When I wrote about the product landscape around Large Language Models a few months ago, I talked about it through the lens of the end-user experience and the tug-of-war inherent to most software applications: On one end of the spectrum is purpose-built products, with maximum proficiency at a certain task - on the other is maximum user convenience through fewer, more general applications. This settles, at least temporarily, into a product landscape determined by all kinds of factors including, as I wrote in my earlier post, the workings of LLMs.
This is partly economics, the study of scarcity, at work for the scarce cognitive load of users. But the economics of more tangible resources, mainly the cost of training and running LLMs, will likely also play a role in the way that the landscape of LLM-based applications emerges. Foundational LLMs, although expensive to build, can be useful for many different tasks. By tuning the models and inserting them in a product environment suitable for each task, they can be even more useful. It makes economic sense for there to be a few platform companies that work on the foundational but costly LLMs, which are used by many companies to build applications for each market.
These market dynamics, with a few platforms used by many companies, are familiar territory in tech. It is no different from other platforms such as cloud services, operating systems, or social networks - all provided predominantly by a few tech giants. Many startups build on top of them while trying to avoid getting crushed. But for each technology the market dynamics might play out differently, and those dynamics have large implications for any startup’s prospects of becoming a large business with that technology.
The doomsday scenario for startups is that of complete lock-in: The startup is dependent on the platform, but the dependency goes one way. As the startup grows, the platform might wonder if it shouldn’t capture that business opportunity for itself. So the platform decides to cut the startup off and offer its own product. Or the startup is crushed by accident, should the platform decide to change its strategy and remove or deprioritize some critical functionality. Both scenarios have disastrous consequences.
What lock-in will there be this time? It mostly depends on the uniqueness of each of the LLMs. Today’s LLMs seem far from irreplaceable to one another, but that could change. LLM platforms would hardly want to sell a commodity, even one that is expensive to build. The platforms might try to differentiate themselves by another scarce, often even unique, resource: data. Unique training data leads to a unique model. We might end up with a landscape of LLMs with capabilities and strengths in different areas, because each LLM platform leverages its access to unique data. Startups that build for a market are then dependent on the LLM platform with unique access to the most relevant data. The importance of unique data will likely play a big role in the market dynamics between platform and startup.
But companies that build on LLMs have to their advantage that they control the user interface - a valuable position to be in. Unlike with operating systems or social networks, they can bring their users with them if they switch LLM platform because it is a tool under the hood. A startup could even use multiple LLMs simultaneously to build their product because it is sensible from a competitive standpoint. LLMs might resemble cloud services in that a startup can switch to another platform without the end user noticing at all. And even if it’s a big job technically, if it is possible that makes all the difference. Time will tell how each market develops.