AI Moats & Google Memo
Societal Risk of Platform Lock-ins
[first rough draft]
In May 2023, a document purporting to be an internal Google Memo leaked on the net that claimed “We Have No Moat and Neither does OpenAI”. They were referring to the idea there was no figurative moat protecting them from competition in the realm of LLMs. The idea was that new LLMs could be trained which would outdo existing chatbots and users could easily switch. It is likely moats will be added before the current technological moat is crossed by any startup or open-source project.
Moats are good for companies but can inadvertently hold back society if they prevent the rise of better technology. Even technology that is different, even if superficially comparable, might have let society develop along a different path. This article suggests the particular social media companies that gained an oligopoly in social media created algorithms that led to what they term a “Viral Inquisitor” rather than the “Viral Editor” many early net pioneers envisioned which might have arisen instead.
To use concepts familiar to those in the AI world, social media technology may have hill climbed society to a local optimum where a stable oligopoly equilibrium keeps technology stagnant until something major disrupts it.
The original hypertext ideas that predated the web contained ways to annotate pages or attach back links. Marc Andreessen originally began to implement the idea in Mosaic, but dropped the feature and so:
The Web Has a Missing Feature You Never Knew About
[…]Few people in this world can point to a decision they made at some point in the past and say with absolute certainty that if they’d taken a different route, it would have changed the world. Marc Andreessen is one of those people. As the co-founder of Netscape, Andreessen made an entirely rational decision to drop a feature that would mean they could ship a working product much faster. However, the result is that he’s spent a lot of time wondering how things would have turned out if the feature had stayed the course.
He thinks “the world would have turned out differently if users had been able to annotate everything — to layer knowledge on top of web pages.” Ted Nelson coined the term hypertext and envisioned features in his ideal Xanadu version of it like backlinks and the ability to transparently track payments for bits of content people cut and paste into other works. This page summarizes some features of it, saying:
The Web as it currently exists is something of a shadow of its potential. Many thinkers, chief among them Ted Nelson, set out systems that could have provided for something more robust and human than the Web, and which would be subject to fewer of the issues that plague it at present: these ideas have not yet been made real.
Unfortunately, Ted Nelson’s vision was so complicated to create that a version of it was not implemented until decades after the web arose and spread. Although society is likely better off having had some version of hypertext for those decades rather than having waited, it is difficult to add certain features now. It is like trying to build a skyscraper on top of a foundation created for a small cottage.
It may be that his approach was not the best path to take, but a better intermediate path might have been possible if some other companies or individuals had been the ones to build the early browsers and server software.
Wikipeda says that in October 1993:
Interpedia was the first-proposed online encyclopedia which would allow anyone to contribute by writing articles and submitting them to the central catalogue of all Interpedia pages[…]
Several independent "Seal-of-approval" (SOAP) agencies were envisioned which would rate Interpedia articles based on criteria of their own choosing; users could then decide which agencies' recommendations to follow.
The project was actively discussed for around half a year, but never left the planning stages, perhaps partly due to the unprecedented growth of the World Wide Web.
It was merely the first such proposal Wikipedia’s writers are aware of. The author of this substack, and likely others, considered such ideas before seeing the first post about Interpedia. The major reason for the project not progressing was the opposite of what Wikipedia suggests: the net did not yet have a critical mass of people online with time to volunteer to create content. Wikipedia arose later when the net population was large enough to support its rise.
Wikipedia is not always reliable. Some of the ideas discussed for Interpedia may have evolved into ways to address various problems with the net like those with Wikipedia’s reliability. Unfortunately Wikipedia’s simplistic approach became entrenched and has stagnated. Although many people champion the idea of open source and non profit platforms in hopes a collective will produce a good product, they can be subject to group think and stagnate just as private products can. The same issues with diffusion of innovation that impede products in society can impede innovations within organizational structures, even those created by technology advocates.
The author and likely others had related ideas we did not publicize due to a desire to find a way to monetize them. Many people like the author also keep some ideas secret out of fear that others will similarly do a poor-quality or limited implementation of the ideas that doesn’t evolve, yet will get entrenched and prevent a better approach from being adopted. Many products stagnate and implement only a shadow of the real potential for the idea, or lead society in a problematic direction.
