Much of the early public discussion about AI focused on what AI could do better than humans and where it would replace human labor. The same thing happened with the emergence of computers in the 1960’s, personal computers in the 1970’s and the internet in the 1990’s. This line of thinking is a natural human response to new technologies perceived as a threat, but it misses the far more important point that the big breakthroughs in productivity and intelligence come from human-AI, human-internet and human-computer teaming.
In this podcast discussion with Gautam Mukunda and Shawn Bice, we had a chance to dig into what the collaborative human-AI intelligence of the future will look like, how AI will change human decision-making and how it will impact employment, science and technology innovation. You can listen to the discussion here.
Of all of the spheres of human activity that are being radically transformed by technology, one of the most consequential is geopolitics. Advances in technology are rapidly increasing national security vulnerabilities, changing the way countries compete and requiring a radical rethink of defense and intelligence strategies. Although the U.S. continues to lead the world in technology innovation, it has been surprisingly behind the curve in this transformation.
The world is in a vulnerable state, and I believe the CFR has never been more needed. Beyond the current pandemic, the rules-based international order established by the U.S. and its allies after World War II is rapidly deteriorating. Democracy is in retreat, and autocracy, nationalism and instability are on the rise. Global problems beyond the reach of individual countries to solve on their own are also on the rise, including climate change, pandemics and cyber threats. It’s time for new ideas about how the U.S. should address these issues and how we can best harness technology to make the world a safer place. I believe the CFR can contribute meaningfully to these discussions and am looking forward getting more involved.
If you are interested in contributing to these discussions, I would encourage you to become a member.
One of the most interesting things I’ve worked on this past year is a project at the Council on Foreign Relations examining the risk and consequences of the U.S. losing part of its science and technology leadership, principally to China.
We had a phenomenal group of people working on our Task Force from the technology, government and academic communities, including Admiral William McRaven, James Manyika, Reid Hoffman, DJ Patil, Eric Schmidt, Raj Shah, Doug Beck and Regina Dugan.
Today our final report was published – you can download it here. For a 2-minute overview, here’s an article I wrote with Congressman Jim Himes that highlights the report’s major ideas and recommendations.
There’s a new economic force at work in the machine learning revolution that is capable of generating increasing returns to scale, much as network effects did in the internet revolution.
This force is automated learning, and its business impact comes in the form of learning effects: the more a product learns, the more valuable it becomes.
Learning effects have the potential to generate enormous economic value, as network effects do, if companies are able to close this loop and make it self-reinforcing: that is, if their products learn more because they have become more valuable.
This happens when more valuable products attract more users or customers, who provide more and richer data of the kind that enables machine learning models to make these products more valuable still, which attracts more users or customers still, and so on, creating a self-perpetuating cycle.
.
Just as network effects determined who the biggest winners of the internet revolution were, learning effects will determine who the biggest winners of the machine learning revolution will be.
Because they enable increasing returns to scale, they will similarly give rise to a set of companies that become runaway leaders – that are capable of pulling away from their competitors and continuing to increase their leads over time.
Offline Origins
Like network effects, learning effects have always existed in the offline world but have become supercharged in the digital world. In the offline world, learning effects are transmitted through humans: as people learn how a product can become more valuable, they modify it accordingly. Human learning, however, is artisanal, and artisanal learning only scales so quickly.
What’s new and different in the machine learning era is that certain kinds of learning have become automated. Software can learn by itself with exposure to new data and become more valuable in the process. This is a big deal economically. It involves the unlocking of new source of economic value that was previously inaccessible.
A Vast New Power
Learning effects have taken off most significantly at large internet platforms given the immense amount of data they control and their aggressive investment in machine learning to accelerate product innovation: Google in search, ads, photos, translate and Waze; Facebook in search, ads and newsfeed; and Amazon in search, ads, product recommendations and Alexa, to name just a subset for each. These companies recognize that machine learning has granted them a vast new power, and they are eager to take maximum advantage of it.
Perhaps the best pure-play example of the power of learning effects is Tesla, which began as an electric car company but was able to deploy machine learning to extraordinary effect across its fleet to become the category leader in autonomous driving. Tesla’s autonomous driving capabilities make its cars more valuable, which attracts more customers and data, which enables it to improve these capabilities further and attract even more customers, and so on. As a result of its learning effects, Tesla’s rate of innovation and value creation in the autonomous driving area have dwarfed what its competitors have been capable of.
Engineered Growth
Network effects and learning effects generate growth in different ways. Network effects tend to generate growth organically through a kind of gravitational accretion, as individual consumers and businesses pursuing their own self-interest decide to join the largest and most valuable networks, making them larger and more valuable still.
