Contents 3

5

7

15

22

26

| Artificial IntelligenceTitle of the document |

• Artificial (AI) is a broad concept which has been around since the 1950s. Ever since the term caught on, the vague definition allowed for theoretical extremes to surface. Some, like , foresee a limitless positive future. Others, like Stephen Hawking and , predict general to be humanity’s last invention. • In this white paper we look at what artificial intelligence is and narrow this down to where we believe value lies. We then use a business model overlay to make a judgement about the potential for disruption.

Artificial Intelligence |

4 | Artificial Intelligence

Artificial intelligence is best summarized as boom, bust, boom, bust, boom. In order to avoid the next bust, it is crucially important to manage expectations. Narrow AI is here to stay, all other forms are likely still very far away.

Artificial Intelligence |

There are many different forms of artificial intelligence and the field of study is very broad. In the end, AI is all about statistics. We argue in this paper that focus should be on artificial narrow intelligence instead of artificial general - or super - intelligence. Narrow intelligence, which focuses on optimizing one specific task, is already being used today and advances in data and computing power in combination with a stronger focus on specific applications are important improvements in comparison with earlier AI boom periods.

Within the artificial intelligence theme we focus mainly on developments in . The potential for this area to grow is very large, given the most recent advances in algorithms and data. Market size ranges from USD mid-teen-billions to USD one hundred and fifty billion by 2025. The vague definition of artificial intelligence leaves room to inflate estimates by means of including or even parts of the car industry. We do not consider the exact size of the potential market to be of interest, we rather focus on how expectations are developing. Judging from that perspective it is clear that AI is hot again.

An issue that emerged during past boom periods is that of timing. It has been proven to be very hard to make good predictions in terms of progress in AI. There is a large probability current predictions are wrong again. We draw the conclusion that, based on the required computing power and current data limits, it will become very hard to live up to the high expectations in the market currently. This statement should not be confused with downplaying the trend though. Narrow artificial intelligence is used today already and it is impacting the way we diagnose, interact and optimize. We believe in incremental progress from this base onwards. We do not believe in theoretical extremes, however, because the technical requirements for such scenarios are simply non-existent currently nor likely in the coming decade.

We build on previous publications by Steef Bergakker that discuss the impact of business models on disruption and integration of new technology. Companies can be categorized as either value chains, value shops or value networks. We argue value chains will likely integrate artificial intelligence into their current processes in order to make these more efficient. Ocado, the online retailer in the UK, is a good example of this, as the company uses AI to optimize all of its logistics.

Value shops on the other hand are the battlefield for AI. Value shops represent specialist goods and services. Often, AI solutions also provide these specialized products, such as translation services. Progress in natural language processing, for example, is a direct threat to today’s translation services. Many more examples of specialized services and goods can be thought of to be replaced by AI in the future. Whereas we argue AI to be sustaining innovation for value chains, we see it as potentially disruptive for value shops. The final business model, value networks, is the holy grail in terms of disruption potential and ‘winner takes all’ outcomes. Current networking companies, like , Facebook and Amazon, can integrate artificial intelligence into their current offering, but we think AI has potential to assist in the creation of new value networks, thereby potentially replacing the current ones.

Artificial Intelligence

Artificial intelligence is a hot topic again, just as it was twice before since the 1950s. But this time it is different…at least that’s what we are being told. What is not different from previous hype cycles though, are the discussions around theoretically extreme scenarios and the claim that this time it is different.

Artificial Intelligence |

In order to make a judgement about the current developments around artificial intelligence, it is important to compare definitions and to narrow these down to what we believe to be core trends. In this part we will discuss what artificial intelligence is, how it was formed throughout history, what we expect in terms of trends and whether this is indeed the end of humanity. We then take this discussion as a base for the next chapters where we discuss the market potential as well as a business model filtering approach.

In 1956 a group of computer scientists came together at a conference organized by Dartmouth College. John McCarthy proposed “a research project on artificial intelligence”, which coined the term and described the project as follows: “an attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves. We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer.” This is an extremely broad working description and it will not come as a surprise that this “summer” turned out to become several decades and the group of “carefully selected scientists” morphed into a new field of study within computer science and statistics that has attracted, and still attracts, thousands of researchers globally.

A more formal definition of artificial intelligence is provided by J. Nilsson1, who defines it as: “that activity devoted to making machines intelligent, and intelligence is that quality that enables an entity to function appropriately and with foresight in its environment.” The interesting part of this definition is that it does not define intelligence as human intelligence per se. Given human intelligence is currently the natural choice, it can be used as a benchmark, but in certain areas benchmarking is done versus systems that exceed human intelligence already (like logistical scheduling and image recognition), at least in speed.

Still, a definition like the one suggested by Nilsson leaves a lot of room for discussion and entails a wide variety of applications. Figure 1 shows a schematic for the field of artificial intelligence. Planning, robotics, natural language processing and many other disciplines can be classified under the AI umbrella. In this paper we will mostly focus on machine learning and narrow this down further to reinforcement learning where applications such as and neural networks are starting to develop. In essence, both those who proclaim AI to be the end of humanity and those who argue AI will provide us with the required tools to grow productivity and cure diseases, focus on machine learning as this is the main area with inherent disruptive capabilities.

1 The quest for artificial intelligence, 2010 Artificial Intelligence

Figure 1 | Artificial intelligence classification

Source: Nazre, Garg, 2015

The fields above can be separated into four tasks that define AI2. These are systems that think like humans (use neural networks and cognitive architectures), systems that act like humans (pass the Turing test3 through natural langue processing and social intelligence), systems that think rationally (optimization, deduction, reasoning and problem solving) and systems that act rationally (intelligent agents that use for example planning, learning, decision making). Replace ‘humans’ by ‘the most intelligent alternative’ and we can combined this line of thinking with Nilsson’s definition.

