As an angel investor and entrepreneur myself, I have been investing in many different companies, and have encountered many different projects, interesting companies and startups over the last two decades. As an expert in Artificial Intelligence, which is a modern hot area for investment, I have put together this quick guide for anyone who is thinking of investing in AI.
In this guide I will present the current narrow reality of AI, where AI has been applied to a very specific application, and present a market overview — who is who, and who are the major players in AI. We will also cover the AI investor perspective — what are the things you should look out for, how do you assess a team, and how do you assess a new project, as there are many interesting AI projects out there. This guide gives you the basic information that you need to cut through the noise that there is right now and help you make a better informed decision. At the end of the guide, I am discussing a bit about future trends in the AI space that may become increasingly important.
As you can see from the AI timeline, AI has been around since the 1950’s, when the term Artificial Intelligence itself was coined, and the prescient ideas of Alan Turing inspired many budding Computer Scientists to start thinking about how they could make a machine learn and reason logically, like human beings. AI progressed rapidly in the 1960’s when a lot of new interesting advances were made — during this time neural networks were conceived and the early successes led led to overpromises and hype, which may seem familiar! Some countries, like Japan, invested heavily in these early AI systems, leading to even more hype and expectations. At this point, many problems started showing up in AI systems and serious concerns about how the technology would scale up for practical problems. Delivery wasn’t matching the expectations and this led to the so-called AI winter, in which the amount of funding for AI was drastically reduced for many years. In the late 1990’s things started getting better. One of the main milestones was the famous chess game against Kasparov by IBM’s Deep Blue system — decades before anyone expected AI to beat a world-class chess grandmaster and champion. The success of IBM started a renewal of interest in AI. I myself experienced this lack of funding in AI for many years, with people asking me why I chose to go into AI and not other fields like high frequency trading or other seemingly more lucrative fields. But perseverance led to the current fruit that we are now reaping. In the early 2000’s, a group of mainly Canadian, British and American researchers started reviving some of the long sought goals of a particular group in AI, called the connectionists, which base solutions on neural networks that are capable of learning and understanding patterns. This revival, enabled by the higher availability of data and the substantial increase in hardware speed and power have led to the current wave of Deep Learning systems which today are the mainstay of AI systems. Gamers and specifically gaming hardware also had a large influence on the success of Deep Learning — specialised Graphical Processing Units that are used in modern video games turned out to be very suitable for the type of learning used by modern AI systems.
Let us now put AI in the context of Amara’s law — that we tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run. When assessing current AI systems in the context of hype vs. reality, we can see that AI is safely past the technology trigger and squarely in the peak of inflated expectations. There are a lot of promises by different AI companies and in fact a recent study showed that around 60% of AI companies that claim to be using AI are not really using much AI in their process — so we are currently in a growing bubble. Unfortunately, we will inevitably soon go through the through of disillusionment, in which a lot of existing AI companies may go bankrupt or otherwise exit with a whimper and will have a few winners that survive and thrive. I do not think that a next AI winter is coming. AI is here to stay, as unlike other past periods, AI is delivering commercially useful results and while the hype is still there, results are also accumulating steadily. I think the future is quite optimistic and those investors who make good investments in promising AI start-ups may be in for a very good return on their investments — and in some cases, truly astronomic returns.
I want to cover some common AI misconceptions, mainly the difference between Artificial Intelligence, machine learning and Deep Learning. Deep Learning is a subset of machine learning, and it is definitely not the be all or end all for AI. There are many other different techniques that supplement and complement Deep Learning such as Bayesian Learning, Monte Carlo Simulation, Genetic Algorithms and so on. Machine learning itself is a subset of Artificial Intelligence. AI concerns itself with other things, apart from machine learning — for example, the presentation of knowledge, how do you get from that knowledge to symbolic processing systems, how do you get efficiently from point A to point B under a variety of constraints and various different algorithms and techniques that are in widespread use. Deep Learning is a direct descendant of methods that were developed in the 1990’s and the underlying theory has changed minimally since then, although now we have practical means of implementing the theory in practice and put it to good use.