The general wide functionality of AI suggests some AI software may turn into widely used platforms that have a similarly widespread impact on society, for better and worse.
The claim that the leading general purpose LLM vendors do not have a “moat” seems to assume LLMs only need to be trained from publicly available data and therefore are all on an equal playing field, and that new techniques will allow comparable models to be trained without the huge computing requirements the current large models require.
That ignores an important idea raised on this substack suggesting the idea of a diverse set of AIs or AI personas that adapt to various subcultures and users. The creation of those diverse AIs ideally would involve input from many users. Although OpenAI has said they wish “Democratic inputs to AI” with data that would be publicly shared, other user feedback data may be kept private.
In general, vendors that have a large user base can take feedback from users to train new models that are better adapted to users. Like the network effect where social media gains utility from having more users, AI systems may gain utility from having more users if their creators take feedback and use it to upgrade their AI’s training. It is an open question how much of an advantage that will provide compared to innovative technology that improves LLMs in other ways.
While there has been research done with small models, nothing has yet achieved the functionality of the larger models despite wishful thinking they will do so soon. It may be that some breakthrough may allow small companies or hobbyists to create models comparable to OpenAI’s GPT-4, or it may turn out to be wishful thinking. Even if small models catch up to GPT-4, OpenAI or other vendors with lots of resources may find ways to make their large models even better and maintain a functionality moat.
There are other ways LLMs embedded within products can raise switching costs to try to keep users even if in theory a competitor’s LLM has improved functionality. This substack raised concerns over the risk Microsoft and Google could decide to make it difficult for third parties to fully replace the AIs they embed in their office suites rather than allowing plug-ins to fully swap out their functionality.
Most people use office suite software from either Google or Microsoft despite myriad competitors that may be arguably better for certain purposes due to switching costs. Many markets become oligopolies that survive until there is a major shift in technology that leads to a new market opening.
The impact of the rise of AI in disrupting existing software markets is more difficult to predict since there are many potential ways an AI can provide utility for one purpose and expand into utility for other tasks.
Society is Prone to Platform Lockins
Technology markets are often not fully competitive, and that can slow the adoption of new technology. Once a market is fully grown and saturated by an oligopoly, even just switching costs can be enough of a moat to deter a better product from trying to enter the market. Investors prefer to use limited funds for companies entering growing markets rather than for the more challenging task of displacing entrenched products. It requires a major leap in technology compared to the existing player before most markets can be disrupted.
Factors like the network effect make incumbents in some niches even more difficult to displace. Facebook may have displaced earlier entrants like Myspace: but it did so before the social media market was fully saturated. It is often only companies with an existing market share for some other niche that can attempt to launch an attack on a related niche. Meta just launched Threads that copies much of the functionality of Twitter, and acquired over 100 million users within a few days, unlike prior Twitter competitors that took a long time to reach a tiny fraction of that.
Some companies remain innovative for many years, while others seem to not rock the boat once they have something that works. Once some companies are entrenched: even if they spend lots of money on R&D, the research does not see the light of day. Xerox PARC may be the most well-known example of a company that did not capitalize on its research, but today Google and Facebook fund many projects that are never released to the public.
Startup Approach Risks Subpar Lockins
Unfortunately, the way most startup funding is done currently risks important niches being held by potentially subpar products. While the Venture Capital industry in theory profits most from the small percentage of companies that hit it enormously big, those are difficult to identify.
Less visionary VCs tend to instead focus on protecting their downside and seek products that have “traction” in the marketplace to ensure they will at least go somewhere. Early-stage capital to develop a product is harder to find, so many companies do not have the funds to build an ideal version of their product. Even if they had the funds, many rationally try to launch quickly to start gaining market share and to deter competition from others who saw the niche.
Tech startups often focus on producing a “minimum viable product” that will gain them sales. Those who develop an MVP may or may not have a broader vision for how a product may evolve in the future and may or may not have built a technical foundation capable of evolving.