Learning effects similarly benefit from consumers and businesses pursuing their own self-interest to purchase the best products, but they are less the result of gravitational accretion than of finely tuned technology and product development efforts that require constant intervention and recalibration in order to tie together data, intelligence, product innovation and user/customer growth.
As a result, even though learning effects are partially the product of automated learning, they are by no means automatic. The data generated from new customers must be of the right kind and of sufficient volume to enable new learning. This learning must be optimized effectively enough to create new product value. And this value must be strong enough and productized well enough to attract more customers. Any break in this chain means there is no self-reinforcing cycle and hence no learning effects.
Runaway Leaders
Perhaps the most interesting question about learning effects is what are the conditions that make them strong enough to create runaway leaders, as these are the companies that tend to create the vast bulk of enterprise value in the technology startup world.
Learning effects don’t always produce runaway leaders. Just because one company has a head start in learning doesn’t mean other companies can’t acquire more or better data or learn more efficiently from similar data to catch up with them and eventually bypass them. It’s an interesting question today, for example, if Tesla is pulling away from the pack in autonomous driving, or if others will catch up in the years ahead.
In order for learning effects to produce runaway leaders, a company must secure a definitive advantage over its competitors in one of the component areas of learning effects – data, intelligence, product innovation or user/customer growth – and leverage this into advantages in the others, such that the company can acquire data, learn, innovate and grow not only more rapidly than its competitors do, but more rapidly than they can.
As with learning effects generally, there is nothing automatic about tying these advantages together. It requires excellent execution.
Typically a company is able to jumpstart this cycle by developing a significant data advantage over its competitors. It then must translate this data advantage into an intelligence advantage as measured by the capabilities of its machine learning models, which requires that its models be as or more efficient than those of its competitors. This intelligence advantage must then tie to a product innovation advantage that is directly correlated with a user or customer acquisition advantage and ultimately with an advantage in the size of its user or customer base. Enough customers have to want to buy Teslas, in other words, because of their autonomous driving capabilities vs. because it’s an electric or cool-looking car, as that doesn’t create a strong enough self-reinforcing cycle. Finally, this user or customer base advantage must enhance the company’s data advantage in the right way to generate additional learning.
.
Generally the narrower the scope of a product and the greater the degree to which machine learning drives its value, the easier it is to tie these advantages together to create a runaway leader.
Wherever it is possible to tie these advantages together, there will likely be ferocious competition, as with network effects, for startups to get an initial head start in competing for scale to achieve escape velocity and become runaway leaders given the huge premium on winning. The early bird that capitalizes on its head start generally gets all the worms. Other birds need to bootstrap alternative advantages in the form of more efficient learning engines or access to large and differentiated datasets in order to have a chance.
Learning Curves: Long, Steep and Perpetual
In order for runaway leaders to be able to maintain their leads over time, there’s an important additional requirement, which is that the learning curves for their products must be long enough and steep enough to enable them to provide increasing product value for an extended period. If the learning curves for their products are short or top off quickly, early leaders will max out on them while they still have viable competitors, and these competitors will be able to catch up. If the learning curves are long and steep, on the other hand, these companies will have sufficient runway to break away from their competitors and maintain their leads over time.
Certain products – particularly those built on highly dynamic datasets – may have perpetual learning curves such that in a rapidly changing world, they can always be meaningfully improved. It’s around these kinds of products that the most valuable runaway leaders will likely develop. Potential examples include search, semantic engines, adaptive autonomous systems and applications requiring a comprehensive real-time understanding of the world.
The Interaction of Learning Effects and Network Effects
Network effects almost always create the opportunity for learning effects, as they involve the generation of ever more data in the form of new network members and interactions. Companies must invest in machine learning to create these learning effects, and they may or may not be successful. They may fail to generate meaningful learning, or they may generate meaningful learning but not learning effects if this learning does not result in more valuable products that lead to the continual acquisition of new data for additional learning.
Conversely, learning effects can create network effects. Tesla, for example, did not benefit from network effects when it was just an electric car company and was not yet focused on autonomous driving. However, once the company outfitted its cars with information sensors to develop autonomous driving capabilities through machine learning, it suddenly began to benefit from network effects: each Tesla became more valuable the larger the fleet became.
Importantly, however, when learning effects create network effects, these network effects do not exist independently of them. They are in effect an expression of the learning effects: learning just happens to take place through a network. If Tesla turned off its machine learning, its network effects would cease to exist.
