We recently launched our latest, ground-breaking AI research: ‘The State of AI 2019 — Divergence’ at the historic Royal Institution in London. It’s an unmissable guide to developments in AI, and their implications, for entrepreneurs, corporate executives, investors and policy-makers. You can watch a video of the keynote talk above or read the transcript below. I explain how:
- among nations and companies, winners and losers are emerging in the race for adoption, the war for talent, and the competition for value creation;
- the landscape for entrepreneurs is changing. Our unique analysis reveals how Europe’s 1,600 AI startups are maturing, navigating unique capital dynamics and bringing creative destruction to new industries;
- as advances in AI technology (reinforcement learning, transfer learning and generative AI) make the impossible inevitable, AI presents profound opportunities and risks for companies and societies.
The State of AI 2019 draws on unique data and discussions with ecosystem participants to go beyond the hype and explain the reality of AI today, what is to come and how to take advantage.
‘The State of AI 2019’ was developed in partnership with our friends at Barclays UK Ventures.
Good morning everybody.
My name is David Kelnar, I’m a Partner and Head of Research at MMC Ventures. We’re a research-led venture capital firm based in London. I have the privilege of leading a team at MMC called the Insights team, which is a research group at the Firm, and our goal is to proactively identify emerging areas of value creation, to understand them deeply, and then to identify, invest in and support the very best early-stage companies aligned to those themes.
William Gibson once wrote:
“The future is already here, it’s just unevenly distributed”.
How right he was.
In 2017 we launched our inaugural State of AI report, in which we highlighted that AI technology – that is, software that can improve with experience instead of just following sets of rules – was at a tipping point and poised to proliferate.
And so it has. Today, as we launch our State of AI report for 2019, we do so in the context of AI becoming mainstream. But as it does so, we see that a divide is emerging. Between nations, across industries and among companies, winners and losers are emerging in the race for adoption, the war for talent and the competition for value creation.
The landscape for Europe’s entrepreneurs is also changing. Europe’s 1,600 AI startups are maturing and bringing creative destruction to new industries. The socioeconomic map of Europe is also being redrawn.
And finally, as new AI technologies make the impossible inevitable, we face divergent futures. AI will have profoundly positive implications for companies and for society, but it also presents real risks. What future will we choose?
So, over the next 45 minutes, I’d like to walk you through the key findings from our State of AI report.
I’ll begin by highlighting how rapid adoption of AI overall is masking a divergence between leaders and laggards. We will see how in the war for talent, although supply is increasing, there’s still a dramatic imbalance in supply and demand.
We’ll see how emerging AI technologies, new hardware and new software, are freeing progress from the constraints of human experience. We’ll see how the landscape for Europe’s AI startups is changing. And, finally, we’ll consider some of these profound implications for companies and for society. So, let’s begin.
The race for adoption
Adoption of AI by large companies is rapidly increasing. Adoption of AI has tripled over the last 12 months. 12% of companies today have adopted AI at scale or in a large part of the organisation. And, in the next 12 months, a further quarter of large companies would have done so.
“Adoption of AI has tripled over the last 12 months.”
2019 is the year when AI ‘crosses the chasm’, from innovators and early adopters, through to the early mainstream.
There are several reasons behind this rapid proliferation of AI:
- widespread awareness of AI is encouraging a lot of test and learn initiatives;
- also, early proof of concept projects are now maturing and being deployed more widely in companies;
- the widespread availability of plug-and-play AI services from global technology vendors is reducing the cost to initiate and scale AI initiatives;
- companies are up-skilling their workforces and engaging with high-quality data scientists and chief science officers to try and address the talent shortfall; and
- companies are realising that there exists this incredible ecosystem of early-stage companies, which put AI at the heart of their value proposition, and these early stage companies can engage with them as suppliers.
Now AI is progressing, not just quickly but across a broad front. One in 10 companies today have 10 or more live AI applications. The most popular include chatbots, general process automation tools and fraud analytics. But many others remain widely popular. Consumer segmentation in marketing and assisted diagnostics in medicine are other use cases gaining rapid ground.
This rapid adoption of AI, however, masks a growing divide between leaders and laggards.