Now that we know what AI is, we can discuss the different forms of AI, as there are plenty of those as well. Often, a separation is made between weak AI and strong AI, where strong refers to everything a human can do and weak, naturally, refers to specific tasks. Another, in our view better suited, specification of AI is the following: - Artificial narrow intelligence - Artificial general intelligence - Artificial super intelligence Narrow intelligence is used to optimize one certain task or specializes in one specific area. An example is playing chess, or arranging timelines on social media platforms according to your interests. General intelligence would match everything a human can do and super intelligence exceeds general intelligence in that it is superior to the most intelligent benchmark. General intelligence is often seen as the holy grail, super intelligence as the feared and unknown future, but we argue the focus from an investment perspective should be on artificial narrow intelligence.

2 Russell, Norvig, Artificial intelligence: a modern approach, 2009 3 The requires that a human cannot distinguish machine interaction from that of another human. Artificial Intelligence |

Since AI is such a broad and vaguely defined topic, an “issue” emerges as observed by McCarthy. He argues that “as soon as it works, no one calls it AI anymore”. This can be clearly observed today as well. No one is discussing or arguing against the algorithms that define search engines like Google, suggest movies, series and music based on your search and purchasing history or virtual assistants like Siri (Apple), Alexa (Amazon) and Cortana (Microsoft). That is AI in its core, but the view of many people is blurred by Hollywood that has brought malicious robots, super-intelligence and other apocalyptic scenarios to the discussion. This, in essence, undermines the power and impact of artificial intelligence in our current society already. Although the boundaries are perhaps a bit unclear, it is important to understand that the core objective of AI is to automate, replicate and optimize intelligent behavior. And it is used already today.

The reason why we are discussing artificial intelligence for the third time since the 1950s is the breakthrough technology developed in 2012 that is related to machine learning and uses Nvidia’s progress in graphical processing units (GPUs). It was discovered that the speed of parallel computing, as performed by a GPU, would enable an increase in performance versus traditional central computing units (CPUs) by a factor 10-100 (depending on test conditions). The increased computing power opened up new possibilities in machine learning, namely to have an algorithm writing its own code instead of a human programmer writing it line by line. This subgroup of artificial intelligence consists of three areas as depicted in figure 1.  Supervised learning refers to the situation where experts have already labeled data, and statistical optimization is used to run an analysis on the labeled groups and extrapolate that knowledge to another set of data. An example would be a group of pictures where dogs have been identified versus birds. The computer will then group all pictures with dogs and birds and can run additional statistics such as the likelihood of a picture showing a dog versus showing a bird et cetera. Email spam detection also works with supervised learning protocols.  In unsupervised machine learning this labelling is not done and the algorithms search for correlations and commonalities in the data. They could for instance group pictures based on the fact that there is a tail or paws versus feathers and wings. The algorithm would not know it is specifying dogs and birds, but it does know commonalities that define the pictures.  The final form of machine learning is reinforcement learning, where data is not labeled, but instead of grouping based on traits, the grouping is done based on what the end-user requires. To come back to the same example of birds and dogs, the end user would comment on a suggestion to add a picture of a dog to the dog group and would correct a bird being placed in the dog group. The algorithm then uses this feedback to learn and (re)write code. It could be defining traits, just as unsupervised learning, and at the same time it is able to add meaning to those traits, like in supervised learning.

Within reinforcement learning, most attention goes to deep learning and neural networks. Whereas the latter try to mimic the design of neuron flows in the brain, deep learning is based on a set hierarchy in the representation of data. The latest innovation is a combination of the two, ‘dynamic program generation’ or ‘deep neural networks’. This essentially uses the best of both the human world (in placing weight on decision criteria) and machines (in structuring data). An example where this technique has been used with great results is a setup to replicate the Bose-Einstein condensate, which

Artificial Intelligence

tries to come as close as possible to a temperature of absolute zero (-273.15 degrees Celsius). In 2001, a group of scientists received the Nobel prize for research in this area that uses lasers to cool an object as much as possible. A deep neural network configuration was used to program the lasers in such a way that the only objective function for the AI was to get as close to absolute zero as possible and the temperature results would be active feedback to re-program the lasers. What took the group of scientists years to accomplish was replicated within one hour by the AI. The interesting part is that the computer scientists didn’t know exactly how the coding of the AI developed, because it differed from the technique used by the researchers in 2001.

Figure 2 shows common applications where a form of machine learning is already in place based on deep learning technology. Dynamic program generation is a newcomer in the reinforcement learning bucket. This extremely powerful approach has only been used with success by a company called Viv, which was recently acquired by Samsung. Dynamic program generation is able to write new lines of code by combining clusters of required inputs in a matter of milliseconds. The program essentially writes itself and then learns from feedback to conclude whether or not the job was completed in a satisfactory way.

Figure 2 | Deep learning applications currently in use

Source: Nvidia, 2016

A mistake that is often made is to consider the self- learning capabilities of deep neural networks as a form of general intelligence. Although extremely impressive, the code that is developed based on a search algorithm does not allow that same program to drive a car, and programs that learn to drive cars are not going to optimize your agenda or order pizza. The systems developed are all forms of artificial narrow intelligence. They are built with the sole purpose of optimizing a particular task.

There are also technological limitations to what an algorithm can do. Purely based on computing power (which does not necessarily make it smart), it is estimated that a computer needs to be able to make 10 quadrillion calculations per second, which is equivalent to what our brains provide with a 20 watts power input. The Chinese Sunway Taihulight was able to produce 93 quadrillion calculations per

Artificial Intelligence |

second, but the USD 275 million machine takes up 1000 square meters of space and is using 25 megawatts of power.

Ray Kurzweil estimates we can reach human-level computing power for USD 1,000 by 2025 as we shown in figure 3. At that point in time, it is theoretically possible to have the computing capabilities of an average human being deployed in many devices. By 2040-2050 artificial intelligence applications could provide the combined computing power of all humans as predicted by Kurzweil. Although the graph is based on work from 2005, the underlying assumptions are still valid, since it builds on Moore’s law, which is still applicable today. We think advances in cloud computing contribute positively, as it is not required for every person to own sufficient computing power themselves; they merely need access to it. But there is more to general intelligence than computing power.

Figure 3 | How much computing power can USD 1,000 buy?