Surprisingly, one of the very first remarkable applications of modern Deep Learning, was the identification of cats in those adorable YouTube videos of cats doing little silly stuff. A group of researchers created an experiment which taught a Deep Learning neural network what a cat is all about — how does a cat move, and what does it look like? And then, let’s try to find all the cat videos on YouTube! Astonishingly, the experiment worked far beyond any expectations! Here is how that newly born Deep Learning AI thought that the stereotypical cat looks like below..
This successful demo eventually led to the creation of Tensor Flow Computing, one of the mainstays of current Deep Learning systems.
You may not realise how much you already use AI in everyday life. For example — one application that you may frequently use is Google Translate. Google Translate used to have many human written knowledge rules that translated, for example English to Italian, Japanese to Chinese, etc. After thorough training, Deep Learning translation systems managed to achieve better results than the human written rules. Nowadays, the translation system has been replaced by a system that is entirely AI based with no human derived knowledge, that takes one language, say English, then takes another language, say Japanese — looks at the differences and similarities between them and automatically creates a translation without any need for human rules through a basic process that is central to all Deep Learning systems called back-propagation. As you can see, this is how Google Translate actually works.
You will see how an English sentence, is converted to a pattern, shown in orange, and is merged to Korean, shown in blue and to Japanese, shown in red. This pattern matching ends up creating sentence fragments that there then joined up by the AI system itself to create the end result, in this case creating a Japanese sentence from an English one. Another misconception I want to clear is about the difference between AI and AGI, Artificial General Intelligence. When most people imagine what AI should look like, they imagine something super intelligent, that can do and learn things on its own without being taught and that is almost sentient. This type of AI is commonly known as Artificial General Intelligence — and rest assured that as of the beginning of 2020, no one has come close to inventing it yet. The current breed of AI that we have is called Narrow AI — which is something that is focused on one specific application and has one specific solution, focused on one domain or task. If you encounter a project that says that it is going to solve AGI, do take it with a pinch of salt, and be prepared to invest for a very long time! Current expert consensus is that it may take anywhere between the next 20 to 100 years for AGI to ever come about. There are other experts who think it will take even longer than 200 years or that it will never be invented. Let’s be optimistic and say that it may take up to another 100 years — you will be in for a very long wait before you will ever see any return on investment.
Now that misconceptions have been cleared, I want to have a look at emerging patterns in AI investments on a regional basis.
- Chinese start-ups have few but very large investments
- EU start-ups have a steadily increasing number of smaller investments
- US start-ups have a steadily increasing number of larger investments
- Average per investment:
As you can see in the table of Emerging Patterns, EU and US start-ups have very different patterns of funding. Chinese start-ups also have a different pattern of funding. For example, Chinese start-ups have very few, but very large investments. In the EU, European AI startups have a steadily increasing number of smaller investments, while US AI startups have an increasing number of larger investments. If a start-up is based in the US instead of the EU, you will roughly get around between a 2.5 to 3.5 increase in the valuation multiplier. Hopefully this gap will start to close as time goes by and more European AI start-up companies become successful.
One of the best sources of knowledge for the AI industry is the CBINSIGHTS AI 100. I thoroughly recommend that you go and download it and look at the different segmentation that is used, as it one of the best industry segmentation and allows you to go and do further research that compliments this presentation.
I now want to present a short overview of the AI global market. This market analysis is based on CBINSIGHTS data from 2019 and OECD data since 2011. Effectively, 12% of all worldwide private equity investments, up to the second half of 2018 — have been into AI. 12% of the world’s investments — which is a lot of money! The US accounts for around 70 to 80% of investments in AI and China for most of the rest. According to my own internal investment models there is a steady growth expectation year on year of almost 200% which is very significant. Such a growth rate can either mean that we are in a hot bubble that may burst at some time in the next few years, or could be something that leads to an AI driven future with many successful companies growing and expanding worldwide.
As a long term angel investor myself, I have a passionate interest in angel investors and early stage start-ups. Angel investors have been involved in around 20% of all global deals involving AI. The average deal size where angel investors were involved was around $2.73 million with the medium size deal being around $1.6 million. The usual pattern seen in other tech start-ups applies. Angel investors typically start off as significant funding sources in the early stage and then larger sources of funding like VC’s start coming in. There are a number of AI startups that have largely skipped angel investor funding and went straight to VC funding — this is a rather unusual funding path — but increasingly common, especially in US West Coast start-ups and Chinese start-ups. Otherwise, the typical funding pattern is very similar to any other tech start-up and angel investors are poised to be make a killing when AI start-up investment opportunities become even more and more available in the near future, and successful exits start delivering lucrative financial results.