Globally, China leads the race in AI adoption. The proportion of companies in Asia-Pacific that have adopted AI is twice that of North America. Similarly, the proportion of companies in Asia-Pacific that have no intention of embracing AI is half that of North America. And specifically, it is China leading the race. Shanghai, Beijing and Guangdong — a lot of these centres are becoming truly global AI hubs. Chinese companies have a number of advantages:
- One of them is an ambitious long-term government policy for AI. China’s ‘next generation plan for AI’ is a long-term blueprint for global leadership in AI by 2030;
- But, so too, Chinese companies enjoy a double data advantage. The first is that more permissive policies regarding use of personal data make it easier for companies to use that data; and
- More broadly, Chinese companies data is typically less siloed. 78% of leading Chinese companies have all of their corporate data in a single unified data lake. That compares with just 43% of companies in Europe.
China’s rapid rise in AI has been a wake-up call for executives and industries globally.
Beyond the global map, we see that adoption of AI across different industries is uneven and in a state of flux.
We can divide industries into early adopters, movers, and laggards.
‘Early adopters’ are industries that proactively invested in AI, have continued that investment and are maintaining their leadership and reaping the rewards.
In 2017 insurance, financial service and high-tech companies highlighted that they intended to increase their spending on AI over the next three years more than any other companies. And so, today, insurance and software & IT service companies have the highest rates of adoption of AI.
“Insurance and software and IT service companies have the highest rates of adoption of AI.”
The second group is ‘movers’. These are industries that have awoken to AI’s potential. In 2017 retail, healthcare and media were in the middle of the pack in AI adoption . But they’re rapidly catching up. Today, four in 10 of companies in those three sectors will have significant live AI initiatives in 12 months’ time.
“Retail, healthcare and media… are rapidly catching up.”
Finally, we have the ‘laggards’. These are sectors falling behind in the race for AI adoption. This includes governments, education companies and charities. AI could transform governments given large datasets and numerous prediction and optimisation challenges. But lack of funding for emerging technologies, sprawling legacy IT estates and difficulty attracting world-class AI talent is holding that sector back.
Perhaps more troublingly, if we look at this group of laggards, there is a real risk that the more vulnerable members of society could be among the last to benefit from AI. And, if we think about slow pace of government adoption, it has the implication that we are all likely to engage with AI not as citizens but as consumers.
We’ve seen that there are significant divergences globally and across industries. But even within industries we seeing a growing gulf among companies’ understanding, learning and investment.
Just two out of 10 companies that are lagging in the adoption of AI say they understand the implications of AI for their company or for their sector. Just two out of 10 leaders say that they don’t.
And the smart are getting smarter. Again, just 20% of laggards say that over the last year they’ve improved their understanding of AI to a great extent, compared with two-thirds of pioneers.
“The smart are getting smarter.”
Perhaps most importantly, leaders are doubling down on AI and extending their advantage in this technology by increasing their investment at a greater pace. While just one in 5 laggards have increased their investment in AI significantly over the last year, nearly 90% of pioneering AI companies have done so.
So, while adoption of AI is growing overall, you can expect the gap between companies’ capabilities to widen.
The war for talent
Increasing demand for AI is dramatically increasing demand for AI talent – and it’s striking to see the dynamics around this. Job postings for AI, as a proportion of total job postings, have doubled in the last 24 months. So, demand for AI talent has never been greater.
“Job postings for AI, as a proportion of total job postings, have doubled in the last 24 months.”
Supply is increasing. Do you know the top emerging field of employment in the US, according to US government statistics? Machine learning. And the second from top is general data science. There are 10 times as many machine learning engineers today than there were five years ago.
“There are 10 times as many machine learning engineers today than there were five years ago.”
But the talent pool remains small. The trouble is that AI originated in academia and demands advanced competencies – in mathematics, computer science and programming. AI developers are seven times more likely to have a PhD than other developers. Because of this, the pool of available talent remains modest. As a result, the imbalance between supply and demand remains significant, with more than two roles available for every AI practitioner today.
This is creating some beneficiaries. First amongst them, of course, is the AI developers themselves. Albeit at the twenty highest paying companies globally, the median salary for an AI engineer is $224,000.
But there are some more significant winners and losers emerging in the war for talent.
The first is that the technology and financial services sectors are absorbing 60% of AI talent. There are 10 times as many developers in those fields as there are in healthcare — despite the potential for AI to transform healthcare. Technology and financial service firms were early adopters of AI, have maintained their commitment of resources, and are now are creating network effects around people to press their advantage in the decade ahead.
“The technology and financial services sectors are absorbing 60% of AI talent.”