Source: Kurzweil, 2005

Artificial intelligence is currently able to perform tasks that requires thinking, but it has not yet mastered doing what humans do without thinking. Kahneman4 classifies two approaches to thinking and decision making, namely system one and system two. System one refers to decision making in auto-pilot mode while system two refers to decision making by explicitly thinking

4 Kahneman, Thinking, fast and slow, 2011 Artificial Intelligence

and weighing arguments. System one is 220,000 times faster than system two. We tend to use system one to make quick snap-shot decisions while we use system two for making reasoned decisions.

Artificial intelligence would speed up the process of making system-two decisions (simply because processing speed can be higher for computers versus the human brain), but the forces that underlie decision making in system-one situations are not as clearly defined or understood as are those for system-two decision making. Humans have five senses: sight, hearing, taste, smell and touch. Of these five senses, AI is currently able to beat human intelligence on sight (image recognition) and sound (natural langue processing). In order to fully understand humans, though, AI needs to grasp the behavior resulting from taste, smell and touch.

The consequence of not fully understanding what drives human decision making is that it could make sense for the AI to do X, while humans actually proceed by doing Y. This can be linked to emotional decision making, which is currently not understood either. Steering of thought and emotions is possible to a certain extent (selectively showing news is an example of it), but as long as we cannot solve questions around emotions, sense and system one thinking, it is not possible to introduce human-like general artificial intelligence. What we would be left with in that case is a computer that is able to compute much faster than humans, without emotional intelligence. This scenario has been qualified as the end of humanity.

| Figure 4 shows some interesting predictions made by AI researchers in the past. The most fundamental limits that prevented these predictions to come true were limited computer power, the lack of statistical work on knowledge and reasoning and the lack of understanding the brains’ functioning. We came across an interesting quote on that latter limitation5: “Ask most people if they want a brain like a computer and they'd probably jump at the chance. But look at the kind of work scientists have been doing over the last couple of decades and you'll find many of them have been trying hard to make their computers more like brains!”

All these limitations resulted in two so-called AI winters, where funding and interest in the topic disappeared after being hyped earlier. The first AI winter lasted from 1974 until 1980. In the 1980s the so-called expert systems entered the discussion which would solve problems specific to logical rules derived from expert inputs. However, overly optimistic predictions about the capabilities of these expert systems resulted in disappointment, which lead to the second AI winter that lasted from 1987 until 1993. Afterwards, some of the initial milestones were finally reached, with Deep Blue winning a computer chess match against world champion Garry Kasparov in 1997, Roomba autonomous cleaners in 2002, by Google in 2008, IBM’s beating humans in the Jeopardy quiz show, Deep Mind’s Alpha Go beating the world champion in the extremely complex ancient game of GO and IBM’s Watson graduating as oncologist. These milestones led to the third spike in interest for AI as we currently observe.

5 Chris Woodford, How stuff works, 2016 Artificial Intelligence |

Figure 4 | AI predictions in history

Source: Wikipedia, MIRI, 2016

After the previous boom-bust periods in AI, you could ask yourself why the discussion this time would be different from that during previous hypes. There are a couple of good reasons why there are more opportunities currently: - Increased computing power and resources - Availability of data - Better algorithms - Narrowing of use cases

Data availability is critical for machine learning. If there is not enough data, the machine will have no resources to learn from. There is a substantial difference between the amount of data today and that during the previous AI hype cycles. We believe another important reason for the ‘this time is different’ claim is the last one in the list above. Although the discussion about general artificial intelligence rises again, as well as the discussions about super artificial intelligence that is going to destroy humanity, we believe the narrowing of use cases is an important difference to past endeavors. Practitioners that ‘get it’ focus on specific use cases and train their systems from that base onwards. Amazon’s Echo solves very specific personal assistance tasks instead of aiming to know the answer to all questions. Nvidia’s algorithms used for autonomous driving modes purely focus on very specific tasks. Algorithms that read road signs differ from those used for lane- keeping and path prediction. In this way it is easier to use deep learning via active feedback and the resulting code becomes more rigid.

An important consideration is that current narrow AI applications will eventually also run at maximum capacity. If you want to do more with AI, it is important to grow computational power accordingly. In order to live up to today’s expectations related to machine learning, it is required to have pure data sets and beat Moore’s law in terms of technological progress. Although not completely unlikely, it puts pressure on current developments in AI and will, inevitably, lead to new disappointments and perhaps a third AI winter. Not for a lack of progress, but rather for an abundance of theoretical applications.

Artificial Intelligence

Artificial intelligence is not new, but it wasn’t practically relevant until recently. AI is used already in many applications ranging from search optimization to personal assistants. Predictions on potential market size have a wide range though and history shows prediction accuracy is low. Managing expectations is key.

Artificial Intelligence |

In this chapter we will go deeper into the market potential of machine learning. Although multiple areas of the AI industry have been present for a while, parts of the machine learning branch have developed only recently thanks to advances in computing power and brain science. Deep learning, artificial neural networks and combinations of these disciplines have given way to interesting new applications. In this chapter we will focus on the market potential of machine learning as we believe this should be the focus from an investment perspective.

There is a wide range of estimates on the potential market size of artificial intelligence. The issue with most estimates is that it is not clear which part of the AI market they represent. Often, the estimates are inflated because robotics are included in the calculations. Research based on data from Tractica and Bank of America Merrill Lynch provides the cleanest estimate in our opinion. Figure 5 provides an overview of their market estimates. Bank of America Merrill Lynch divides this USD 127 billion market opportunity by 2025 into software, hardware and services. It will not be a surprise that the services layer is expected to grow fastest, followed by hardware.

Figure 5 | AI revenue per segment, world markets 2016-2025

Source: Bank of America, Merrill Lynch, 2016

Although the classification of many market analysts differs from the one we use in figure 1, it can be concluded that half of all efforts currently go into deep learning. Neural networks and deep neural networks represent about ten percent of the AI pie currently. Supervised and unsupervised machine learning make up roughly ten percent as well and the other thirty percent is divided amongst the other disciplines with a focus on image recognition and natural langue processing, as shown in figure 6. In our definition, deep learning is part of machine learning. In figure 6 it is shown as a separate technology silo though. Deep learning capabilities are expected to grow fastest in the coming years.