The reverse, however, is not true. Network effects can give rise to learning effects that can exist independently of them. Facebook’s core network effect of people wanting to be part of the same social network that their friends are, for example, generates lots of new data that machine learning models can learn from. One area where Facebook has invested significantly in machine learning and succeeded in generating learning effects is improving the relevance of its newsfeed. Newsfeed relevance is a different kind of value than the core value around which the company’s network effects are based, although the two clearly reinforce each other. If Facebook stopped growing its user baser, it could continue to generate increasing value by improving the relevance of its newsfeed through these learning effects.
Since network effects and learning effects are both functions of customer value, whenever they exist side by side in a product, they always reinforce each other, as each makes the product more valuable in a way that attracts more customers and data.
The most formidable kinds of runaway leaders that tend most strongly toward natural monopoly – Facebook and Google are excellent examples – are those that benefit from network effects and learning effects working in tandem, as their mutual reinforcement means these companies run away from the pack much faster and are generally impossible to catch, provided they also benefit from perpetual learning curves.
Startups vs. Incumbents
Incumbent internet platforms have unsurprisingly been the big winners of the machine learning revolution to date because of their vast data assets and their significant investment in this new technology. Their early dominance has led skeptics to wonder if machine learning is a game that startups can win at all given their relative data disadvantages.
There are huge new datasets and data-rich applications created every day, however, in domains where these and other platforms have little or no presence, which provide an abundance of new opportunities for startups.
In addition, there are many large datasets sitting in organizations that startups are best suited to access because they are better able to provide these organizations with innovative applications to take advantage of them.
And although startups make lack the early edge in data, they always have the advantages of focus and adaptability. Where I believe these advantages will make the biggest difference in machine learning is that machine learning applications are engines, and startups have the ability to build and tune these engines most precisely to maximize learning effects. They have the ability not only to maximize the amount of learning and hence value they create from new data, but to complete this loop and maximize the amount of data in the form of new customers they create from new learning.
Only by constantly tightening and amplifying these loops can companies grow rapidly from learning effects and hope to achieve escape velocity to become runaway leaders. As a general rule, startups tend to be better at this than incumbents.
This article was originally published on Techcrunch
If there is one technology that promises to change the world more than any other over the next several decades, it is arguably machine learning. By enabling computers to learn certain things more efficiently than humans and discover certain things that humans cannot, machine learning promises to bring increasing intelligence to software everywhere and enable computers to develop ever new capabilities – from driving cars to diagnosing disease – that were previously thought impossible.
While most of the core algorithms that drive machine learning have been around for decades, what has magnified its promise so dramatically in recent years is the extraordinary growth of the two fuels that power these algorithms – data and computing power. Both continue to grow at exponential rates, suggesting that machine learning is at the beginning of a very long and productive run.
As revolutionary as machine learning will be, its impact will be highly asymmetric. While most machine learning algorithms, libraries and tools are in the public domain and computing power is a widely available commodity, data ownership is highly concentrated.
This means that machine learning will likely have a profound barbell effect on the technology landscape. On one hand, it will democratize basic intelligence through the commoditization and diffusion of services such as image recognition and translation into software broadly. On the other, it will concentrate higher-order intelligence in the hands of a relatively small number of incumbents that control the lion’s share of their industry’s data.
For startups seeking to take advantage of the machine learning revolution, this barbell effect is a helpful lens to look for the biggest business opportunities. While there will be many new kinds of startups that machine learning will enable, the most promising will likely cluster around the incumbent end of the barbell.
Democratization of Basic Intelligence
One of machine learning’s most lasting areas of impact will be to democratize basic intelligence through the commoditization of an increasingly sophisticated set of semantic and analytic services, most of which will be offered for free, enabling step-function changes in software capabilities. These services today include image recognition, translation and natural language processing and will ultimately include more advanced forms of interpretation and reasoning.
Software will become smarter, more anticipatory and more personalized, and we will increasingly be able to access it through whatever interface we prefer – chat, voice, mobile application, web, or others yet to be developed. Beneficiaries will include technology developers and users of all kinds.
This burst of new intelligent services will give rise to a boom in new startups that use them to create new products and services that weren’t previously cost effective or possible. Image recognition, for example, will enable new kinds of visual shopping applications. Facial recognition will enable new kinds of authentication and security applications. Analytic applications will grow ever more sophisticated in their ability to identify meaningful patterns and predict outcomes.
Startups that end up competing directly with this new set of intelligent services will be in a difficult spot. Competition in machine learning can be close to perfect, wiping out any potential margin, and it is unlikely many startups will be able to acquire data sets to match Google or other consumer platforms for the services they offer. Some of these startups may be bought for the asset values of their teams and technologies (which at the moment are quite high), but most will have to change tack in order to survive.