The final dynamic is that the brain drain from academia to industry is real. The proportion of academic papers in AI that have a corporate affiliation has increased from 2% to 50% over the last decade. Now, there are many advantages to the movement of individuals into industry. They’re freed from the burden of creating and securing research grants. They can innovate with greater resource and at a greater pace. And, more broadly, the movement of academics into industry is catalysing the impact of AI in the real world.
“The brain drain from academia to industry is real.”
But so too there are some disadvantages just to be mindful of. For one thing, we may lack teachers for the next generation of practitioners. Secondly, we’re seeing the concentration of a lot of AI expertise in a small number of organisations. And thirdly, of course, this causes the movement of value from the public good to private gain. It’s worth remembering that AI arose from experimentation in academia. And if some presence in academia is not maintained, in the long-term society will suffer.
Th advance of AI technology
We’ve seen that demand for AI rapidly accelerating. And we’ve seen that supply of talent remains constrained. But AI technology itself, of course, has not been standing still. In particular, new advances in approaches to AI are freeing progress from some of the constraints of human knowledge.
I’ll touch on two approaches, and a new AI technique, that will have profound implications for society over the decade ahead.
It’s worth putting in context that traditional AI — so-called ‘supervised learning’ — relies on large amounts of training data. Systems evaluate a new input in the context of the training data they’ve received in the past. But what if we want to work in a domain where no training data is available? That’s where these new approaches to AI come alive.
Reinforcement Learning is an emerging alternative approach to AI where instead of amassing large amounts of training data, we specify just a goal. And through exploration, and trial and error, the system will try and achieve it. And we will reward the system for positive progress toward that goal.
Reinforcement Learning is a powerful way of enabling AI agents to interact with their environment. Amazingly, as far back as 1997 it was shown that if we provide a robot with knowledge only of the way in which its joints can move, and the directions in which it is able to move, the robot can learn to walk — without any knowledge of walking or any knowledge of its environment. And since then, and particularly in the last 24 months, progress in Reinforcement Learning has been remarkable.
“In the last 24 months, progress in Reinforcement Learning has been remarkable.”
Alpha Go Zero is a Reinforcement Learning-based AI, designed to play the ancient game of Go. Armed only with the rules of Go, and with no training data about prior games, by playing against itself within four days Alpha Go Zero was better than most human players and within 40 days was arguably the strongest Go player — human or artificial — that’s ever existed.
In the last 12 months in particular, and through 2019, we’re seeing progress now not just with individual agents, but with Reinforcement Learning enabling groups of agents to collaborate.
“Reinforcement Learning is enabling groups of agents to collaborate.”
This is a clip from a game called ‘Defence of the Ancients 2’ (DOTA 2), which is an online multiplayer game in which teams of five agents collaborate to compete against other teams.
OpenAI Five is an AI-based DOTA 2 team and, in the last 12 months, it’s demonstrated its ability to beat many human players. And indeed, at the end of last year, only the very best human players globally have been able to defeat it.
Reinforcement learning has numerous applications in cybersecurity. For example, it can be used to support an anomaly detection. Indeed, we recently backed an exciting early stage cybersecurity company called Senseon, which is using the latest advances in reinforcement learning to spot aberrant network behaviour.
It’s also of value in autonomous vehicles that have to adapt to unpredictable environments. And by enabling us to adapt to new domains, in the longer term it could be very valuable for Space Exploration. More broadly, what’s significant about reinforcement learning is that it potentially enables us to“achieve superhuman performance in really challenging domains with no human input”.
The second approach I’d like to talk to you about is Transfer Learning.
With traditional AI systems, we have to start from scratch every time, amassing training data and more, and that makes it slow and expensive.
With Transfer Learning, we apply learnings from a different — but similar — problem to the challenge at hand. And, by doing so, we can usually get better initial results, a faster pace of improvement, and better long-term outcomes. To use an analogy, I’ve never driven a bus, but having driven a car, I probably have a better shot at doing so.
Interest in Transfer Learning has grown sevenfold in the last 24 months, and for very good reason.
Interest in Transfer Learning has grown sevenfold in the last 24 months.
2018 was a breakthrough year in a variety of domains due to Transfer Learning. The first area was in the processing of natural language. Our understanding of natural language through AI has always been quite shallow. We can understand the meaning of words a bit, but less so sentences, and still less so paragraphs.
“Our understanding of natural language through AI has always been quite shallow. Transfer Learning is changing that.”
