Introduction and Health Warning
I wrote this short discussion paper during my internship with Guy Spier at Aquamarine Capital.
I am no expert in artificial intelligence. But as a second year Natural Sciences student at Cambridge University in the UK, I have had some exposure to computer science. And it has become clear to me that most people in the investment world may know even less than I do, and, moreover, may be woefully underestimating the impact that the coming wave of AI will have on their portfolios.
The purpose of this document is to prompt conversations on the impact of AI on the competitive moat of different companies and industries.
In addition to this being a first draft, you should know that I am no expert in the industries I will discuss: I come at both topics – AI and industry analysis – from the perspective of an intelligent layperson, but with the benefit of fresh eyes. I have listed the resources I used at the end of the paper.
You will, no doubt, read things that you disagree with and will also see gaps in my knowledge. But with that in mind, please send me your thoughts – as I will revise and update this discussion based on your and other readers’ feedback, before recirculating this back out to you: My email is [email protected]
What is ‘Artificial Intelligence’ and why should anyone care?
I imagine many of us (particularly my generation) have, at some point, thought along the lines of ‘What would we do without smartphones?’. I believe that this is how a substantial proportion of companies and individuals may come to view artificial intelligence (AI). Officially beginning in 1956, the field of AI has experienced its ‘winters’ and ‘summers’, with its state of affairs today firmly in the latter season: venture capital funding into AI reached $12bn in 2017. Jeff Dean, a veteran at Google and now their Head of Artificial Intelligence, stated that “We’ve gone from less than a thousand people at Google who were trained in machine learning to more than 18,000 people today”. We can try and understand why dependency on AI may develop by first looking at what AI actually is: this is where our first problem arises.
AI has a notoriously varied definition, ranging from the esoteric to the practical. Facebook’s AI Research has a commitment to “advancing the file of machine intelligence and creating new technologies to give people better ways to communicate”, whilst Google aims to “create smarter, more useful technology and help as many people as possible”. However, these do not dispel much ambiguity. A useful definition that I came across, which is also more applicable to the sorts of AI companies may utilise, is as follows: AI is “focused less on endowing ‘intelligence’ to make machine’s human-like, but more to design agents that perceive and act to satisfy some objective without being explicitly programmed to do so.”
The most common form of AI is known as Machine learning (ML). In a normal program, we hard-code patterns into the machine so that, when they are spotted, the machine knows what to do – this is called pattern matching. But these patterns need to be very specific: for example, if I wanted my machine to recognise trees in pictures, I would have to pattern match every possible colour, size and orientation of tree. This impracticality is solved by ML: thousands/millions of examples of trees are shown to the machine and it adjusts its own parameters until it eventually gets the right answer consistently. This is how it ‘learns’. Deep learning (an example of ML) takes this a step further by mimicking the architecture of the human brain to make sense of patterns when there may be
unwanted noise, missing details or other confusions. These principles can be extended beyond picture recognition to almost any conceivable goal. This is why companies have such diverse definitions – they are objective orientated.
From this understanding of AI, we can begin to see why its applicability is so vast. However, valuing that applicability becomes challenging, not just because of the ambiguity in its definition, but also due to the fact that many of these applications are yet to be conceived of, let alone implemented. Estimates of the global value of AI in 2025 range from $35bn to over $70bn. Regardless, the primary takeaway from this is that AI is likely to have a very significant impact and that its reach into society should not be underestimated. Thomas Watson, founder of IBM, famously stated that “I think there is a world market for maybe five computers.”
How may AI affect different industries?
I will split the industries that AI may affect into 2 categories: those where AI interacts directly with individuals, and those where AI works solely with systems: databases, the environment etc. Of course, even in interactions with individuals, AI does plenty of systems work; however, I believe there are certain features and consequences that are unique to the former category and will highlight them as such.