So where are the best AI startups in the world? Obviously anyone has the chance to create the next big thing in AI in their garage or somewhere remote in the world, since all you need is basically a laptop, a little bit of computing hardware resources based on the cloud, adequate data and expertise and off you go! However, geographical location does increase the chances of a successful outcome in many cases with perhaps the exception of deep tech and cross-industry solutions that are less sensitive to the physical location of the start-up.
When you analyse the top 100 AI start-ups, you can see that there were $11.7 billion in total aggregate invested across 367 deals and 11 unicorns that are currently identified as being AI companies. The average total funding for each company was around between $52.8 million to $129.5 million per AI company — not at the early stage, but at a more mature stage — at around series B stage. US and Chinese AI start-ups have been the best funded in the last 12 months and it looks like this trend is going to continue.
We have also split the world into main geographical locations that are most amenable to AI start-ups. In the top tier, Tier 1, we have USA and China, that completely stand apart from the rest of the world. In Tier 2 we have the UK, which is a distant second to the US and China. In tier 3, we have the EU represented by Germany and France together with Canada, Israel and India. Canada has a long history of AI research and it is one of the places where you will find the AI researchers that kept Deep Learning techniques going on. In Tier 4 we have Switzerland and Spain, together with Japan and South Korea. Switzerland is one of the countries with a long history of AI research, having a lot of experience in Fintech and also a country that invented a lot of the AI techniques for time series data that are widely used around the world. Japan is one of the countries that has been supporting AI research since the early 1980’s and 1990’s and has a long deep history of supporting AI. In Tier 5, there’s Singapore, Hong Kong and Taiwan, with its chip manufacturers, Australia, Sweden, Ireland, Finland, Netherland and Austria. The top funded areas in AI have been cyber security, healthcare, and enterprise business applications. The European Artificial intelligence landscape is dominated by London and Berlin, Paris, Madrid and Stockholm together with other countries to varying degrees. The European country of Malta has created an EU style certification programme for AI systems that may inspire other countries over time.
So what are the specific aspects of AI start-ups? There are some things that are very specific towards AI start-ups and some others that are common as with any other start-up. One of the main AI specific aspects is that AI start-ups need data. With the current state of the art in AI, AI machine learning algorithms need a lot of data. Datasets have become very important, which has led to some liking them to a sort of new gold. There is a well known phrase in computer programming that garbage in, garbage out — so Quality Assurance is very important for AI algorithms — if you do not train the AI algorithms in a good way, you are not going to get good quality results. Unfair bias can also creep into your dataset, so this is something that is very important to make sure that the data set is fair and that the AI start-up is not producing misleading results. Consultants and experts that will need to check the dataset and help companies build up the datasets can be very expensive — in fact they are one of the main sources of bottlenecks for AI startup growth. So when evaluating AI start-ups, those start-ups do not depend on one source of data, or that have enough of their own customers to generate the datasets for them or in collaboration with them is one advantage that is inbuilt. In fact there is this virtuous cycle that creates a very positive feedback loop in which a company has access to some data that is useful, which is then used to provide a basic useful service, and this service leads to the creation of more data — in a never ending cycle of improvement. The flexibility of the AI platforms is also very important, since we do not really know what the consequences of AI technology will be, and how exactly things will pan out — so AI start-ups are expected to have to pivot a lot. Being flexible prevents excessive rework from being done or catastrophic moments where you must do everything from scratch.
What do promising AI start-ups need to have? From my past experience, I believe that AI start-ups need to have all the following characteristics:
- A solid founding team
- Expert AI skills
- Access to Domain Experts
- A flexible platform
- A means to acquire or generate a dataset or datasets as required. This may not be always applicable for deep tech start-ups
All these aspects need to be present in order to have a very good chance of success. If one of them is missing that is probably a big red flag. If more than two are missing you probably should not invest in such a start-up until they acquire that particular skillset or missing asset.