Artificial Intelligence

Figure 6 | AI revenue by silo, 2015 (circle) and projected growth until 2025

Source: Tractica, 2015, Bank of America Merrill Lynch, 2016, Robeco

On the hardware side there are two fast-growing areas. Storage hardware is expected to grow from USD 336 million currently to USD 20 billion by 2025. Computing power is expected to grow from USD 126 million to USD 70 billion by 2025. The reason why this is growing fast is that parallel computing via the GPU has increased speed versus serial computing through a CPU.

Intel and AMD are the main companies developing central processing units (CPUs), which are set up in such a way that problems are broken into a series of discrete instructions that are executed sequentially by the central processing unit. Companies like Nvidia developed a process in which there is no reliance anymore on sequential processing, rather parallel processes are used. This speeds up the computational process and this has sparked use cases for artificial intelligence and deep learning/ neural networks, given the enormous amount of computations required to run these applications. Intel tries to compete with Nvidia by means of their Knights Mill product which specifically targets the deep learning market, but the advantage of Nvidia is large.

We expect to see more companies develop their own GPU for specific tasks (like Amazon and Baidu are currently doing). We expect cloud solutions to be offered soon as well, where the pure computing power will be performed in the cloud. Companies like Amazon and Microsoft are positioned well to benefit from that trend.

Artificial Intelligence |

Most benefits of artificial intelligence are to be realized in the industrial end markets rather than on the consumer side. Many applications focus on data optimization and specific expert tasks employed at industrials. Consumer products mainly focus on personal assistants, but there is little need for consumers to use specific data optimization algorithms.

Figure 7 shows another perspective on the end market, namely specific industry use cases. Most AI effort goes into advertisement, investments, retail and media. Almost all of the applications within these areas focus on the optimization of certain objectives. For instance in advertising, AI is used to generate recommendations based on past search results and other people’s inputs and price advertising spots accordingly. Other growth areas for search engines are speech to text and text to speech. Especially within Asia, natural language processing is the key area of focus at this point in time due to the complexity of written queries as a result of using characters instead of the alphabet. The potential impact is very large because more people can start interacting through their devices when they don’t have to bother with written text anymore.

In terms of market expectations, Goldman Sachs estimates that by 2025 the biggest impact will be on energy (which they estimate at USD 140bn). Retail and healthcare are estimated to save up to USD 54bn per annum and the financial services sector is expected to represent USD 34-43bn in annual cost savings and new revenue. The final sector that can benefit from AI, according to Goldman, is agriculture, where a USD 20bn addressable market is predicted by 2025. Use cases all revolve around capturing and interpreting data for optimization, in other words; statistics on steroids.

Figure 7 | Artificial intelligence revenue end markets 2015

Source: Tractica

It is estimated that 32 percent of AI applications and research is initiated from the US. Western-Europe is estimated to contribute about 23 percent to AI research. Interestingly, 33 percent of development comes from Asia Pacific. Including the Latin American markets as well as Africa, emerging markets make up about 43 percent of the AI market. It is often argued that emerging markets are the big losers of artificial intelligence because much of the low-income labor will be replaced by AI and robotization, while these countries often lack the infrastructure required to benefit from AI. However,

Artificial Intelligence

we would argue that emerging markets can benefit considerably because that lack of infrastructure enables enormous leaps in technology. Moreover, not all low cost labor will be replaced because the cost benefit analysis in early years will lean towards manual labor rather than automation.

When looking at the type of end markets and use cases for AI in Asia, one could argue that the focus is more on the consumer side, which is different from developed markets. Tencent, Baidu and Alibaba spend a considerable amount of money on AI in order to make their consumer platforms better. Personal assistants are likely to see faster penetration in Asia than in Western markets. This holds especially in Japan, where the government is pushing hard to work on AI and robotics in order to cope with an aging population; an issue that China also needs to overcome soon.

Narrow artificial intelligence and machine learning are real. It is used today and we already interact with it in many day-to-day processes. Yet, the theoretical examples that are currently being discussed often ascribe, in our opinion, too much power to AI. Two important hurdles need to be taken before this technology can indeed fulfill all theoretically envisioned roles.  First of all, the available computing power needs to increase dramatically. Building one supercomputer that can perform super human computations is not enough. That technology needs to become embedded in standard equipment (instead of only available to a hand full of researchers) and needs to be affordable. Limited capacity to transfer data requires data filtering on the device before it is repackaged and sent to data processing centers. The required processing improvements imply stretching the boundaries of Moore’s law. Even if it were possible to do this in terms of technology, it is still questionable if it can be done profitably. Current theoretical use cases for AI don’t only require Moore’s law to be sustainable, they also require technology to accelerate faster than Moore’s law, which is quite a challenge.  Besides pure computing power, there is another hurdle to overcome. Machine learning is based on statistics and therefore depends on data. It learns from and the lack of data has prevented progress in the past. However, if the wrong data is used for analysis, the wrong outcome is generated too. A hurdle now presents itself in terms of how clean the underlying data is from which machine learning is taught. It might not be an issue when personal assistants reply incorrect answers to questions asked by the owner, but it is crucial self-driving cars are trained on correct data and use the latest available, clean inputs for route optimization. Common data issues are: noise, systematic errors, outliers, measurement issues, duplications or missing data, selection bias, ambiguous labeling and ambiguous linking. Many of these issues are well known in the field of statistics, but there is no consensus in terms of how to correct for each of the issues. It will take time and a lot of human thinking/training to ensure correct data. However, there is a lack of skilled data analysts which could slow down progress or influence which industries will benefit from AI first, as depicted in figure 8.

Artificial Intelligence |

Figure 8 | Top 10 industries hiring big data experts

Source: Forbes, 2016

Judging from figure 9, neither experts nor non-experts have been able to accurately predict developments in AI in the past. Most predictions estimate artificial general intelligence within 15 to 25 years starting from the date of the prediction. Consensus at the moment seems to lie between 2030 and 2040. As we argued before, we don’t consider artificial general intelligence to be the most important focus point from an investment perspective, and rather discuss narrow intelligence applications that are already having an impact today. But we believe showing the graph on predictions is a good way to calm down the hype.