AI interacting with individuals
The spearhead of this form of AI (and industrial AI in general) are technology companies. We display a deceptively high amount of our character online through our choices – what we search for/like/buy. AI can take this data and create profiles of individuals. However, what is often overlooked is the accuracy of these profiles. Both the amount of data provided and the frame of mind in which we make decisions online means that they are often truer reflections of character than those provided by conventional marketing surveys. When answering a survey, we are acutely aware that our actions will be recorded and analyzed, which subconsciously affects our responses regardless of how honest we try to be. However, when browsing the web or shopping online we are often less conscious that this is still being done, allowing our more accurate preferences to appear. This an example where AI not only improves the efficiency of a human task (researching consumer tastes), but can obtain a level of accuracy that a human is prohibited from achieving by the very nature of the process.
Access to ‘who people are’ gives firms like Google, Facebook and Amazon much of their value. Traditionally, this information has been sold to advertisers or used to improve product suggestions. Whilst this business model is now relatively familiar, the growing dependency that firms have on an ever decreasing number of companies to reach consumers may have profound consequences for consumer goods as a whole. Crucially, this power stems from the near constant interaction that these firms’ AIs have with individuals. Every time we ask Siri a question or like something on Facebook, we are presenting the outside world with the opportunity to market to us, which places immense power in the hands of the companies that control who gets to market. It is this power, and the ubiquity of the interactions, that separates AIs that interact with individuals from ones that interact purely with systems – the latter focuses on optimization and pattern recognition whilst the former uses these to understand consumers and the market.
The natural progression of this idea leads to a worrying picture: technology companies marketing their own products alongside others and having the ability to enter consumer goods market with significantly less difficulty than traditional small firms. If I ask Alexa for my usual beer, what’s to prevent it from adding a can of Amazon Beer to my delivery for free as a taster? This ease with which this extra competition can suddenly materialize may require investors to revise how they view the consumer goods landscape. I think this issue also goes a step further – Paul Sheard explains in his paper on “The Economic Impact of AI” that the value of the network these technology firms provide increases as demand rises. Evidence of this enormous demand is best visualized by ‘social graphs’ – graphs displaying the interconnection between individuals, places and things on social media websites. This is a famous depiction of Facebook’s social graph: 2 billion monthly users, ~70% of all Internet users outside China/Russia.
This lead to a ‘winner-takes-all’ effect where a few firms dominate the supply. A consequence of the consolidation of firms with a sizable consumer reach is the hiking of prices for advertising space; naturally, this will harm small firms. This seems analogous to the days when the 2 or 3 channels that were on televisions were the only way that the mass market could be reached.
Paradoxically, however, these technology companies are also the very reason that many small firms exist in the first place, namely that they provide unlimited ‘shelf space’ and the ability to sell anything at a much lower cost. Anyone can create a website and Google will direct any user to it that asks. This leads to huge success stories like the Dollar Shave Club: a company that created a subscription based model to distribute razors and grooming products for $1 per month and was bought by Unilever for $1bn in 2016. These 2 effects seem to contradict: on the one hand, technology companies have the ability to market products that they choose, including their own, and so harm small companies that cannot afford their prices; on the other hand, the large firms also provide smaller ones with the capacity to sell their products widely in the first place through unlimited shelf space.
How can these inconsistent effects be reconciled? If we look closer, much of the described power that technology firms wield already exists – they already control which ads go where. Marketing their own products would only be an extension of this and, crucially, does not affect the ability of better quality products to be sold as well: Amazon does not (and probably will never) prevent a legitimate item from being sold on its platform. This is how the Dollar Shave Club example fits in: their success derives from the superiority of their service – a subscription based, cheap, delivery model that tackles many of the problems consumers face when buying their grooming products – rather than the size of their advertising budget. This suggests that AI may allow companies like Google and Amazon to squeeze out inferior businesses faster by providing quick, easy alternatives via marketing their own products, but that a superior service or business model may still achieve success without the requirement of significant marketing. After all, the Dollar Shave Club had modest advertising and was officially launched on a Youtube video.
Although consumer goods are a natural first example of an industry that concerns itself with individuals’ characteristics, another sector is financial services. AI allows for spending habits to be tracked and the right products to be suggested at the right time. Payment fraud is more easily identified and incorrect card declination can be minimized. Wells Fargo has started using AI to provide more “personalized customer service”, which includes some of the above features. This shift towards AI may be where mobile-only banks such as Monzo start gaining the upper hand. Not only will the services provided be tailored to you (helping with saving, providing advice on loans, high security), but they are not burdened with the additional costs of traditional banks (building costs, larger workforce etc). Insurance companies also benefit. GEICO has introduced a virtual assistant, Kate, that helps customers enquire about policy coverage and other questions. But beyond chatbots, AI can allow insurers to access and interpret vast swaths of data on individuals, which can be collated to analyze trends as well – from the propensity of an individual to smoke to the likelihood of theft in a certain area.