Transfer Learning is changing that. In 2018, a Transfer Learning program was able to achieve better results with a hundred examples of training data than a previous program had achieved with ten thousand. In a domain where advances are usually measured in fractions of a per cent, a Transfer Learning-based system could reduce error rates in use cases like answering questions and extracting entities by 25%.
Transfer Learning is also important is because it’s enabling real-world AI systems at scale. It’s very difficult, for example, to train a model on an industrial robot because it’s really slow. With Transfer Learning, we can simulate the behaviour of that robot in a computer, and then use Transfer Learning to determine what the differences might be in its performance in the model, and simulation, as opposed to real life. Progress in areas like autonomous vehicles may very well depend on the dramatic advances that we’re seeing in transfer learning.
“Progress in areas like autonomous vehicles may.. depend on the dramatic advances that we’re seeing in transfer learning.”
Finally transfer learning is significant as a step towards artificial general intelligence, or AGI. AGI is the idea that we can create an AI that could perform pretty much any intellectual task that a human does. We are years away from AGI. But, by enabling systems with greater adaptability, Transfer Learning is an important step along the way. Indeed, transfer learning may be the key to us achieving AGI one day.
“We are years away from AGI. But…Transfer Learning is an important step along the way.
Finally, in terms of emerging technology, generative AI will transform media in the years ahead. Generative Adversarial Networks, or GANs are a novel AI technique that enables the creation of lifelike pictures and media.
Developments in GANs in recent years have enabled the creation of media so lifelike, it’s virtually impossible to distinguish from what’s real.
None of these people ever existed. All of these pictures were hallucinated by a Generative Adversarial Network, and developments continue to be rapid.
“Developments in GANs in recent years have enabled the creation of media so lifelike, it’s virtually impossible to distinguish from what’s real.”
Until recently, one drawback of GANs was that it was quite hard to guide the outcome, more specifically, that you wanted. Even that’s now changing and we can create incredibly convincing media, increasingly according to specification.
GANs will transform the media sector in the years ahead. The ability for companies to create lifelike-looking content, from inception, at low-cost and scale will effectively democratise the creation of content.
And GANs can do more. For example, we can splice onto video new audio, and resynchronise the speaker’s lips to that audio. So, an agency could take existing footage of a brand ambassador and adapt their speech and video — so they speak in one of 30 languages for overseas markets.
Now as well as having many benefits, GANs have some risks for society. I’ll talk about those a little bit later.
The disruptors: Europe’s AI startups
Every paradigm shift in technology unleashes a wave of new, innovative early-stage companies that put that technology at the heart of their value proposition. AI is no different.
We individually reviewed more than 2,800 purported AI startups across Europe, in the 13 EU countries most active in AI that together comprise 90% of EU GDP. I’m excited to share with you some unique findings from that research.
“We individually reviewed more than 2,800 purported AI startups across Europe.”
Firstly, mind the gap! About 40% of the companies that we looked at showed no material evidence of AI being core to their value proposition. So we distilled a map of the 1,600 disruptors in Europe driving entrepreneurship in this sector in Europe – and found some remarkable things.
Secondly, AI entrepreneurship is beginning to go mainstream. In 2013, just one in 50 startups had embraced AI. Today, one in 12 startups in Europe is an AI startup. Entrepreneurship is proliferating as companies take advantage of triggers for entrepreneurship, as well as general enablers of AI.
“Today, one in 12 startups in Europe is an AI startup.”
Third, the ecosystem is maturing. Just two years ago, only one in 20 startups had progressed from the angel and seed stages of their funding through to the later, growth stages of their journey. Today one in six of Europe’s AI startups are increasingly mature. The maturing of the ecosystem will have profound implications. Expect to see:
- a growing number of exits, as AI startups achieve critical mass and are required by incumbents;
- the recycling of capital and talent as those exits occur;
- some high-profile failures, sadly, as well capitalised companies falter;
- greater competition — not just between startups and incumbents but between startups and scale ups; and more.
“Today, one in six of Europe’s AI startups is increasingly mature.”
I’d like to take you on a brief tour of some European countries to highlight to you some of the national dynamics in European AI.
The first is that the UK is the powerhouse of European AI, with the third of the Continent’s AI startups and more than twice as many as the next most active country. The UK is drawing on many assets — including:
- the largest Internet economy in the G20 (as a percentage of GDP);
- a quarter of the world’s top 25 universities as a hub for talent;
- a growing roster of successful AI exits (including SwiftKey, Deepmind, Magic Pony, and more);
- many successful scale ups; and
- one the world’s great financial service hubs.