A lot of people are working to accomplish ever-better applications for artificial intelligence and a lot of investments are flowing into the field of study, especially related to machine learning. But that is no guarantee predictions come true! Usually the impact of new technology is overstated in the short run, but understated in the long term. Because the bar is set much too high for artificial general intelligence, it blurs what has been accomplished from a narrow intelligence perspective. The risk is that people/investors become disappointed with the progress in the short run, which could result in another AI winter.

Artificial Intelligence

Figure 9 | Median estimate for human-level AI graphed against date of prediction

Source: MIRI, 2012

Artificial Intelligence |

Applying a business model overlay allows us to focus on areas that have truly disruptive potential versus those places where AI will simply be integrated into the day-to-day business.

Artificial Intelligence

Discussing market potential as we did in the previous chapter provides an overview of which areas could eventually grow and how big the pie is estimated to become. By implementing a business model overlay, we hope to provide the reader with a blueprint that can be used to further specify the potential when applied to specific companies. We build on the work by Steef Bergakker6 who describes three of the most important generic business models: - Value chain - Value shop - Value network

We believe it is helpful to categorize companies into one of the above mentioned generic business models. As discussed in the paper by Bergakker, companies can use combinations of business models, but often a dominant model can be identified. Once the business model overlay is used, it becomes possible to assess the likelihood of disruption as a result of new technology. Figure 10 shows the main characteristics of value chains, value shops and value networks. A value network is the most important breeding ground In terms of possibilities for disruption.

Figure 10 | Value chains, value shops and networks - the main characteristics

Source: Robeco, 2016

6 Steef Bergakker, Business model disruption, 2016 Artificial Intelligence |

Value chain companies are all about making stuff and getting it from their warehouses to their clients. The impact of AI on companies in a value chain is not disruptive in our view. Artificial intelligence is not going to replace the process of making stuff, rather it is going to improve that process in order to make the business model more efficient. A good example of this would be the UK online grocery company Ocado. In order to run their warehouses, pick items and schedule distribution routes, the company uses optimization tools resulting from machine learning applications. There are no human schedulers involved in warehouse logistics optimization anymore. The AI deployed by Ocado goes even further than not including human schedulers anymore. Human schedulers can simply not grasp what the logistics optimization software is doing due to the complexity and multi-layered approach of the AI software.

We think value chains will be impacted by AI, in the sense that they have to invest in it in order to remain competitive. We do not expect artificial intelligence to replace these processes. We like the analogy of IBM which describes the future of AI as IA, intelligence augmentation. Augmenting differs substantially from replacing.

Value shops represent companies that fulfill expert tasks. Companies that work on artificial intelligence solutions are represented in this category as well. Companies like Viv.ai, X.ai, sense.ly and Skytree are all examples of expert companies that offer a specific solution through the usage of artificial intelligence. But also companies like Intel and Nvidia, that deliver expert solutions that enable machine learning, are value shops.

Besides companies that offer specific solutions for tasks that can be solved via artificial intelligence, other expert services are also represented in the value shop business model. Artificial intelligence has a large potential to disrupt. Since value shop companies sell solutions for specific issues, they are at risk of being replaced by artificial intelligence programs. For example, a company that offers translation services for multinationals is a typical example of a value shop, but that business is at risk of being replaced by natural language processing solutions created through artificial intelligence. Another example is radiologists. Image recognition software is likely to replace those highly paid skills. Where we saw artificial intelligence as complementary to value chain companies, we would classify value shops in the ‘at risk’ category, ironically coming from companies within the same business model bucket.

Companies that have networking effects are in the sweet spot in terms of business model opportunities. Facebook, Google and Amazon all benefit from networking effects. The larger the number of users, the higher the value. In effect, the value per user is the square of the number of users7. Value networks are usually very disruptive to value shops and value chains. Airbnb is an example of such a disruptive network in the hospitality sector that is pushing out typical value shops as represented by traditional hotels. Companies that are able to shift from value shops into value networks usually find themselves in the sweet spot. In terms of artificial intelligence, we find that many of the leading companies

7 Bergakker, 2016 Artificial Intelligence

are networks. Facebook, Google, Microsoft, Baidu, Apple and Yahoo! are great examples of companies that fit in the value network business model bucket while at the same time being known for their progress in deploying and researching AI.

Artificial intelligence can have two kinds of impact on value networks. First of all, artificial intelligence will be used by value network companies. Search optimization, big data analysis, speech recognition and sentiment analysis are examples of artificial intelligence applications currently employed. On the other hand, artificial intelligence in itself can lead to the creation of a value network. Value networks connect two parties and benefit from scale effects. If some AI protocol would function as the mediator for people and organizations, it in itself would become a tool to build a network. Once this network grows big enough, it could become a competitive threat to other value networks. The most practical application of this that we foresee is an artificial intelligence protocol that listens to what its users want and is able to connect them to those that are able to offer those services. When asking Amazon’s Echo to order a pizza, it connects the person who asks for pizza to the company that can supply pizza. In this case the artificial intelligence protocol behind this network is owned by Amazon, but it is not unimaginable that new networks appear and AI lowers the barriers to create network effects.

Figure 11 | Impact of digitization and connectivity on generic business models (also applicable to AI)

Source: Robeco

Artificial Intelligence |

Value chains are likely to integrate AI, value shops are at risk of being replaced, but winners are also found in this category, and value networks are the holy grail.

Artificial Intelligence

The current artificial intelligence landscape is large, as can be seen in figure 12. Most of the companies mentioned in figure 12 are startups. Machine learning, and natural language processing attract most funding from venture capitalists as well as established big technology companies.

Figure 12 | AI industry overview

Source: Venture Scanner, 2016

We recommend using the business model overlay as presented in this paper and the paper by Bergakker in order to categorize companies according to their generic business model. Winners are networking companies and value shops that focus on artificial intelligence services. Although most companies currently working on artificial intelligence are startups (and university spin-offs), we argue the large technology companies are best positioned to end up with the competitive advantages that artificial intelligence can bring. Google, Facebook, IBM, Apple, Yahoo, Microsoft, Amazon, Baidu and Alibaba have large portfolios of artificial intelligence startups. These networks buy up companies that best fit into their current product and services offerings. The one who owns the data and provides ease-of-use is likely to be a long term winner.