These AI built personal profiles can also be used in tandem with various databases to find solutions to a multitude of problems. Normally, humans ask a search program to find an answer to a specific question such as ‘what model is this car?’. However, unlike traditional searching programs, AI is particularly useful in areas where the problems are so diverse, it is not immediately obvious which solutions may be applicable – similar to picture recognition. One of the industries that is most affected by this is healthcare. Virtual health assistants could help with diagnoses and even recommend preventative measures if a certain lifestyle increases the risk of specific illnesses – this is where personal profiles can help. Apple’s ‘Health’ app is an example of where this is beginning to happen. This not only frees up doctors’ time, but also means treatments can be suggested earlier. IBM Watson, pioneered by David Ferrucci, is now being used to help diagnose and treat lung cancer. Other databases that AI can search include legislation: a Stanford student developed a Chabot called DoNotPay that helped appeal over $4mn in parking fines in 21 months with a 64% success rate.
AI interacting with systems
Moving on from profiling, AI has significant potential in interacting with systems: by systems, I mean things like the environment around it, networks etc. The most common form of this is autonomous cars – whilst this has been widely covered, there are a few perspectives that haven’t seen as much light. An interesting viewpoint, which I came across in Alex Lebkücher’s white paper on the German Automotive Industry, looks at how patent applications for AI in automobiles have been distributed. Whilst most people consider Google and Tesla to be the leaders in autonomous driving, the majority of patents for the technologies have been submitted by German and Japanese manufacturers like Toyota and Volkswagen. Silicon Valley companies do not even feature in the top ten patent appliers – this is often overlooked when analyzing the autonomous automobile industry and may well play a significant role in deciding which companies end up reaping the benefits of the technology.
In fact, it seems that AI may affect the industries that surround the automobile industry, such as the taxi industry and insurance, even more. For car manufacturers, AI is most disruptive in their assembly lines, where automation reduces costs and increases efficiency. In contrast, autonomous cars will likely eliminate the requirement for taxi drivers altogether. Furthermore, given that the marginal cost to produce an extra autonomous car will eventually be less than the cost of buying one, there may be a consolidation of car manufacturers and taxi companies. Evidence of this can be seen in Uber’s initiatives in the autonomous car field.
In addition to cars, a less investigated industry that AI may disrupt is logistics. Two possible ways this may happen are on the storage side and the delivery side. AI’s effect on storing items is not unique to logistics, but is significant: the ability for goods to be catalogued and retrieved at much lower cost has been a huge aid to companies. The primary result of this is scalability – where previously a company may have required 10-20 more individuals to open another warehouse, now it may require 1-2 robots. In addition to this, AI’s skill at analyzing patterns could allow it to suggest which products are likely to be in higher demand when and even where the best places are to store them. This process of identifying problems before they even happen extends onto the delivery side. AI has the ability to scan entire areas and not only find the quickest routes, but also predict which routes may become more congested at which times. With the advent of autonomous trucks through companies like Tesla and Uber, action can be taken without a human present: if routes become too congested or an accident occurs, backup trucks can be dispatched instantly with no extra cost. Bradley Jacobs, CEO of XPO Logistics, acknowledges this by stating that “Automation…is providing the answers to secular trends …which touches contract logistics, last-mile logistics, brokerage, expedite, truckload and less-than-truckload.” This form of solution, prediction and optimization, is applicable to a vast number of industries including telecommunications and engineering: General Electric is now using AI to track wear and tear in its machinery as well as boost efficiency of its wind farms.