For AI start-up teams, the team is one of the most important factors of determining whether to invest or not in an AI start-up, as with most other start-ups in general. The expertise in AI is in acutely short supply. Globally there are around 300,000 qualified people , and the number of PhD’s is even less. There were estimates that initially there were around 3000 PhD level AI talent worldwide. So those who has done PhD’s in AI many years ago, like myself, are quite a rare breed! Due to the lack of supply and the high demand, the premium on AI related salaries is quite high. Lots of studies and salary surveys have showed that for teams in an AI start-up, a 30–150% premium on top of an average salary needs to be accounted for — so the HR cost can be quite expensive, even more than in the average tech start-up. HR will be one of the main cost items of any AI start-up, together with cloud computing or hardware computing resources that the AI or machine learning algorithms need. Then there is the usual consideration for the team composition but obviously in an AI start-up you do need to have a little bit more tech and engineering expertise. You cannot really have an AI start-up made solely out of marketing and sales people, like other start-ups can manage to do. Some AI problems are very hard and solving them require world class teams to be able to solve them, so the harder and the more exotic the problem, the more expertise and intellectual acumen the team will need to have apart from commercial skills and experience.
One problem that AI start-ups face, similar to what other tech companies face, is the so called big tech’s kill zone. This problem is very strongly felt in promising AI start-ups that have promising patent portfolio or technology that if allowed to grow may eventually threaten the dominance of big tech and other established market players. The sell or die effect is very noticeable once an AI start-up becomes important and develops something that is very interesting. What typically happens is the one or more of the larger established players will come and make an acquisition offer. If this offer is refused, what usually happens is that particular giant develops its own alternative solution and open sources it to make it available at such a low price point that they will literally take that small start-up out of business. This is called the kill zone effect and unfortunately it is a very commonly experienced effect, so this is something that investors need to take into consideration. This can obviously work both ways, as it can be seen as an opportunity for an early exit but it can also be seen as a threat — if you do not want an early exit and then a big giant comes along and makes such an offer — what do you do then? Do you risk being taken into the kill zone or do you just accept the money and continue by joining them? Or do you forge ahead on your own independent path, fully knowing that the rebuffed giant may be planning to take you out of business soon? This factor should be one of the items discussed with the founder team prior to an investment or when enough funding and traction has been gained to make such an event likely to happen in the near future.
The race to acquire AI start-ups is heating up. As you can see from the CB Insights graph, the number of acquisitions overtime is on the increase. Of course, you should jump onto the bandwagon whilst the timing is still right and obviously keeping in mind the points in this article to make the right choice.
Naturally every advanced technology that has the potential for immense disruption presents its own opportunities and challenges. A lot of people have started imagining about AI and its unintended consequences. Stuff that was recently the purvey of science fiction has now become science fact. There are serious concerns about whether we are abdicating our responsibility to a machine that can learn. As with any new advanced technology, the risk of unintended consequences with AI are very real. Ethical and safety guidelines can provide a way of ensuring that such risk is kept within acceptable ethical boundaries. One of the first people to think about the ethics of AI and making sure AI is done in the right manner is the science fiction writer Isaac Asimov, who created Asimov’s laws of robotics during the period from 1942 to around 1985. Asimov has though a lot about the various nuances surrounding AI and has published his laws in various different science fiction stories which explore these rules. There are four laws — the so called zeroth law (law zero) which is that a robot may not harm humanity, or by inaction, allow humanity to come to harm. The first law is that a robot may not injure a human being, or through inaction, allowing a human to come to harm. The second law is that a robot must obey lawful orders given to it by human beings, except when it comes into conflict with the first law. Finally, the third law is that a robot must protect its own existence apart as long as such protection does not conflict with the first or second laws. These rules seem to be quite self-evident, yet in practice have a lot of loopholes and exceptions. In reality even though Asimov’s laws have been practically guidelines for ethical AI for more than seventy years, up until recent times, people did not really think much about them as they seemed to be very far off in the future and inapplicable to the present situation. With the rapid adoption of modern AI, AI researchers and policy makers started realising that there is a need for a practical ethical framework — AI systems in the meantime are being deployed without much thought about their unintended consequences.
OECD Principles on Artificial Intelligence - Organisation for Economic Co-operation and Development
The OECD Principles on Artificial Intelligence promote artificial intelligence (AI) that is innovative and trustworthy…
- AI should benefit people and the planet by driving inclusive growth, sustainable development and well-being.