For a public equity investor this has important implications. It is not possible to invest in unlisted startups and at the same time the investment in technology giants will only provide little exposure to the artificial intelligence theme. For the sake of completeness, figure 13 shows which of the large companies is best

Artificial Intelligence |

positioned to benefit from AI.8 Interestingly, some of the companies that tell everyone they lead in terms of AI come out relatively low in the matrix.

Figure 13 | Customer feedback on the companies they use for each AI technique

Source: UBS, 2016

Given the fact that there are no clear winners in AI, it is best to consider using a basket approach until true winners stand out. Instead of looking at the companies that eventually provide artificial intelligence services, it is perhaps better at this point in time to look at companies that provide resources used in that process, as shown below.9

Table 1 ‘Shovel’ suppliers during the

GPU Nvidia, Intel Processors/chips Qualcomm, NXP, ASML, Xilinx Computing power (cloud) Amazon, Microsoft, Google, Alibaba, Juniper Vision/sensors Cognex, Keyence Voice recognition Nuance communications, Mmodal Data Relx, Teradata, Hortonworks, Cloudera, Skymind, Confluent, Databricks

Source: Robeco

Challenged companies are mostly found in the value shop segment, as services offered are being replaced by artificial intelligence solutions. Within value chains we believe companies that do not invest in artificial intelligence have a larger probability of losing out in the long run, as their competitive advantages are likely to deteriorate.

8 These names are mentioned for illustrative purposes only and are not meant to be an investment advice in any way. 9 These names are mentioned for illustrative purposes only and are not meant to be an investment advice in any way. Artificial Intelligence Jeroen van Oerle Marco van Lent Trend analyst Portfolio Manager Robeco Robeco Trends Investing Global Growth Trends Equities

Artificial Intelligence | Important Information Robeco Institutional Asset Management B.V., hereafter Robeco, has a license as manager of UCITS and AIFs from the Netherlands Authority for the Financial Markets in Amsterdam. This statement is intended for professional investors. Therefore, the information set forth herein is not addressed and must not be made available, in whole or in part, to other parties, such as retail clients. Robeco disclaims all liability arising from users other than those specified herein. Without further explanation this presentation cannot be considered complete. It is intended to provide the professional investor with general information on Robeco’s specific capabilities, but does not constitute a recommendation or an advice to buy or sell certain securities or investment products. All rights relating to the information in this presentation are and will remain the property of Robeco. No part of this presentation may be reproduced, saved in an automated data file or published in any form or by any means, either electronically, mechanically, by photocopy, recording or in any other way, without Robeco’s prior written permission. The information contained in this publication is not intended for users from other countries, such as US citizens and residents, where the offering of foreign financial services is not permitted, or where Robeco’s services are not available. The prospectus and the Key Investor Information Document for the Robeco Funds can all be obtained free of charge at www.robeco.com. Investment involves risks. Before investing, please note the initial capital is not guaranteed. The value of the investments may fluctuate. Past performance is no guarantee of future results. Historical returns are provided for illustrative purposes only. The price of units may go down as well as up and the past performance is not indicative of future performance. If the currency in which the past performance is displayed differs from the currency of the country in which you reside, then you should be aware that due to exchange rate fluctuations the performance shown may increase or decrease if converted into your local currency.

Additional Information for investors with residence or seat in France Robeco is having the freedom to provide services in France. Robeco France has been approved under registry number 10683 by the French prudential control and resolution authority (formerly ACP, now the ACPR) as an investment firm since 28 September 2012. Robeco France is only authorized to offer investment advice service to professional investors.

Additional Information for investors with residence or seat in Germany This information is solely intended for professional investors or eligible counterparties in the meaning of the German Securities Trading Act.

Additional Information for investors with residence or seat in Italy This document is considered for use solely by qualified investors and private professional clients (as defined in Article 26 (1) (d) of Consob Regulation No. 16190). If made available to Distributors and individuals authorized by Distributors to conduct promotion and marketing activity, it may only be used for the purpose for which it was conceived. Therefore, the information set forth herein is not addressed and must not be made available, in whole or in part, to other parties, such as retail clients. Robeco disclaims all liability arising from uses other than those specified herein.

Additional Information for investors with residence or seat in Spain The Spanish branch Robeco Institutional Asset Management BV, Sucursal en España, having its registered office at Paseo de la Castellana 42, 28046 Madrid, is registered with the Spanish Authority for the Financial Markets (CNMV) in Spain under registry number 24.

Additional Information for investors with residence or seat in Switzerland This document is exclusively distributed in Switzerland to qualified investors as such terms are defined under the Swiss Collective Investment Schemes Act (CISA). RobecoSAM AG has been authorized by the FINMA as Swiss representative of the Fund(s), and UBS Switzerland AG, Bahnhofstrasse 45, 8001 Zurich, postal address: Badenerstrasse 574, P.O. Box, CH-8098 Zurich, as Swiss paying agent. The prospectus, the key investor information documents (KIIDs), the articles of association, the annual and semi-annual reports of the Fund(s), as well as the list of the purchases and sales which the Fund(s) has undertaken during the financial year, may be obtained, on simple request and free of charge, at the head office of the Swiss representative RobecoSAM AG, Josefstrasse 218, CH-8005 Zurich. If the currency in which the past performance is displayed differs from the currency of the country in which you reside, then you should be aware that due to exchange rate fluctuations the performance shown may increase or decrease if converted into your local currency. The value of the investments may fluctuate. Past performance is no guarantee of future results.

The performance data do not take account of the commissions and costs incurred on the issue and redemption of units. The prices used for the performance figures of the Luxembourg-based funds are the end-of-month transaction prices net of fees up to 4 August2010. From 4 August 2010, the transaction prices net of fees will be those of the first business day of the month. Return figures versus the benchmark show the investment management result before management and/or performance fees; the fund returns are with dividends reinvested and based on net asset values with prices and exchange rates of the valuation moment of the benchmark. Please refer to the prospectus of the funds for further details. The prospectus is available at the company’s offices or via the www.robeco.ch website. Performance is quoted net of investment management fees. The ongoing charges mentioned in this publication is the one stated in the fund’s latest annual report at closing date of the last calendar year.