However, if we take this line of thought to its natural conclusion, we find that AI raises some unresolved questions. Given that processing power will continue to become cheaper (the most common reference for this is Moore’s Law: the idea that the processing power of computers doubles every few years) and the competitive pressure from other AI-using companies, AI will likely become more ubiquitous in industries that require optimization and efficiency. But unlike AI that answers queries or profiles individuals, optimization AIs have an end point: there is only so much optimization possible in a delivery route or a network within the framework of natural random error. When this point is reached (likely in the distant future), how do companies differentiate themselves to customers? Must their competitive advantages come from areas other than quality of service such as lower costs, brand names or patents? If so, does this not skew in favor of more established firms that have the reputation and finances to implement them? I do not have the answers to these questions but I think they are worth considering.
AI in asset management
So far, we have considered AI’s impact on the companies that we research. However, AI’s influence on the way research is conducted and its impact on the asset management industry should not be overlooked. Quantitatively, the uses of technology are clear and, in many cases, already present: screening for companies; financial statement analysis; comparisons with competitors. This is where AI is most likely to penetrate the industry – passive investing and algorithmic trading lends itself well to the repetitive, data driven nature of AI. Along with quantitative hedge funds, this is also notably seen in the PE/venture capital sphere: Hone Capital, the Silicon-Valley base arm of CSC Group, uses AI to support the investment team in seed investments.
It is the qualitative side of investment research, that which the active investor devotes much of their time to, that provides a greater challenge for AI. Investors like to know not only how profitable a company is but where their profitability is derived from and how sustainable those factors are. Aspects like identifying and categorizing barriers to entry or evaluating management decisions and integrity are not only crucial to understanding a business but can also be highly subjective. Examples I have used in previous sections of AI-friendly tasks (profiling; optimization; autonomous vehicles) are all repetitive and easily definable. This is not so with a question like ‘what are the causes of barriers to entry in the CPU-chip market?’ Whilst a correct answer may exist, it is much more difficult to train an AI to find it: millions of iterations of similar problems are required to be fed into the program. Not only is the availability of such similar problems insufficient to produce accurate results but the economics of different industries vary to such a large degree that any sort of generalization across them would be impractically demanding.
That is not to say that AI does not have a place in the industry. Alongside its ability to make company recommendations and provide safety checks, a more unconventional use of AI may be in aiding ‘Intelligent Cloning’. The term denotes looking at what other successful investors have done and using their ideas. Mohnish Pabrai is a great proponent of this and the concept was discussed at the ‘Zurich Project 2017’ by Peter Coenen. AI could screen ideas from other investors and run them through a filter before presenting them. But, importantly, this is not the whole picture. I doubt that many investors would be satisfied with blindly following other portfolios – they will want to conduct their own analysis before committing. This lies at the heart of AI’s potential: alongside human work. Eric Chang of Microsoft Research Asia stated that “The AI plus human intelligence (HI) model holds the most promise. Our competition is not machines. It is the other people plus machine teams out there.” AI presents an opportunity to streamline relevant ideas to the investor faster, whilst still leaving them to the art of qualitative analysis and interpretation.
Overall, I believe that AI working in conjunction with humans will allow investors to expand their field of search. It may also provide a useful safety net to counteract the strong psychological influences that the industry thrusts upon its participants. Falling in love with a stock or panicking after a significant market drop are thorns in the side of the value investor. Many, such as Guy and Mohnish, use checklists to fight these impulses. Results from an AI may also aid investors in the battle against irrationality.
How much of this is excessive hype? As with other ground-breaking technologies, whilst the opportunities are certainly substantial, the human mind has an uncanny ability to run away with itself. AI requires data and time and it is important to put that in perspective. Researchers believe that the threshold required for an AI to beat a human at Go is roughly 10-million rules. Writing and training that much code takes a long time and that is for a simple, rule based game. In an ImageNet image recognition competition, 50,000 people took 3 years to label over 100 million pictures. Although this is not a small feat, it is considerably less nuanced than investment management. In the depths of the Internet boom, there were predictions that users would double every 3 months for years. This implausibly high growth didn’t happen but that is not to say that the Internet was not revolutionary in almost every aspect of life.
