The killer robots can wait

Which enterprises will win out in the AI race and what factors will drive their success?

In the 1950s, pioneers of AI sought to replicate human intelligence in machines. There were two primary avenues: rule-based AI composed of “if X then Y systems”, and neural networks that tried to mimic the way humans learn. The two main components required for neural networks were data and computing power. These were not strong in the 1950s, the deficit remained through the pre-silicon age, and by the 1980s neural networks had taken a back seat to rule-based systems.

Today, we have a near unimaginable abundance of computing power. It’s sobering to think that the moon landing 50 years ago was achieved with one thousandth of the on-board computing power of the first iPhone. Now we are witness to a confluence of excellence in computer science, engineering and neuroscience that has created fertile ground for huge advances in neural network-based machine and deep learning. The final ingredient is massive amounts of data from which the machines can learn. A search for the winners might start with a useful equation:

Data volume x sources of signal x available processing power = leadership

That’s especially true when to most people the real life manifestation of AI is computers that are good at games, and when AI and sci-fi have converged in popular culture.

In some ways we reached a mountaintop of reinforcement learning when DeepMind’s AlphaGo conquered the world’s most complex game. In the grandest yet simplest terms, we are watching a key development from machine learning to learning machines.

We are in the land of neural networks iterating through back propagation and making sense of unstructured and often unlabelled data. We are still between a decade (maybe) and a century (way more likely) from the development of artificial general intelligence, and thus can restrict our fear and ambition to that presented by human and corporate actors. The killer robots can wait. In the meantime, business leaders and regulators don’t need to simply understand how to discover new technology, but how to apply the right (in every imaginable sense of the word) technology in the right situation.

AI will accelerate capability gaps within and between companies, sectors and countries as the datarich become the AI-rich who will operate on accelerated cycles of innovation. More high-quality data fuels AI and produces better products, increases traceable customer interactions, and produces still more and better data. Amazon Go’s staff and checkout-free stores are the product of verified identity and supervised learning that informs smart cameras. Increasingly, the aggregated signals from all Amazon interactions will boost the relevance and efficiency of the next transaction.

There is a truth about AI and enterprise: entities that see and capture the most relevant data will win. The question is whether they will win in a:

• Narrow application — every brain scan ever made or every online ad served
• Broader application — everything anyone ever bought or watched
• Broadest application — everything anyone ever said or did and every place they went

In the West there will be multiple winners in narrow AI, like IBM in medical imaging perhaps, or Google and The Trade Desk in programmatic advertising delivery. In broader terms, particularly as it relates to commerce, Amazon is a clear leader. In business operations Microsoft may be the same, as the two Seattle giants crunch ever bigger data sets in the Amazon Web Services and Azure clouds.

The most broadly defined application — understanding human behaviour and emotional states — creates the most unease. Here, Facebook and especially Google are way, way ahead. Both companies know more about their billions of users than is known by friends, family, lovers, employers and governments.

What makes Google (and Amazon) especially interesting is that, in contrast to their walls around reach and frequency data, their cloud services businesses increasingly offer “open source to go”. Consequently, in creating new knowledge of their own, thousands of businesses are increasing the knowledge of Amazon and Google. They are not alone. In China, Alibaba and Baidu lead AI development and implementation. We remain in the age of discovery for AI, but also in a time of its practical application to problems as diverse as healthcare and transportation. Just as significantly, AI will create a landscape of ethical complexity for which we may be poorly prepared. The most commonly discussed ethical use case surrounds decision-making by autonomous vehicles. If a vehicle sees a child in its path and has enough time to swerve but not to stop, how does it trade injury to the child against possible injury to people in other vehicles or on the pavement? Ethics also play a part in managing both conscious and unconscious bias in decision-making over the provision of services like health insurance and activities like social resource allocation. In short, AI has the potential to widen the digital divide and accelerate social inequality as well as the chance to be a driver of significant positive change and societal benefit.

