are subject to a lot of variability. We want conversation interfaces. The best AI still flunks 8thgrade science. Perception is only one small component of intelligence. AI hasn’t figured out common sense reasoning. He showed how machines that lack common sense – in one notable case a puppy made a poop on the floor and the Roomba started running an hour later - with predictably messy results!
Security for IoT is an AI arms race
An interview with Nicole Eagan CEO of privately held Darktrace focused on the cybersecurity challenges. There have been attacks on internet connected cappuccino makers in train stations, in order to gain access to the rail provider’s corporate network. There are early signs of using AI for attacks – there is an area of AI called generative adversarial networks (GANS) which take two neural networks which compete against each other and get smarter and smarter. There’s a generator and a discriminator network that go back and forth. This GAN format could enable an attacker to create their own neural network to fight against itself and develop a weapon that when dropped into a network could be highly sophisticated. For defense, AI uses historical data sets to classify and categorize spam or malware and maybe even predict future attacks using supervised machine learning. Self-learning or unsupervised learning is based on the principles of the human immune system. When something that’s “not self” hits the human body there’s a rapid response.
Darktrace was deployed in a casino, where the internet connected thermostat in the fish tank was used to get onto the casino’s network to search for the high roller database – they found some data and tried to pull it back through the thermostat to the cloud, but Darktrace found it. While it was difficult to put an agent on the device or create a rule prohibiting access to certain data, the immune system scanning detected the anomalous behavior on the network. There is no way to anticipate everything, but AI is the way to combat this.
Machines are smart, but they augment us, not replace us
Former World Chess Champion Garry Kasparov wrote a recent op-ed about how AI will change how humans collaborate with machines. In his view, AI is not a harbinger of dystopias. The term Augmented is better than Artificial, because this is technology that enhances our intelligence. There will be unintended consequences – but jobs don’t disappear they evolve. When people thought about chess with computers they think about solving chess as a problem. Humans inherently make mistakes. We should not compete with machines in closed systems – Go or Chess are systems where machines can dominate.
A conversation with Garry Kasparov
Kasparov noted that Claude Shannon has talked about the vision for Type A and Type B machines. Type A machines are programmable, Type B machines emulate human thinking – that’s what was thought to be the future. Today we are seeing the dawn of a Type C machine. Alpha Zero played 60 million games against a Type A machine in 4 hours, accumulating more knowledge in that time than perhaps in human history. A smartphone app is more powerful than Deep Blue 20 years ago. People mistake that machines in business are completely profitable, but there is always room for human value-add. There is a paradox that more sophisticated machines require less skills to operate. Elevators at one time had 70,000 people in the Operators Union. It took a strike in 1945 before people decided to push the elevator call buttons themselves, which had been around for over 40 years. In many cases, machine predictions are not always reliable.
Exploring the Cone of Possibility
Amy Web is a Professor at NYU School of Business and author of The Signals are Talking. In her view, because we’ve anthropomorphized computers, we have misperceptions. AI is already here but it didn’t show up how we expected it, and because of this we are missing signals. The first era of computing was tabulation, the second was programmable systems, and the third era is about systems that make decisions for us. The present modality is not what’s important; it’s the potential.
There are the “Big 9” companies: these are Tencent, Alibaba, Baidu, Amazon, IBM, Alphabet, Facebook, Microsoft and Apple. The first AI winter in the 60s occurred because of a lack of compute power. In the 90s the concept of neural nets emerged from looking at parallels in biology.
China is ahead of the US in investing in AI, and because they don’t have constraints around privacy, they are really good at facial recognition. They are working on payments through facial recognition. China is building out a future in a way that could change the balance of power. Imagine if you are jaywalking and the system publicly shames you – or the social ranking system in China – this is a radically different way of using data to control society. Because people generate data there are people whose lives are completely changed.
Network effects, not AI, a greater threat to innovation and job creation
David Sigal from Two Sigma Investments spoke on the future of jobs. New roles will require creativity, empathy and dexterity – there are capabilities that machines do not have. One of the challenges here is that human labor is taxed with payroll taxes and other requirements. Governments need to think about how we can incentivize businesses for human labor. One of the challenges of businesses with network effects is that Uber has made it very difficult to start a new car service. These dynamic favors larger businesses – they are not traditional monopolies, but the network effect and data advantage risks stifling economic growth and creativity. While Amazon, Google and Uber generate a lot of economic benefits, we do need to be concerned over the impact on creative growth.
Share it