In the same vain, it is difficult to envisage areas where AI will not have some impact. Peter Diamandis, Founder and CEO of X Prize Foundation and one of Fortune’s ’50 greatest leaders of all time’, describes it as travelling towards a “world of near perfect data” where we “democratize the ability for everyone to have equal access to services ranging from healthcare to finance advice”. But beyond replacing some labour and streamlining businesses, AI promises to profoundly change the way companies view their objectives. When problems like scaling a product, reaching a market and even understanding consumer tastes become as easy as calling someone on the other side of the world or Googling a statistic, other challenges like idea generation and company strategy can be focused on. This shift in priorities is an effect produced by previous radical technologies from the steam engine (getting somewhere else ceased to be a problem) to the telephone (communication became considerably easier). The consequences of this are unknown – it may cause the frequency of brilliant ideas to increase as people have more time to ponder them, or it may simply permit greater leisure time for individuals. Perhaps it will bring us closer to Keynes’ prediction of a 15-hour work week.
People mentioned in this paper:
Guy Spier: https://en.wikipedia.org/wiki/Guy_Spier
Satvik Subramaniam: https://www.linkedin.com/in/satvik-subramaniam-bb783623/
Paul Sheard: https://www.spglobal.com/cn/who-we-are/our-company/our-people/leadership/Paul-Sheard.html
Alex Lebkücher: https://de.linkedin.com/in/alex-lebkücher-153495136
David Ferrucci: https://en.wikipedia.org/wiki/David_Ferrucci
Jeff Dean: https://en.wikipedia.org/wiki/Jeff_Dean_(computer_scientist)
Bradley Jacobs: https://en.wikipedia.org/wiki/Bradley_S._Jacobs
Peter Diamandis: https://en.wikipedia.org/wiki/Peter_Diamandis
Peter Coenen: https://www.linkedin.com/in/petercoenen/
Eric Chang: https://www.microsoft.com/en-us/research/people/echang
Books on AI:
Artificial Intelligence and Machine Learning for Business: Steven Finlay https://www.amazon.com/Artificial-Intelligence-Machine-Learning-Business/dp/1999730305
The Executive Guide to Artificial Intelligence: Andrew Burgess https://www.amazon.com/Executive-Guide-Artificial-Intelligenceapplications/dp/331963819X/ref=sr_1_1?ie=UTF8&qid=1524228953&sr=8- 1&keywords=artificial+intelligence+and+business+management
The AI Business: Patrick Winston & Karen Prendergast https://www.amazon.com/dp/0262730774/ref=sxts_sxwdspuwylo_rv_3?pf_rd_m=ATVPDKIKX0DER&pf_rd_p=3534076942&pd_rd_wg=x1P67&pf_rd_r=F4SZZM0P90AJ12K M0MBP&pf_rd_s=desktop-sx-topslot&pf_rd_t=301&pd_rd_i=0262730774&pd_rd_w=z5VzH&pf_rd_i=ai+business&pd_rd_r=a640658f-7768- 4e01-8b99-271d39956c5b&ie=UTF8&qid=1524229120&sr=3
Human + Machine: Paul Daugherty & H. James Wilson https://www.amazon.com/Human-Machine-Reimagining-WorkAge/dp/1633693864/ref=sr_1_1?s=books&ie=UTF8&qid=1524229183&sr=1- 1&keywords=human+%2B+machine
Prediction Machines: The Simple Economics of Artificial Intelligence: Ajay Agrawal, Joshua Gans & Avi Goldfarb https://www.amazon.com/Prediction-Machines-Economics-Artificial-Intelligence/dp/1633695670/
Superintelligence: Nick Bostrom https://www.amazon.com/Superintelligence-Dangers-Strategies-Nick-Bostrom/dp/0198739834
The Master Algorithm: Pedro Domingos https://www.amazon.com/Master-Algorithm-Ultimate-Learning-Machine/dp/0465094279/
Rise of the Robots: Martin Ford https://www.amazon.com/Rise-Robots-Technology-Threat-Jobless/dp/0465097537/
Influential individuals in AI:
Elon Musk (Tesla & SpaceX): https://en.wikipedia.org/wiki/Elon_Musk
Jeff Dean (Head of AI – Google): https://en.wikipedia.org/wiki/Jeff_Dean_(computer_scientist )