In the report we provide a functional map of the UK’s nearly 500 AI startups and feature 16 of, we think, the UK’s most exciting early-stage AI companies.
“The UK is the powerhouse of European AI, with the third of the Continent’s AI startups and more than twice as many as the next most active country.”
Germany and France are flourishing AI hubs, with a growing roster of successful AI scale-ups, high quality talent, and growing venture investment. Indeed, venture capital inflows to these countries in 2018 increased 39% and 29% year-on-year respectively.
“Germany and France are flourishing AI hubs.”
Spain’s contribution exceeds its size. Despite having a population half the size of Germany, Spain has nearly as many AI startups. Interestingly, Spain has the second highest level of immigration in EU — and immigration is correlated with entrepreneurship. So, it may be that extensive immigration in Spain has deepened the already broad pool of talent in the country.
And finally, Italy and Sweden punch above their weight in ‘core technology’. By core technology, I mean AI companies that don’t serve a particular sector or business function, but instead are providing a horizontal AI technology — maybe computer vision or language or autonomous systems. While the average across Europe is for countries to have one in ten of their startups focusing on core tech, in Italy and in Sweden it’s one in five.
For Sweden, that might be because of a ‘halo effect’ of broader technological activity. Sweden is known as the ‘unicorn factory’ for the number of amazing early stage companies that grow from it, and it’s likely that the success of these companies serves as powerful hub for attracting talent.
This is an ecosystem in flux. What we’ve seen is that while the UK is the powerhouse of European AI, other countries may extend their influence in the decade ahead. In the last five years, the UK’s share of Europe’s AI startups (by volume) has very slightly reduced. Brexit could accelerate this trend. We’ve seen that AI practitioners are highly skilled, few in number and can take their pick from the many opportunities they receive. If free movement ends, if low-friction visa programs are not forthcoming, or if rhetoric is unwelcoming, the UK’s access to talent could become inhibited. Germany, France and other countries may extend their influence in future.
“The UK’s share of Europe’s AI startups (by volume) has very slightly reduced. Brexit could accelerate this trend.”
We can also see, from a sector analysis of Europe’s 1,600 AI startups, that entrepreneurs’ focus is highly uneven. One point I’d highlight is that healthcare has become a focal point for AI entrepreneurship, with one in five companies that are focused on a sector focusing on the healthcare domain.
There are good reasons for this:
- AI offers profound opportunities for process automation and cost reduction in healthcare, by enabling us to do in software, for the first time, what in the past could only be done by people;
- There’s also incredible need. European countries’ spending on health care, as a percentage of GDP, has doubled to 10% and systems are a breaking point;
- Thirdly, there is increasing openness to innovation from healthcare stakeholders. Indeed, just in October last year the UK government published its vision paper for the development of the NHS, which proactively put AI and emerging technologies at its heart; and
- Finally, there’s a new cohort of bold entrepreneurs seeking to enact change at a systemic level.
We also found that AI companies have unique capital dynamics. AI startups attract more money, and at higher valuations, than other software companies. In terms of the amount of money they raise, the difference can be between 15% and 50% over time and, as a result, valuations tend to be higher as well. This is because of both good fundamental reasons and also some practical reasons.
AI startups attract more money, and at higher valuations, than other software companies.
The fundamental reason is that the journey to a minimum viable product for an AI company can be long, given the technical demands of developing AI. And we’ve seen that the cost of AI talent is very high, so AI companies do need to make sure they’re well capitalised.
But at the same time there’s a bit of an effect here of the imbalance of the supply and demand for capital in the space. There’s a lot of venture capitalist interest in AI and a relatively small number of investable prospects, and as a result that dynamic is increasing valuations.
What’s striking is: that trend has declined over time, as AI entrepreneurship has gone more mainstream. I would expect that difference in valuation and round sizes to continue to mitigate and decline in the years ahead.
Finally, we spoke with hundreds of entrepreneurs in the UK, and beyond, to understand their challenges. What we found is that three were consistent:
- firstly, ability to attract and retain world-class AI talent;
- secondly, the ability to develop the training data they need for their systems; and
- the difficulty of productising AI of moving AI from the lab to live.
“We spoke with hundreds of entrepreneurs in the UK, and beyond, to understand their challenges: talent, data and productising AI.”