- AI systems should be designed in a way that respects the rule of law, human rights, democratic values and diversity, and should include appropriate safeguards — for example, enabling human intervention where necessary — to ensure a fair and just society.
- There should be transparency and responsible disclosure around AI systems to ensure that people understand AI-based outcomes and can challenge them.
- AI systems must function in a robust, secure and safe way throughout their life cycles and potential risks should be continually assessed and managed.
- Organisations and individuals developing, deploying or operating AI systems should be held accountable for their proper functioning in line with the above principles.
- Consider some of the unintended consequences before unleashing AI onto the world.
Luckily, a lot of bright people have been recently thinking about modern ethical AI guidelines and the OECD principles has recently published globally accepted guidelines on the development of AI systems that are trustworthy and will bring about value to human society. AI systems need to be designed in a way that respects the rule of law and that bring positive value to society, benefiting people by driving inclusive growth and sustainable development. There are a lot of different rules and guidelines that people should follow. One of the main rules is to not to weaponize AI and never give AI the power to make a decision that will impact human life in a severe way. This is a red line that should never be crossed as the potential unintended consequences that could happen are generally too bad to contemplate. While we are far away from creating AI technology that can give rise to the Terminator, this is something that investors should really look out for in their due diligence. Ethical issues and the founding team values need to be examined for alignment with the investors.
I’m finally going to look at the future trends in AI, especially since AI is now finally here to stay. I am going to make seven predictions about AI in the short and medium future. My first prediction is that AI is going to drive society to change. There is going to be more management and less manual labour. There is going to be a focus on assistance rather than replacement of people with AI systems. More leisure time available for humans to connect with each other and enjoy the fruits of our automated labour. Second, I expect AI to be embedded everywhere, from the most exotic to the most mundane applications. It is going to be everywhere around us. Why? Because generally, the smarter things behave, the better and easier our life will be. Third, I think that the AI hype bubble will burst in the next 5 years, by 2025, I think most of the AI startups that are not really based on a solid foundation will be gone — like what happened in the world wide web bubble of the early 2000’s. Fourth, I think that ethical AI can and will impact the development of AI positively. I think driverless cars for example, will lead to far fewer fatalities in the medium and long term but they need to be implemented in an ethical way, within an ethical framework that does not just throw these driverless cars on the street without much thought, but that there will be proper human oversight and human overrides and control over such systems. Medical AI will be reducing misdiagnosed errors and thus lead to more lives saved, and higher life expectancy. Medical AI can also substantially reduce the time before cures and vaccines are created when new diseases crop up from time to time. Fifth, I think existing tech giants will do their best, unfortunately, to prevent emerging start-ups from becoming a threat. I think the kill zone effect is going to become more and more pronounced over the next decade. Vice-versa, the opportunity presented by this aggressive yet defensive play can also be exploited by investors by seeking out good early exit terms. Six, I predict that hardware advances are going to drive innovation and accelerate software innovation. Software has become quite interestingly advanced at the moment and new algorithms will undoubtedly drive AI forward, however there is severe limitation placed on the growth of AI by the hardware itself. I do predict that there will be new hardware architectures and applications that will be coming on the market which will complement the software. Architectures like neuromorphic computing and others that may have yet to be invented, will be one of the key ways in which innovation is going to be driven. Last but not least, I think that whoever invents the best AI is going to dominate the market to such an extent that they will even effect geopolitical power.
Today’s AI applications focus on very narrow tasks. I think that as AI progresses, it’s going to get broader in scope and more applicable. We are going to see more intelligent systems of the science fiction type — maybe not for now, but in a few years’ time. We may need to wait a little bit until we see them, but this is going to be something which is almost inevitable. Combined together, all these advances will reshape business, society, markets and industries. And you never know, maybe AI is the last thing that humans need to create. Once AI systems become intelligent enough to design and generate their own AI systems, maybe we can just sit back and relax and enjoy life, do more things than previously possible than anyone else. The future looks bright, and I think AI investment is the way to go.
You have seen what to look out for in promising AI start-ups. The composition of successful teams, what the main characteristics of AI project should be, and also roughly how much money AI start-ups need in terms of funding and an idea of their return potential. I hope that you are going to be investing in AI and you will be part of this exciting new world of the future!