Andrew Ng (Chief Scientist – Baidu): https://en.wikipedia.org/wiki/Andrew_Ng
Martin Ford (Author): https://en.wikipedia.org/wiki/Martin_Ford_(author)
Yann LeCun(Chief AI Scientist – Facebook): https://en.wikipedia.org/wiki/Yann_LeCun
Investors using AI:
Nick Granger (Man AHL): https://www.ahl.com/teams/nick-granger
Babak Hodjat (Sentient Investment Management): https://www.sentientim.com/management/babak-hodjat/
David Andre (Cerebellum Capital): http://davidandre.com
Vasant Dhar (SCT Capital Management): https://en.wikipedia.org/wiki/Vasant_Dhar
Manoj Narang (MANA Partners): https://www.linkedin.com/in/manoj-narang-400a011/
Ray Dalio (Bridgewater Associates): https://en.wikipedia.org/wiki/Ray_Dalio
Crispin Odey (Odey Asset Management): https://en.wikipedia.org/wiki/Crispin_Odey
Bill Ackman (Pershing Square Capital): https://en.wikipedia.org/wiki/Bill_Ackman
Mohnish Pabrai: https://en.wikipedia.org/wiki/Mohnish_Pabrai
Neil Woodford (Woodford Investment Management): https://en.wikipedia.org/wiki/Neil_Woodford
Warren Buffett (Berkshire Hathaway): https://en.wikipedia.org/wiki/Warren_Buffett
Terry Smith (Fundsmith): https://en.wikipedia.org/wiki/Terry_Smith_(businessman)
History of AI – Harvard University http://sitn.hms.harvard.edu/flash/2017/history-artificial-intelligence/
Definitions of AI – Forbes https://www.forbes.com/sites/bernardmarr/2018/02/14/the-key-definitions-of-artificial-intelligence-ai-thatexplain-its-importance/#629f196d4f5d
How does AI work? – Futurism & World Economic Forum (Video by Facebook) https://futurism.com/1-evergreen-making-sense-of-terms-deep-learning-machine-learning-and-ai/
Investment into AI – South China Morning Post, Statistica & Tech Emergence
The Economic Impact of AI – Paul Sheard Facebook’s social graph – Medium https://medium.com/@johnrobb/facebook-the-complete-social-graph-b58157ee6594
Uses of AI by technology companies – Medium https://expertise.jetruby.com/how-5-of-the-most-innovative-tech-companies-are-using-ai-in-2017- 85ae92a5d9f2
AI at Google – The Australian https://www.theaustralian.com.au/business/technology/googles-jeff-dean-were-using-ai-to-build-machinelearning-systems/news-story/c0120711ad2368cc8d9938f7575a78ba
Jeff Dean on AI – Datanami & The Verge https://www.datanami.com/2018/03/13/jeff-dean-thinks-ai-can-solve-grand-challenges-heres-how/
AI in the legal profession – American Bar Association https://www.americanbar.org/publications/youraba/2017/july-2017/artificial-intelligence-and-the-future-oflaw-practice.htm l
AI at GEICO – GEICO https://www.geico.com/web-and-mobile/mobile-apps/virtual-assistant /
AI in Healthcare – Medium https://medium.com/@Unfoldlabs/the-impact-of-artificial-intelligence-in-healthcare-4bc657f129f5
AI at Wells Fargo – Fortune http://fortune.com/2017/02/10/wells-fargo-artificial-intelligence/
Bradley Jacobs on AI in logistics – JOC.com https://www.joc.com/aro/bradley-s-jacobs-ceo-and-chairman-xpo-logistics
AI in Transportation – Tech Emergence https://www.techemergence.com/ai-in-transportation-current-and-future-business-use-applications/
AI at General Electric – Technology Review https://www.technologyreview.com/2017/06/27/150784/general-electric-builds-an-ai-workforce/
The German Automotive Industry – Alex Lebkücher AI in asset management – CFA Institue, Accenture https://capitalmarketsblog.accenture.com/machine-assisted-investing-how-ai-is-changing-private-equity
Intelligent Cloning – Peter Coenen https://thevaluefirm.com/wp-content/uploads/2017/11/Thoughts-on-Intelligent-Cloning.pdf
Peter Diamandis on AI – Inc, Impact Theory & Diamandis Blog https://www.inc.com/tess-townsend/diamandis-artificial-intelligence.html