Once they have reached the minimum viable product that lets them achieve a spot in an oligopoly controlling a market, there is less pressure for them to innovate. Society benefits from the creation of these companies: but their impact in the long run may depend on who is running them and the corporate culture they create.
In the realm of forecasting the future, there are 2% of forecasters labeled superforecasters who do a far better job than others at predicting the future. Those who succeed with a startup idea have managed to forecast that one idea is good: they may or may not be the equivalent of superforecasters who have the mindset to continue innovating broadly.
Some entrepreneurs are innovative in many ways, others merely were in the right time and place with a one-hit wonder and do not necessarily have a broader vision. Those leading such a company are only guaranteed to have been able to invent that product, they may be “one hit wonders” and not be innovative in general.
Fortunately for society those who innovate with complex products like artificial intelligence seem more likely to be people who are possess the traits of superforecasters like “better at inductive reasoning, pattern detection, cognitive flexibility, and open-mindedness".
Sometimes potential new platform opportunities can be predicted in advance when a new technology or need can be predicted to arise. In 1999 many people were claiming the “operating system wars” were over with a seeming steady state reached with Microsoft Windows leading and Apple with a minority share ensuring at least limited progress. At the time this author was telling people that the rise of new handheld form factors would lead to an opening for a new operating system war, as it did several years later leading to IOS and Android claiming the market.
Unfortunately, the future evolution of Artificial Intelligence and what niches may become entrenched platforms with moats is not easy to predict. It is such a general-purpose technology that general purpose AI’s like ChatGPT, or the API can serve many potential niches. Specific applications may have niches for specialized products to develop a moat, but those niches may be broad enough to overlap with other niches where products are seeking to get entrenched and created moats.
Alan Kay argued “The best way to predict the future is to invent it.” It is only in hindsight that we will see who invented the moats that arise. Investors and entrepreneurs will just need to take the risk to try to capture a moat and deal with the uncertain prospects for where competition may arise.
Indirect Product Paths
In 1993, someone who wanted to see most people carry around high quality digital cameras in 30 years would have focused on creating a stand-alone digital camera they would evolve over the years. Most likely wouldn't have focused on integrating their camera with cell phone technology as the way to get their technology adopted. In that case the cell phone platforms adopted other technologies which were able to spread along with them. It is possible some eventually widely used AI technologies will similarly turn out to have tagged along as an added feature for a different product. The open question is whether platforms for AI will similarly aid the rise of related technologies by having component parts where bits of technology can be provided by third parties that will evolve good solutions, or whether they will limit the ability to add on third party technologies.
Complicated Moats And Early Releases, Crypto Example
Some moats have proven very difficult to cross despite widespread interest in doing so. Many wish to replace government currencies, but have underestimated or misunderstood the moats that protect them. Its useful to be aware of the problems they encountered to avoid repeating the same mistakes when hoping to apply AI. The adoption of new technologies may also potentially lead to first movers creating moats similarly difficult for later entrants to displace.
Although major government fiat currencies like the US dollar suffer from some amount of inflation, overall, they maintain far more price stability than cryptocurrencies like Bitcoin. Initially government currencies were accepted by many people due to being being backed by metals like gold, or other commodities. Now the US dollar is essentially backed in a decentralized fashion by the billions of goods it can be exchanged for. It is used in vast numbers of purchase transactions each day, with prices for each item collectively backing the currency and helping to stabilize its price.
The Federal Reserve and the US Treasury are centralized entities that have a disproportionate level of influence on the dollar and choose to attempt to interfere with its price. However most of its value over a short term is stabilized by a decentralized process. Bitcoin created a predictable process for the number of coins, but that isn’t the only aspect of price stability. It didn’t address the decentralized process for price stability.
The US dollar benefits from a network effect where the existence of a vast numbers of users gives it price stability. If a new currency were widely used in transactions the way the US dollar is, that should provide comparable price stability. The problem is how to get to that wide usage.
Unfortunately, when a new cryptocurrency launches that is not backed by anything, like Bitcoin, it does not have a similarly large base of transactions or products whose prices are set using its units. Its value is purely speculative and prone to potentially large fluctuations. Those fluctuations deter its use as a transactional platform and undermine the ability to reach the widespread usage that would provide stability.