The material and information in this document are provided “as is” and without warranties of any kind, either expressed or implied. Robeco and its related, affiliated and subsidiary companies disclaim all warranties, expressed or implied, including, but not limited to, implied warranties of merchantability and fitness for a particular purpose. All information contained in this document is distributed with the understanding that the authors, publishers and distributors are not rendering legal, accounting or other professional advice or opinions on specific facts or matters and accordingly assume no liability whatsoever in connection with its use. In no event shall Robeco and its related, affiliated and subsidiary companies be liable for any direct, indirect, special, incidental or consequential damages arising out of the use of any opinion or information expressly or implicitly contained in this document

Additional Information for investors with residence or seat in the United Kingdom Robeco is subject to limited regulation in the UK by the Financial Conduct Authority. Details about the extent of our regulation by the Financial Conduct Authority are available from us on request.

Additional Information for investors with residence or seat in Hong Kong Investment returns not denominated in HKD/USD are exposed to exchange rate fluctuations. Investors should refer to the fund’s Hong Kong prospectus before making any investment decision. Investors should ensure that they fully understand the risk associated with the fund. Investors should also consider their own investment objective and risk tolerance level. Any opinions, estimates or forecasts may be changed at any time without prior warning. If in doubt, please seek independent advice. The content of this document is based upon sources of information believed to be reliable. This fund may use derivatives as part of its investment strategy and such investments are inherently volatile and this fund could potentially be exposed to additional risk and cost should the market move against it. Investors should note that the investment strategy and risks inherent to the fund are not typically encountered in traditional equity long only funds. In extreme market conditions, the fund may be faced with theoretically unlimited losses. This document has not been reviewed by the Securities and Futures Commission. This document has been distributed by Robeco Hong Kong Limited (‘Robeco’). Robeco is regulated by the Securities and Futures Commission in Hong Kong.

Additional Information for investors with residence or seat in Singapore This document is not intended as a recommendation or for the purpose of soliciting any action in relation to Robeco Capital Growth Funds or other Robeco Funds (the “Fund”) and should not be construed as an offer to sell shares of the Fund (the “Shares”) or solicitation by anyone in any jurisdiction in which such an offer or solicitation is not authorised or to any person to whom it is unlawful to make such an offer and solicitation.

Nothing in this document constitutes accounting, legal, regulatory, tax or other advice. Any decision to subscribe for interests in the Fund must be made solely on the basis of information contained in the prospectus (the “Prospectus”), which information may be different from the information contained in this document, and with independent analyses of your investment and financial situation and objectives. The information contained in this document is qualified in its entirety by reference to the Prospectus, and this document should, at all times, be read in conjunction with the Prospectus. Detailed information on the Fund and associated risks is contained in the Prospectus. Any decision to participate in the Fund should be made only after reviewing the sections regarding investment considerations, conflicts of interest, risk factors and the relevant Singapore selling restrictions (as described in the section entitled “Important Information for Singapore Investors”) contained in the Prospectus. You should consult your professional adviser if you are in doubt about the stringent restrictions applicable to the use of this document, regulatory status of the Fund, applicable regulatory protection, associated risks and suitability of the Fund to your objectives.

This document is not directed to, or intended for distribution to or use by, any person or entity who is a citizen or resident of or located in any locality, state, country or other jurisdiction where such distribution, publication, availability or use would be contrary to law or regulation or which would subject the Fund and its investment manager to any registration or licensing requirement within such jurisdiction. Investors should note that only the sub-funds listed in the appendix to the section entitled “Important Information for Singapore Investors” of the Prospectus (the “Sub-Funds”) are available to Singapore investors. The Sub-Funds are notified as restricted foreign schemes under the Securities and Futures Act, Chapter 289 of Singapore (“SFA”) and are invoking the exemptions from compliance with prospectus registration requirements pursuant to the exemptions under Section 304 and Section 305 of the SFA. The Sub-Funds are not authorised or recognised by the Monetary Authority of Singapore and Shares in the Sub-Funds are not allowed to be offered to the retail public in Singapore. The Prospectus of the Fund is not a prospectus as defined in the SFA. Accordingly, statutory liability under the SFA in relation to the content of prospectuses would not apply. The Sub-Funds may only be promoted exclusively to persons who are sufficiently experienced and sophisticated to understand the risks involved in investing in such schemes, and who satisfy certain other criteria provided under Section 304, Section 305 or any other applicable provision of the SFA and the subsidiary legislation enacted thereunder. You should consider carefully whether the investment is suitable for you. This document may contain projections or other forward looking statements regarding future events or future financial performance of countries, markets or companies and such projection or forecast is not indicative of the future or likely performance of the Fund. Neither the Fund, its investment manager nor any of their associates, nor any director, officer or employee accepts any liability whatsoever for any loss arising directly or indirectly from the use of this document. The information contained in this document, including any data, projections and underlying assumptions are based upon certain assumptions, management forecasts and analysis of information available as at the date of this document and reflects prevailing conditions and our views as of the date of the document, all of which are accordingly subject to change at any time without notice and the Fund and its investment manager are under no obligation to notify you of any of these changes. Prospective investors should not view the past performance of the Fund or its investment manager as indicative of future results.

