Platforms
Many internet-age companies have scaled to gargantuan sizes. 8 of the top 10 companies by market-cap are big tech, up from 2 just two decades back. For the successful ones, this scale-up happens far quicker than classical business sense would admit. Consider Instagram; when it was acquired for 1 billion USD by Facebook it had a 13-member team who had been building a photo sharing platform for just two years.
An effective lens to study such companies is to see them as platforms, or as aggregators as Ben Thompson puts them. Ben characterizes aggregators with three defining properties: (a) a direct relationship with users, (b) zero marginal costs of serving users, and (c) demand-driven multi-sided networks with decreasing acquisition costs. Nobel Prize winning economist Jean Tirole has also formally characterized such platforms as two-sided markets where one side can be charged low or even negative prices. In essence, such platforms directly interface with users who they often charge nothing but from whose presence they profit by interfacing with other businesses. As an example, Google charges nothing for making the entire web searchable while earning from positioning third party ads. This is fantastic value creation that is enabled by the internet which has driven cost of goods sold (COGS) close to zero while opening a global market.
Consequently, many of us are massive consumers of free services from platforms such as Gmail, WhatsApp, Facebook, Google search, YouTube, Twitter, and Instagram. And because these are compelling deals, a large number of our friends, family, colleagues, influencers, intellectuals, entertainers, and politicians are also on the same platforms, making a transition to an alternative hard. For instance, in January 2021, when WhatsApp announced a change in their privacy policies, there was a lot of hullabaloo about migrating to alternatives such as Signal. But in the end, there was no meaningful dent to WhatsApp’s user base, which stands today at a staggering half a billion in India alone.
As these platforms have scaled in response to and in anticipation of user demand, many techno-social challenges have emerged. Most of these revolve around the copious amount of user data these platforms collect and use, sometimes in ways that aren’t entirely transparent. Much worry has been expressed about this data profiling by governments and intellectuals. As Shoshana Zuboff writes in her impassioned book “Our lives are rendered as data flows”. A second major issue has been questionable ways of harvesting human attention. These platforms are incentivized to promote stronger and divisive emotions to maximize engagement. Anger is one of these divisive emotions as Scott Alexander characterizes in his essay toxoplasma of rage. Finally, there have also been concerns about biased censorship and the potential for moral policing and political manipulation. For instance, Twitter censored a story of national interest in the US from a news agency that is over 200 years old, only to reverse its decision after the political sensitivity of the story had died down.
In sum, there is widespread worry about the effect of large platforms to both social institutions and individuals. This worry from an American point of view is well summarized by the title of a recent article in The Atlantic- “Why the past 10 years of American life have been uniquely stupid”. The byline of the article, ‘It’s not just a phase’, suggests that we are in this for the long haul and thus are in need of foundational solutions.
Protocols
One such solution is government regulation, possibly by breaking apart big tech companies. This is being discussed actively in many countries. But it is unclear if standard notions of monopolies apply to these platforms. Users today are not forced to go and search on Google, but they do so out of choice. These platforms are thus essentially valuable to users. Any regulatory solution must ensure that this value, on a global scale, is preserved.
Mike Masnick champions protocols as one potential solution, in his article titled “Protocols, Not Platforms: A Technological Approach to Free Speech”. Mike notes how the early internet was built upon open protocols such as SMTP for email, IRC for chat, and HTTP for web content itself. This allowed anyone to build products on top of the open protocols, which was maintained by working groups representing governmental, academic, and commercial entities. Mike’s hope is for protocols to remain open and well supported, enabling users to switch between any number of competing platforms or products all supporting the same underlying protocol.
While primacy of protocols is a foundational point, the implementation of open protocols enabling user agency in choosing between products is not quite straightforward. Consider for instance email with SMTP. Indeed, a user can use any of the dozen or so popular email clients to receive email following the SMTP open standard. However, Gmail as a platform, while supporting SMTP, has a huge user base acquired initially by offering unprecedented mailbox space on the cloud for free. This is right out of Jean Tirole’s playbook. So, while the metaphorical pipelines that carry user email are compatible with SMTP, the pipelines themselves belong to Google, and so too the users, their attention, and often their data.