The need for many users actively purchasing goods using a new unbacked currency to stabilize it acts as a moat preventing competition. It is an even more complicated moat to cross than the one social media has due to the network effect. Unlike with the network effect where functionality improves consistently as users are added, the initial users in this case tend to engage in speculation that exacerbates the issue.
David Chaum originally proposed the idea of an electronic cash in a 1982 paper after having created what are foundational ideas related to blockchains in his 1981 thesis. Despite attempts to create a commercial digital version of cash over the next few decades, it was not until 2008 that Bitcoin was invented and became the first widely held version of digital cash. Unfortunately, Bitcoin and later cryptocurrencies have not fulfilled the hopes many have had for the idea, likely including its creator who hoped for their use in micropayments.
The problem is Bitcoin had not found a way to cross the moat to price stability but was launched despite that. Its speculative usage led to wide adoption, but its transactional use did not grow to stabilize it. Venmo was launched around the same time as Bitcoin and also addressed the issue of digital payments, yet Venmo has vastly more usage in real world transactions. Some like the author saw Bitcoin as a great leap in technology, but a very poorly considered product launch if the goal was to create a widely used payment mechanism useful for things like micropayments.
It had solved difficult technical problems, but hadn’t addressed larger problems. It hadn’t yet solved the issues of providing price stability prior to widespread usage, nor did it seem to have a path to cross the chasm to widespread acceptance. It has been mostly driven by those who envision what the world could be like if it were used widely, who haven’t yet found out how to cross the chasm to get there. They were addressing a technical and ideological preference, not addressing a real user need with a good user experience.
The first product in a niche doesn’t always get things right, and the still unmet hope is someone will find a “killer app” or rethink the whole user experience and product approach in a way that leads to a product able to cross the chasm to the mainstream. Unfortunately the speculative usage of cryptocurrencies has led to skepticism about them which risks slowing the adoption of whatever new cryptocurrency arises that does have a realistic plan to get widespread adoption.
More importantly, many countries are considering the potential for government run digital currencies which have features that are contrary to the hopes of many cryptocurrency advocates. The spread of Bitcoin without a way to cross the moat to price stability, and the chasm to wide usage, provided the concept to governments which have a built in user base and way to launch their products and get the widespread adoption private cryptocurrencies haven’t yet achieved.
If Bitcoin, and similar cryptocurrencies had not been released to be used by early adopters, it may be that a better currency would have been invented that addressed the issue of price stability somehow and had a better user experience in a private currency that managed to succeed. Or perhaps not: no one has yet invented that idea so there is no way to know if it would have been invented earlier if Bitcoin and its successors hadn’t distracted people’s attention from addressing the user experience and marketing problems.
Products like Venmo, Cash App and Paypal focused on solving direct user problems to transfer money from one person to another. To most everyday users who merely wish to pay for a product in a store, asking them to use Bitcoin was like asking them to take their existing store of value in dollars and convert it to yen and give it to a store that takes the yen and exchanges it for dollars. They aren’t interested in having a different store of value. Bitcoin advocates didn’t provide a simple payment system that hid details a user didn’t care about if they didn’t wish to maintain a store of value in Bitcoin.
An unintended side effect of cryptocurrencies being released without a stabilizing process was their use as speculative collectibles leading some folks to gamble on their future value, which spread the usage of Bitcoin as a store of value but decreased its initial intended utility for transactions. Bitcoin didn’t provide simple functionality the way Venmo and the other payment approaches do for those who aren’t interested in such speculation.
It may be we would have had widespread usage of a stable digital currency by now. The idea of smart contracts was also around before the rise of blockchains, in the mid 1990s: but they may also have been held back by what might be also arguably premature release of solutions that have interesting technology but haven’t yet successfully achieved widespread adoption since there wasn’t a fully path considered to cross the chasm.
It is unclear if there is the potential for similarly complicated moats to arise after an AI becomes entrenched as a platform, or if network effect moats are the most likely worst case moats. However there does appear to be the risk that some new types of flawed AI technology will similarly be released prematurely that are flawed as products and slow the adoption of better products. If a product is too flawed, it may even lead to a user backlash agains the product category and skepticism of any new products attempting to enter that market.