Additional Information for investors with residence or seat in Shanghai This material may not be copied or used with the public. This material is prepared by Robeco Investment Management Advisory (Shanghai) Limited Company (Robeco Shanghai for short) and is only provided to the specific objects under the premise of confidentiality. This material must not be wholly or partially reproduced, distributed, circulated, disseminated, published or disclosed, in any form and for any purpose, to any third party without prior approval from Robeco Shanghai. The information and/or analysis contained in this material have been compiled or arrived at from sources believed to be reliable but Robeco Shanghai does not make any representation as to their accuracy, correctness, usefulness or completeness and does not accept liability for any loss arising from the use hereof or the information and/or analysis contained herein. Neither Robeco Shanghai or its affiliates, nor any of their directors, officers or employees shall assume any liability or responsibility for any direct or indirect loss or damage or any other consequence of any person acting or not acting in reliance on the information contained herein. The information in this material may contain projections or other forward-looking statements regarding future events, targets, management discipline or other expectations which involve assumptions, risks, and uncertainties and is only as current as of the date indicated. Based on this, there is no assurance that such events will occur, and may be significantly different than that shown here, and we cannot guarantee that these statistics and the assumptions derived from the statistics will reflect the market conditions that may be encountered or future performances of Robeco Shanghai. The information in this material is based on current market conditions, which will fluctuate and may be superseded by subsequent market events or for other reasons. The information contained herein may not reflect the latest information on account of the changes and Robeco Shanghai is not responsible for the updating of the material or the correction of inaccurate or missing information contained in the material. Robeco Shanghai has not yet been registered as the private fund manager with the Asset Management Association of China. This material was prepared solely for informational purposes and does not constitute a recommendation, professional advice, an offer, solicitation or an invitation by or on behalf of Robeco Shanghai to any person to buy or sell any product. This material should not be viewed as a recommendation to buy or sell any investment products or to adopt any investment strategies. Nothing in this material constitutes investment, legal, accounting or tax advice, or a representation that any investment or strategy is suitable or appropriate to your individual circumstances, or otherwise constitutes a personal recommendation to you. Robeco Shanghai is a wholly foreign-owned enterprise established in accordance with the PRC laws, which enjoys independent civil rights and civil obligations. The statements of the shareholders or affiliates in the material shall not be deemed to a promise or guarantee of the shareholders or affiliates of Robeco Shanghai, or be deemed to any obligations or liabilities imposed to the shareholders or affiliates of Robeco Shanghai.

Additional Information for investors with residence or seat in Australia This document is distributed in Australia by Robeco Hong Kong Limited (ARBN 156 512 659) (‘Robeco’) which is exempt from the requirement to hold an Australian financial services licence under the Corporations Act 2001 (Cth) pursuant to ASIC Class Order 03/1103. Robeco is regulated by the Securities and Futures Commission under the laws of Hong Kong and those laws may differ from Australian laws. This document is distributed only to “wholesale clients” as that term is defined under the Corporations Act 2001 (Cth). This document is not for distribution or dissemination, directly or indirectly, to any other class of persons. It is being supplied to you solely for your information and may not be reproduced, forwarded to any other person or published, in whole or in part, for any purpose. In New Zealand, this document is only available to wholesale investors within the meaning of clause 3(2) of Schedule 1 of the Financial Markets Conduct Act 2013 (‘FMCA’). This document is not for public distribution in Australia and New Zealand.

Additional Information for investors with residence or seat in the Dubai International Financial Centre (DIFC), United Arab Emirates Robeco Institutional Asset Management B.V. (Dubai Office), Office 209, Level 2, Gate Village Building 7, Dubai International Financial Centre, Dubai, PO Box 482060, UAE. Robeco Institutional Asset Management B.V. (Dubai office) is regulated by the Dubai Financial Services Authority (“DFSA”) and only deals with Professional Clients and does not deal with Retail Clients as defined by the DFSA.

Additional Information for investors with residence or seat in Brazil The fund may not be offered or sold to the public in Brazil. Accordingly, the fund has not been nor will be registered with the Brazilian Securities Commission - CVM nor have they been submitted to the foregoing agency for approval. Documents relating to the fund, as well as the information contained therein, may not be supplied to the public in Brazil, as the offering of the fund is not a public offering of securities in Brazil, nor used in connection with any offer for subscription or sale of securities to the public in Brazil.

Additional Information for investors with residence or seat in Colombia This document does not constitute a public offer in the Republic of Colombia. The offer of the fund is addressed to less than one hundred specifically identified investors. The fund may not be promoted or marketed in Colombia or to Colombian residents, unless such promotion and marketing is made in compliance with Decree 2555 of 2010 and other applicable rules and regulations related to the promotion of foreign funds in Colombia. The distribution of this document and the offering of [Shares] may be restricted in certain jurisdictions. The information contained in this document is for general guidance only, and it is the responsibility of any person or persons in possession of this document and wishing to make application for the fund to inform themselves of, and to observe, all applicable laws and regulations of any relevant jurisdiction. Prospective applicants for the fund should inform themselves of any applicable legal requirements, exchange control regulations and applicable taxes in the countries of their respective citizenship, residence or domicile.

Additional Information for investors with residence or seat in Panama The distribution of this fund and the offering of Shares may be restricted in certain jurisdictions. The above information is for general guidance only, and it is the responsibility of any person or persons in possession of the prospectus of the fund and wishing to make application for Shares to inform themselves of, and to observe, all applicable laws and regulations of any relevant jurisdiction. Prospective applicants for Shares should inform themselves as to legal requirements also applying and any applicable exchange control regulations and applicable taxes in the countries of their respective citizenship, residence or domicile. This document does not constitute an offer or solicitation to any person in any jurisdiction in which such offer or solicitation is not authorized or to any person to whom it would be unlawful to make such offer or solicitation.

Additional Information for investors with residence or seat in Peru The fund has not been registered before the Superintendencia del Mercado de Valores (SMV) and are being placed by means of a private offer. SMV has not reviewed the information provided to the investor. This document is only for the exclusive use of institutional investors in Peru and is not for public distribution.

Additional Information for investors with residence or seat in Uruguay The sale of the fund qualifies as a private placement pursuant to section 2 of Uruguayan law 18,627. The fund must not be offered or sold to the public in Uruguay, except in circumstances which do not constitute a public offering or distribution under Uruguayan laws and regulations. The fund is not and will not be registered with the Financial Services Superintendency of the Central Bank of Uruguay. The fund corresponds to investment funds that are not investment funds regulated by Uruguayan law 16,774 dated September 27, 1996, as amended.

Additional Information for US offshore investors The Robeco Capital Growth Funds have not been registered under the United States Investment Company Act of 1940, as amended, nor the United States Securities Act of 1933, as amended. None of the shares may be offered or sold, directly or indirectly in the United States or to any US Person. A US Person is defined as (a) any individual who is a citizen or resident of the United States for federal income tax purposes; (b) a corporation, partnership or other entity created or organized under the laws of or existing in the United States; (c) an estate or trust the income of which is subject to United States federal income tax regardless of whether such income is effectively connected with a United States trade or business.