So, both regulating platform companies as monopolies or defining open protocols don’t seem to solve the entire problem. Would a combination of the two work, i.e., can governments regulate openly built protocols? Thankfully, this is not just management speak, but an idea which has been tried and shown to succeed in India. In a series of innovations over the last decade and half, led by public minded technocrats including Nandan Nilekani, India has deployed government backed protocols at population scale. The first one was a unique identity for all individuals called Aadhaar. This unique ID enables multiple services at population scale such as electronic know-your-customer (eKYC), digital benefit transfer for government subsidy, digital lockers for documents like vaccine certificates. A second example is Unified Payment Interfaces (UPI) for digital payments which has powered transactions worth 1 trillion USD in financial year 2022 at significantly lower transaction costs than credit cards. A third and developing example is that of account aggregators which enable institutions like banks and hospitals to share user data with consent to third party institutions such as a lender or an insurance provider.
In each of these cases, the government has backed the protocol and put in place regulations, often in response to commercial usage. Given the scale these efforts have reached, it is valuable to study them from the framework of aggregation theory. Again, from Ben’s theory the three things that characterize an aggregator are direct access to users, low marginal costs of service, and two-sided markets. For a government backed protocol, access to users comes through regulation and the government’s already existing people connect. So, no dark patterns are necessary to “acquire” users. On marginal costs of service, systems like Aadhaar and UPI are state-of-the-art digital services which benefit from economies of scale of datacenters and ubiquity of smartphones. Open-source software and voluntary effort of public minded individuals also contribute. On two-sided markets, in the case of a democratic government, these services are offered free to the users and are absorbed as a public good by the public exchequer. However, private enterprises that interface with the protocols, such as payments apps using UPI, benefit immensely and champion these protocols further. So, government backed protocols supply an interesting contrast to big tech platforms. And in doing so, government builds a techno-legal state capacity, where state capacity is defined by Brink Lindsey as “the ability to design and execute policy effectively”.
Predictive algorithms
Government backed protocols have succeeded in India in defining a tech stack for identity, payments, and consented data sharing. But so far these interventions have had nothing to do with platforms such as YouTube, Facebook, or Twitter. What do these platforms have apart from users acquired and served with low costs? Answer - they have highly accurate predictive algorithms, i.e., algorithms that can predict which movie we would enjoy, which tweet we would retweet, which product we would add to the cart, or which partner we may date.
The rise in the importance of predictive algorithms has been dramatic over the last decade. One way to appreciate this is to contrast the innovations at Instagram and TikTok. When Instagram was launched in 2010, its primary USP was the addictive nature of a scrolling image feed and the network effects that set in as people got hooked. When TikTok launched in 2016, there already were several other apps that had tried out the scrolling feed idea on short form videos. A crucial differentiator for TikTok was its personalized recommendation engine. The engine allowed discovery of highly relevant videos from all creators as opposed to creators followed by a user. This also set in network efforts, but at much wider scales than platforms like Instagram.
Advances in the field of Artificial Intelligence (AI) and specifically Deep Learning (DL) have been at the center stage of improving accuracies of predictive algorithms such as recommendation systems for videos. At their core, these algorithms use various sources of data to learn to “represent” content and users, and then to learn “relevance” of a piece of content to a user. The algorithms that are most effective in learning to represent and check relevance are often simple but require huge amounts of user data to learn from. The reputed AI researcher Rich Sutton called this the bitter lesson - “Simple AI leveraging compute power [and data to operate on] beats clever AI built using human knowledge”. This bitter lesson plays right into the advantage of aggregators. Aggregators have direct access to users, can maintain a continuous stream of data from them due to low marginal costs, and can leverage the data through predictive algorithms in two-sided markets. The more the users, the more the data, the better the predictive algorithms, and hence greater the ability to attract new users. This reinforcing cycle is also why after over a decade of investment in Bing, Microsoft hasn’t made a large dent in web search.
More than any other aspect, it is this astonishing accuracy of predictive algorithms at scale that characterizes the dominance of aggregators today. Amazon, with a huge user base has very accurate sense of the products people buy at different locations and can leverage this to co-develop products itself. Google and Facebook with detailed user profiles can target advertisements far more precisely than any other platform. And this trend will only accelerate because of sustained growth in three axes - users are putting out ever more digital fingerprints for predictive algorithms to train on, deep learning models used in predictive algorithms are becoming larger and more accurate, and hardware for compute and storage is becoming cheaper and more powerful.
Again, like with platforms, predictive algorithms are not all bad. A personalized drug discovery may happen through predictive algorithms, or enormous amounts of carbon emission can be saved with green routing. But the design of the platforms in which predictive algorithms are embedded often leads to poor trade-offs. For instance, Facebook is incentivized to get teens addicted on social media. YouTube is incentivized to appeal to our moralistic selves rather than our reflective selves. Twitter is incentivized to enable 280-character bickering and trolling. And scarily predictive algorithms trained to maximize these incentives are inhumanely effective.