To help with this we’ve developed our AI Playbook, which is a blueprint for how companies can develop and deploy AI. It covers the six core competencies for building an AI capability — from strategy through to people, data, development, production, and regulation & ethics. It’s available via our website.
To help with this we’ve developed our AI Playbook, which is a blueprint for how companies can develop and deploy AI.
Our AI future: the implications of AI
I’d like to finish by touching on some of the profound implications of AI for companies and society.
It’s a little easier to understand the implications of AI if we first abstract its benefits. AI offers:
- innovation: that is, new products and services. Sometimes, AI makes the previously impossible possible — as with autonomous vehicles.
- secondly AI can offer efficacy: enabling us to do what we could do before, but to a higher standard.
- in other cases, AI provides velocity: the ability to do what we could do before, but do it faster; and finally
- scalability: by enabling us to do in software what previously only people could do, we break free from the constraints of human capacity.
It’s a little easier to understand the implications of AI if we first abstract its benefits.”
When we understand these benefits at the heart of AI, I think its implications become a little bit clearer.
Firstly, for companies, expect to see an influx of new market participants. Historically access to a range of sectors from healthcare to transport has been limited to subsets of the global population. Consider healthcare and receiving a medical diagnosis. In the West, receiving medical care is expensive or inconvenient. But in developing economies it can often be non-existent. Why? Because receiving a medical diagnosis requires an experienced professional. Training that professional takes time and money. And because GP’s can’t be in more than one place at once, access is limited as well. But we’re entering the era of automated medical diagnosis where, at least for a range of conditions, it should be possible to make an automated diagnosis. Of course, the marginal cost of an AI diagnosis will be zero. Access won’t be a problem either. By the end of next year, 60% of, for example, the population of Africa will have a smartphone. The combination of this collapse in cost, and increase in access, will result in an influx of new market participants — individuals, who can access these services for the very first time. That will have profound implications for those individuals, for the companies providing these services, and of course for the value chains that they’re disrupting.
“A collapse in cost, and increasing access, will result in an influx of new market participants.”
A second implication will be: shifts in sector value chains. By that, I mean where and the extent to which profits are available in a given industry. 42% of all insurance premium revenue globally, of any kind, is from car insurance. But in the decade ahead we’re going to see the rise of autonomous vehicles. And with the rise of autonomous vehicles we’ll see a fall in the frequency and severity of road accidents — which is remarkable. But alongside that we will see a collapse in the price and profitability of car insurance. In fact, it’s forecast that in the UK, the profitability from car insurance policies will fall by up to 80% in the years ahead. How will the insurance industry adapt the loss of 80% of its profitability from 40% of its revenue? Expect to see similar profound shifts in sector value chains elsewhere.
“How will the insurance industry adapt to the loss of 80% of its profitability from 40% of its revenue?”
Thirdly, expect new competitive success factors. Every paradigm shift in technology provides new competencies but demands new capabilities, and so too in the age of AI. We can expect to see certain companies extend their advantage by developing the right competencies. In the age of AI these will include:
- the organisational vision to embrace AI;
- use of large proprietary data sets to train machine learning algorithms;
- reshaping of organisational pyramids to attract data scientists and other AI professionals;
- a mindset shift from software that gives binary outcomes to probabilistic recommendations;
- the need to be able to diligence AI suppliers; and more.
“Every paradigm shift in technology provides new competencies but demands new capabilities.”
Companies that can master these success factors the age of AI will thrive. Those that can’t will lose market share and, I think, relevance surprisingly quickly.
Finally, with regard to companies’ new business models, the combination of AI, subscription billing models, and ‘X-as-a-service’ will obviate a range of old business models and create new ones. In particular, AI will transform the economic fabric of car ownership and insurance. Did you know cars spend 96% of their lives parked and unused? It’s crazy, but of course private vehicle ownership has historically been the only way that we’ve been able to have the kind of spontaneity and security and privacy that we desire. But, again, the rise of autonomous vehicles will change all of that. With the cost of the driver removed, and the cost of the vehicle and fuel more effectively amortised over a greater number of journeys in a given period, the marginal cost of a car journey will fall.
‘The combination of AI, subscription billing models, and ‘X-as-a-service’ will obviate a range of old business models and create new ones.”