AI, certainly to the extent of machine learning applications, has become almost commonplace in marketing and advertising:

• The sophistication of search engines makes them unrecognisable from a decade ago. Their predictive capabilities and use of image and speech recognition have transformed the experience.
• The same underlying techniques, combined with filtering and sophisticated cluster analysis, have done the same for predictive content recommendations in media and ecommerce applications. The same is true for “personalised” advertising and customer service delivery.
• Speech recognition and natural language processing enable chatbots, so-called “conversational” AI, and sophisticated sentiment analysis.
• Areas such as dynamic pricing — which have long existed, most notably in airline ticketing — are becoming pervasive as crawlers learn more about competitive pricing and the price sensitivity of different customer cohorts.
• The efforts of Google, Facebook, Twitter and others to police the exponential proliferation of content and comment require incredible resources. They have come a long way, but not far enough.

The common factor is scale and speed at orders of magnitude that are replicable only by the tiniest subset of advertisers.

We should be careful to distinguish the automated and even the algorithmically informed, of which there are a lot, from the autonomous, of which there are few. Those that have begun to deploy autonomous systems face an unexpected challenge: explaining how it works, not just demonstrating that it works. Furthermore, autonomous systems that cross long-established silos can prove challenging to the prospective buyer. It’s equally useful to distinguish between predictive systems that look back to forecast a future state and prescriptive systems that will recommend next actions. The marketing industry and its supply chain will need to reorganise around the data of tomorrow rather than around the channel silos of today.

Transacting media is an activity that obviously will benefit from machine and deep learning. It represents an immense data set and a near infinite range of outcomes. It defies precision and predictability in the hands of humans. The machines are doing better and will do better still; even here ethical issues will arise, as it’s far from impossible that brand owner A could learn enough about brand owner B’s strategy to gain competitive intelligence. It’s even more likely that AI will create targeting decisions that discriminate in areas of age, gender, race and location, or worse, prey on addiction, depression or other afflictions.

AI and China

In Kenneth Pomeranz’s book The Great Divergence, he refers to the process through which the western world overcame premodern growth and overtook empires such as China. Consider global GDP. Up until the 1750s, China dominated and dwarfed India and the imperialist powers of Northern Europe. This only changed with the advent of the industrial revolution in western Europe in the early 19th century and decades later in post-Civil War United States. This was the catalyst for the great divergence between the West and China. Some of the most important inventions of any age — including the compass, gunpowder and printing — came from China, but industrialisation did not follow. In the last 25 years China has, of course, industrialised up to and beyond any other nation, but success in the development and application of AI can potentially aid China in creating a new generation of economic leadership. The Chinese have typically taken the long view, and many believe the last 200 years have been an anomaly.

In the grandest yet simplest terms, we are watching a key development from machine learning to learning machines

The availability of data in vast quantity alone is likely to support the Chinese endeavour. Data can be assessed based on breadth (quantity of data, how many individuals), quality (structure and labels of the data) and depth (the number of data points about each user). Surprisingly, the breadth of data from China and the United States is about the same; although the Chinese population is larger, American companies pull data from across the globe. Depth of data may tell a different story.

To paraphrase remarks made late in 2018 by Kai-Fu Lee, chairman of Sinovation Ventures: Chinese internet users channel a much larger proportion of their daily activities through their smartphones. From grocery shopping and taking out loans, to messaging and ride-hailing, these transactions and interactions take place via a smartphone — and often through a single platform like Alibaba and Tencent. Consequently, the ability of AI to observe and anticipate behaviours rises significantly.

The Chinese have developed a powerful iterative method of business innovation based on massive amounts of data in a highly competitive market where players don’t mind occupying the same space. Chinese VCs also surpassed the US in 2017 as the greatest investors in AI technology, making up 48% of global investment in AI. Additionally, the expectation of privacy is severely limited in China, allowing companies to use more sensitive data to train AI algorithms more quickly and to be more effective. As China expands its internet model and networking technologies to other countries, as it already has in Vietnam and Tanzania, it gains access to new data to fuel its algorithms. To China, the arms race in 5G is just as much about gaining access to data as it is about creating a new revenue stream.

If China is successful in expanding its internet model, it is likely that it will gain an early advantage in AI, and as a result, Chinese firms will reap the lion’s share of funding. If this happens, it will be difficult for Western companies to compete with not only the breadth of data that China has access to, but the depth and the quality as well. There is no sign yet that the ambitions of the Chinese leaders are being hampered by consumer privacy and other data use restrictions, but tolerance for the possibility of unfettered capitalism may be tested soon.

Read more from Atticus Journal Volume 25
This is an excerpt from Opportunity and hazard: 2020 and beyond 2.5 MB