This effectiveness has led to a lot of angst. The popular public intellectual Yuval Harari has spoken at Davos about how predictive algorithms are setting us humans to be hacked by machines, because we know far less about ourselves than do these algorithms. There is also concern with motivated actors gaming the predictive algorithms on these platforms to drive their own agendas such as spreading disinformation. This can get particularly slippery if these actors also use advanced AI technologies to generate realistic multimedia content. Finally, there are concerns with these AI platforms causing income disparity. As Diane Coyle writes in her book Cogs and Monsters, “An economy of tech millionaires or billionaires and gig workers, will not be politically sustainable”.
People
At the center of such income disparity are people. So how should platforms, protocols, and predictive algorithms evolve for better outcomes for people? What role should government regulation have? These are some of the defining questions facing our generation.
Platforms are intrinsically valuable to users. Forcing deep regulation, such as breaking them up, may prove net negative. Instead, platforms can evolve in an obvious yet foundational manner to be people-centric: A platform that captures user data, and grows beyond a certain user base say 1 million active users, should be required by regulation to make parts of the users’ data it collects available to the user. Further, the platform should allow any third party, that the user consents to, to also have access to the data. For instance, if a student wanted to prove to her teacher that she had seen all lecture videos on YouTube, she should be able to set up a consent for her teacher to run an automated check on her viewing history for a specific channel. This is a basic requirement that we take for granted in more traditional dealings. For instance, it would be very strange if a professor declined to show the attendance sheet to a student, or a university declined to share course grades with an external admission committee. GDPR’s right of access clause is a toned down version of this regulation, wherein a user has to initiate manual processes to access data in custom formats. To be effective, two conditions need to be met: (a) such access must be programmatic with minimal or zero user intervention, and (b) the data shared should be in openly published formats enabling reuse. Such programmatic and standardized access opens a host of possibilities for how people can use these platforms. As an example I may share my YouTube viewing history to a startup which specializes in making recommendations for videos for my region. The startup may use personalized recommendation engines like YouTube does today or may combine it with manually curated content from curators I choose. Crucially, if I do not like their recommendations, I would be able to switch to a different provider or default back to YouTube’s engine.
Protocols when regulated do enable efficient population scale public goods such as in the India stack. Another potentially valuable class of public goods are publicly available predictive algorithms trained on data pooled together by citizens. The training and the hosting of the models can be done by a government backed public institution which should earn citizens’ trust. A part of this trust will be dependent on adopting state-of-the-art privacy technology combining advances in cryptographic algorithms, privacy preserving federative learning, and secure hardware. Further, standardized protocols would be needed for how citizens can share data from multiple platforms and consume the outputs of predictive algorithms. An example of such a public good relevant to social media platforms is a predictive algorithm to find disinformation. Given the multi-modal and omni-channel nature of social media and the high velocity of information spread, it would be valuable to pool together user feedback across platforms to train a single algorithm. Another example is traffic modelling and routing. Real-time traffic conditions are important public information that should be available to public institutions without dependence on entities such as Google Maps, Magellan, and Garmin. Another area which can benefit from public models is healthcare. Given how health data is split up into silos across hospitals, it is hard for any one entity to train a predictive algorithm say for a blood pathology test. A public institution pooling data with strict privacy would go a long way in introducing AI in healthcare to assist our over-burdened healthcare professionals. A similar case can also be made in the financial sector. While large banks can afford to have predictive algorithms to identify risk associated with a mortgage, small lenders may depend on public models trained on confidential user and enterprise data, such as transaction and tax history.
Predictive algorithms such as recommendation engines are riding the wave of innovation in Deep Learning (DL) over the last decade. Increasingly, more of this innovation is coming from companies rather than universities. But quite strangely, a lot of the research outcomes in DL have been open-sourced soon after invention. This is in part because of the competitive space of hiring and retaining top AI researchers who are often driven by peer recognition through research papers and transparent benchmarking. Thus, much of the machinery used to build predictive algorithms such as programming frameworks, compilers, and hardware runtimes have been open-sourced. Many of the algorithms have also been open sourced as studied by the Center for Research on Foundation Models at Stanford University. Foundation models refer to algorithms that are trained on massive amounts of data which can be fine-tuned for specific tasks with smaller amounts of data. A good example is the BERT language model released by Google in 2018 which within the last 4 years has been cited by 40,000 research papers. The BERT model was trained on a large amount of text obtained from the web and can be relatively easily fine-tuned for a task such as estimating sentiment of a given utterance. The center at Stanford makes a case that these foundation models have been pivotal in the proliferation of AI, across text, audio, images, and video modalities. They also raise concerns that these foundation models need to be broad-based supporting more contexts, races, and languages. This importance of open-source AI technology has not received much attention from governments. European and American policies on AI make no mention of it. Going ahead, governments need to create public institutions supported by researchers to analyze, promote, and improve on open-source AI technology. Government of India’s Bhashini project is an example of such an effort with the goal of building open-source AI technology for the long list of Indian languages.