Cryptocurrencies are a work in progress that so far have solved some problems, while raising others. Some aspects of most cryptocurrencies like implementation details are centralized in a small number of developers trusted to get things right. In theory there are ways for the users to ensure they do, but in reality they suffer from the issues real world democracies have like rational voter ignorance and apathy. Its likely eventually the various problems will be resolved, but advocates who saw the promise of the technology often underestimated the effort required to address the remaining issues. There are risks the same thing will happen with those hoping to apply AI technology that is also still a work in progress.
Hammers Are Not The Only Tools
The rise of Bitcoin and cryptocurrencies exposed many people to the idea of decentralized processes for the first time, even though many computer scientists and others had been exploring them for years. Unfortunately the saying that if the only tool you have is a hammer, everything looks like a nail applied to the rise of blockchains. There are many ways to build decentralized processes, sometimes even with centralized implementations. Unfortunately many people focused purely on using this cool new “hammer” they had found and didn’t explore other alternatives to find what was best for a particular problem.
Some Bitcoin enthusiasts even adopted the idea of Bitcoin maximalism: the idea that its the only cryptocurrency people should support, that only a hammer is needed. For society it seems the best approach will be competing cryptocurrencies with different features for different purposes. The eventual goal should be a framework to address the problem of allowing people to have their assets stored in whatever form they wish and to be able to transfer a certain quantity of value to other(s) who may wish to store their assets in some other form. The first step though is finding any solution to the user experience issues that will lead to crossing the chasm to widespread public adoption of even one cryptocurrency.
Those who saw the utility of blockchains focused on seeking applications for the technology, rather than trying to solve problems where it may be merely one piece of the puzzle. A Wired article in 2019 addressed the issue of “What’s Blockchain Good for, Anyway? For Now, Not Much”, noting for example:
Stevens’ team focused on projects that touted blockchain as a way to identify fraudulent and tainted goods in supply chains. They predicted 90 percent of those projects would eventually stall.
Many of the issues involved in addressing these issues have nothing to do with the blockchain technology which is only addressing part of the problems. There are many potential uses for the concept of smart contracts, but people may similarly overestimate the difficulty of applying the concept to domains when human level flexibility and nuance is needed to interpret contracts that are used in the real world so the process may take longer than people wish.
Merely because in theory a technology can apply to a problem, that doesn’t mean its required or the best choice, or that the technology is truly yet up to the task. There new wave of AI risks falling into the same trap. AI will often be capable of a task to some degree, but the question will be whether its the best option or merely the hammer everyone knows now exists. Another page on this Substack deals with that issue regarding the problems LLMs have with logic, which might be better addressed by symbolic AI. There is a risk that LLMs may be a fast way to implement some product and get it entrenched as a platform that we are locked into, yet with the technologies limitations hindering the progress of the platform.
Work on the diffusion of innovations suggests even free new ideas aren’t always easily spread. People working on new products are always concerned about how to cross the chasm from those who are early adopters to those who resist change. The tendency of people to anthropomorphize AI may be a mixed blessing. In some niches it may be seen as comparable to having a human perform a task and easily accepted.
People have discovered however that with robots there is an uncanny valley where robots that seem close to human, but not quite, aren’t well accepted. In other niches of AI the tendency to subconsciously anthropomorphize the software may people to struggle to interact with them smoothly since they don’t quite know how to do so and lead to a similar difficult with acceptance.
The flaws in the current AI technology may also lead to products that almost work, but not quite, in ways that aren’t easy to fix. Early adopters often accept beta or even alpha release quality software that has flaws in exchange for the utility that does work, if the problems persist for too long they may give up on it. Many users may be cautious about adopting software that has great utility, but where the subtle flaws may take too long to find ways to fix. Just as some people have lost patience waiting for the promised utility of blockchain, even if that potential may one day arise, there is the risk of a backlash against some potential uses of AI if they launch prematurely before the technology is fully up to the task.
Thanks for reading Society and AI Substack! Subscribe for free to receive new posts and support my work.