This end of the barbell effect is being accelerated by open source efforts such as OpenAI as well as by the decision of large consumer platforms, led by Google with TensorFlow, to open source their artificial intelligence software and offer machine learning-driven services for free, as a means of both selling additional products and acquiring additional data.
Concentration of Higher-Order Intelligence
At the other end of the barbell, machine learning will have a deeply monopoly-inducing or monopoly-enhancing effect, enabling companies that have or have access to highly differentiated data sets to develop capabilities that are difficult or impossible for others to develop.
The primary beneficiaries at this end of the spectrum will be the same large consumer platforms offering free services such as Google, as well as other enterprises in concentrated industries that have highly differentiated data sets.
Large consumer platforms already use machine learning to take advantage of their immense proprietary data to power core competencies in ways that others cannot replicate – Google with search, Facebook with its newsfeed, Netflix with recommendations and Amazon with pricing.
Incumbents with large proprietary data sets in more traditional industries are beginning to follow suit. Financial services firms, for example, are beginning to use machine learning to take advantage of their data to deepen core competencies in areas such as fraud detection, and ultimately they will seek to do so in underwriting as well. Retail companies will seek to use machine learning in areas such as segmentation, pricing and recommendations and healthcare providers in diagnosis.
Most large enterprises, however, will not be able to develop these machine learning-driven competencies on their own. This opens an interesting third set of beneficiaries at the incumbent end of the barbell: startups that develop machine learning-driven services in partnership with large incumbents based on these incumbents’ data.
Where the Biggest Startup Opportunities Are
The most successful machine learning startups will likely result from creative partnerships and customer relationships at this end of the barbell. The magic ingredient for creating revolutionary new machine learning services is extraordinarily large and rich data sets. Proprietary algorithms can help, but they are secondary in importance to the data sets themselves. The magic ingredient for making these services highly defensible is privileged access to these data sets. If possession is nine tenths of the law, privileged access to dominant industry data sets is at least half the ballgame in developing the most valuable machine learning services.
The dramatic rise of Google provides a glimpse into what this kind of privileged access can enable. What allowed Google to rapidly take over the search market was not primarily its PageRank algorithm or clean interface, but these factors in combination with its early access to the data sets of AOL and Yahoo, which enabled it to train its algorithms on the best available data on the planet and become substantially better at determining search relevance than any other product. Google ultimately chose to use this capability to compete directly with its partners, a playbook that is unlikely to be possible today since most consumer platforms have learned from this example and put legal barriers in place to prevent it from happening to them.
There are, however, a number of successful playbooks to create more durable data partnerships with incumbents. In consumer industries dominated by large platform players, the winning playbook in recent years has been to partner with one or ideally multiple platforms to provide solutions for enterprise customers that the platforms were not planning (or, due to the cross-platform nature of the solutions, were not able) to provide on their own, as companies such as Sprinklr, Hootsuite and Dataminr have done. The benefits to platforms in these partnerships include new revenue streams, new learning about their data capabilities and broader enterprise dependency on their data sets.
In concentrated industries dominated not by platforms but by a cluster of more traditional enterprises, the most successful playbook has been to offer data-intensive software or advertising solutions that provide access to incumbents’ customer data, as Palantir, IBM Watson, Fair Isaac, AppNexus and Intent Media have done. If a company gets access to the data of a significant share of incumbents, it will be able to create products and services that will be difficult for others to replicate.
New playbooks are continuing to emerge, including creating strategic products for incumbents or using exclusive data leases in exchange for the right to use incumbents’ data to develop non-competitive offerings.
Of course the best playbook of all where possible is for startups to grow fast enough and generate sufficiently large data sets in new markets to become incumbents themselves and forego dependencies on others, as for example Tesla has done for the emerging field of autonomous driving. This tends to be the exception rather than the rule, however, which means most machine learning startups need to look to partnerships or large customers to achieve defensibility and scale.
Machine learning startups should be particularly creative when it comes to exploring partnership structures as well as financial arrangements to govern them – including discounts, revenue shares, performance-based warrants and strategic investments. In a world where large data sets are becoming increasingly valuable to outside parties, it is likely that such structures and arrangements will continue to evolve rapidly.
Perhaps most importantly, startups seeking to take advantage of the machine learning revolution should move quickly, because many top technology entrepreneurs have woken up to the scale of the business opportunities this revolution creates, and there is a significant first-mover advantage to get access to the most attractive data sets.