I think you can expect to see the emergence of ‘transport-as-a-service’, where large fleets of autonomous vehicles are summoned by us, from our smartphone, and used on a point-to-point basis and probably paid for with a fixed low monthly fee for reasonable use. This will have profound implications — not just, of course, for car manufacturers themselves (as we see private vehicle ownership decline) but also for downstream market participants. The nature and distribution of petrol or charging stations, and local repair centres, could all change in the era of autonomous fleets.
We’ve seen several implications that AI can have for companies. But let me touch on the implications of AI for society.
AI will have innumerable benefits for society, including: improved healthcare; increased industrial and agricultural output; greater access to transport; more personalised retail experiences; greater day-to-day convenience and …the list goes on.
But so too there are real risks. And it’s important that these are identified and mitigated.
To me, one the most important is the risk that biased AI systems could increase social inequality.
In theory, AI systems can free us from the constraints of human bias by drawing objective patterns from large amounts of data. The trouble is that AI systems are trained using large amounts of training data, and that training data reflects decades of historic bias, particularly regarding race and gender.
In a training data set used to train a lot of the most popular facial recognition systems, 75% of faces are male and 80% that are white. These kinds of incomplete or problematic datasets are causing AI systems to behave problematically, particularly in relation to minorities. To give an example, popular facial recognition systems that offer gender classification misgender just 1% of lighter-skinned men, but misgender 35% of darker-skinned females.
“In a training data set used to train a lot of the most popular facial recognition systems, 75% of faces are male and 80% that are white.”
AI systems are being used to make decisions that have significant ramifications for people’s lives, in areas ranging from recruitment through to credit. And if we fail to eliminate bias from AI systems, individuals can suffer economic loss, lots of opportunity, and social stigmatisation. Put another way, “if we fail to make ethical and inclusive AI we risk losing gains made in civil rights and gender equality under the guise of machine neutrality”.
And finally, artificial media could undermine trust. We saw that generative AI offers the ability to create increasingly lifelike pictures and video, and that recent developments in generative AI make it virtually impossible to distinguish that from real life. That will have some troubling implications. One of these two videos is fake. I’m going to play the video to you and ask you to put your hand up for which one. Ready?
Want to see it again? Put your hand up if you think video of Theresa May is fake. Okay, very roughly I’d say about 65% of the audience. Put your hand up if think it’s Vladimir Putin. About a third. Okay, two thirds of you were right. The video of Theresa May was artificial. But the very fact that it’s this hard to tell speaks volumes.
Generative media will supercharge the development of fake news. But I think the longer-term implications are perhaps the more insidious and need to be carefully managed. Adversaries have long since realised that, over time, more effective than direct action against an enemy, is to sow confusion and discord. With generative AI, in an era where any media could be artificial, all media is open to challenge. How will we know what to believe?
“With generative AI, in an era where any media could be artificial, all media is open to challenge. How will we know what to believe?”
In the book ‘Nineteen eight-four’, George Orwell described a totalitarian state which persecuted independent thought. The Party ‘told you to reject the evidence of your eyes and ears. It was their final, most essential command’. Ironically, in the future that may be necessary, and with the next generation of AI society will grapple with challenges of truth and trust.
So, let’s sum up.
We’ve seen that:
- growing adoption of AI, overall, is masking winners and losers in the race for adoption, and the war for talent;
- incredible new technologies — reinforcement learning, transfer learning and generative AI — are freeing progress from the constraints of human knowledge and enabling breakthroughs with far-reaching implications;
- AI entrepreneurship is coming of age, and proliferating, and that we’re entering the age of healthcare AI; and
- AI will have powerful implications, reshaping companies’ business models and sector value chains, and with many benefits as well as risks for society.
I’ll finish with a final thought. 24 months ago, an AI system called AlphaGo, which was a predecessor to AlphaGo Zero that I described earlier, was playing a five-game challenge match against perhaps the world’s greatest Go player: Lee Sedol. In game two, on move 37, something truly extraordinary happened. AlphaGo made a move so unexpected, so brilliant, so mysterious, and so bold that it had probably never been considered before in the two-and-a-half thousand-year history of the game. Move 37 will be studied for years to come. AlphaGo went on to win that game and series.
Just watch the expression of Lee Sedol after that move is played. The chap you’re seeing here is the human operator, about to place the tile on behalf of AlphaGo. The second person you’ll see will be Lee Sedol.
In the 20th Century we got used to computers learning from us. In the 21st Century, as AI presents new ideas, new ways of thinking and new aesthetics, we will start to learn from them. Artificial intelligence slowly reshaping organic intelligence might be the greatest divergence in human history.