Partnerships
There is hope that platforms, protocols, and predictive algorithms evolve to be more people centric. But this requires deeper and trusted partnerships between big tech, governments, and budding entrepreneurs.
Big tech companies should be encouraged to continue innovation in AI and related areas. The value generation from these companies for users and shareholders is unprecedented. As hobble into post-pandemic economic and political uncertainties, we need to improve our technology enabling every individual to be productive and content. We need to improve outcomes in healthcare, transportation, climate change, and societal justice. And technology innovation, especially from big tech, has a key role in these areas. Another foundational component where big tech should continue to contribute is infrastructure, specifically data centers around the world. As Microsoft’s Brad Smith writes in this book Tools and Weapons, data centers today are what banks were earlier, providing globally available infrastructure that is widely trusted. The economies of scale of large datacenters provide value for all stakeholders. Further innovation in confidential computing, data provenance, and multi-cloud deployments will continue to drive adoption with trust.
Governments are visibly spending more time in anti-trust and regulatory action. There are often valid reasons for this as per the laws of the land. However, governments may do well to engage more deeply in AI technology with new public institutions. The definition of protocols for consented data access, joint training of predictive algorithms as public goods, and support for foundation models are examples of such deep engagements. Governments may benefit from aggregation theory just like big tech companies do, and thus have outsized impact on people with technology-driven public goods. Governments may also want to grow into the vacuum left by the reasonable trust deficit users have with large platforms today.
Entrepreneurs have a rich space to innovate given the rapid advances in AI technology. While there remains constant opportunity to build the next big platform, there also exist opportunities to build products catering to smaller sets of individuals with specific needs. Management guru Seth Godin may be right in his analysis that, “The mass market is being replaced by multiple micro-markets and the long tail of choice.” One may think of many examples of micro-markets. Consider for example an app for elderly people to digitize their medical records (printouts, forms, X-rays, prescriptions, etc.) super easily. Or an app for a family to hear Shakespeare’s Hamlet performed in the voices of the family members. Or a custom search engine for journalists indexing only high quality longform text. Or a chatbot for a teen who wants to get out of TikTok addiction. Or an app that helps me to memorize and correctly chant old Sanskrit compositions. Or an app for an author to dictate a book while mixing in questions like “What did Sherlock have to say about violins”. The list of such applications is indeed long, reflecting the organic heterogeneity in our societies. For entrepreneurs, the barriers to innovate on such applications is historically one of the lowest with availability of open-source foundation models, on-demand compute on the cloud, and many free-to-access educational resources.
Play
What is the essential philosophy that would drive partnerships between big tech, governments, and entrepreneurs to evolve platforms, protocols, and predictive algorithms to be people-centric?
I personally see things from the standpoint of Vedantic philosophy which proclaims a singular cosmic reality that has been apparently splintered into individual conscious experiences that you and I live. These apparent experiences are merely play creating the stage of our journey of self-discovery. While being simultaneously simple and obtuse, the philosophy at its core maximally empowers every individual with abundance and purpose.
In the play of technology as we stand today, we have a choice. We can either build large and complex AI systems with tricky objectives of gaming behavior to enhance engagement, or we can build them with the objective of empowering and supporting the journey of each individual towards self-discovery. This choice for society would not be cohesive. There would be individuals and organizations who prize one objective over the other based on respective value systems. The question is - Can we build the operating conditions within which empowering every individual positively becomes the dominant choice. While this question is a perennial one, the scale and power of platforms and predictive algorithms makes this specific time in human evolution rather momentous.
Excellent👍. Inspiring & thought provoking. Technological advancements create haves & have nots and widen disparity. The impact of advancement in data science on common man is many folds. But any such advancements can be made beneficial to human-kind by making it inclusive. Classical example is UPI revolution in payment space Vs Card payment.
Congratulations 👍All the best
Contemporary well though out beautifully articulated presentation with focus on emerging trends