The rise of SaaS has been the story of software the last 10+ years: recurring revenue, shortened product cycles, ease of deployment and management via the cloud, access anywhere (with internet), and monster shareholder accretion.

What’s not to love. In fact, there’s so much love for its core ideas that you’d be hard pressed not to see how it’s infected and affected the industry spectrum:

  • Peloton: monthly membership fee that unlocks cycling content and classes that are accessible anytime and anywhere
  • Amazon Prime: monthly membership fee that gives you access to… alot
  • PreFix: monthly fee for access to lower cost handymen and and regular preventative maintenance
  • Apple: newer to the strategy but benefiting from 2x the market capitalization despite a lack of growth in earnings because of their rising Rundle
  • Private Equity: investment firms like Vista Equity and Thoma Bravo have been riding the tide that is SaaS to much success
  • Microsoft: away with old school licenses, make way for subscriptions to Microsoft Office and Teams

And that infection has now reached Machine Learning. How?

With Machine Learning APIs.

What are APIS and how to build API? - Learn Steps

‍

Problem

The field of ML is at the crossroads of some major trends including:

  • A shift towards less openness as the power, accuracy, and reach of algorithms outpaces regulation or researchers' intentions
  • Increasing disharmony amongst international coalitions on how to regulate the weaponization of recent inventions LINK LINK
  • Technological advancement and sophistication to the point where implementing breakthroughs increasingly places demands on talent specialization in order to take advantage of it

These trends in turn have and will continue (to varying degrees) to translate to:

  • Less open sourcing of code for the community, competitors, and bad actors to experiment with and immediately productionalize
  • Increasing corporate development costs to translate the theoretical into the practical
  • This takes the form of both higher headcount costs and lost time that could be allocated towards other ventures
  • Another implication is the increasing upfront costs of a project in proportion to its variable cost (assuming the company is even able to implement the architecture which isn’t a sure thing)
  • Let’s not forget the challenge of then productionalizing and scaling a model that is implemented. Non-trivial to say the least.

ML API’s to the Rescue

But what if a company had access to ML “software" that could:

  • At the outset of the project, immediately be tested to get a sense of not only its viability but also its value
  • Rapidly scale upon ramping of the use case
  • Always be up-to-date with the latest, greatest, and most performant architecture (even if that architecture isn’t made fully available to the community)
  • Guarantee uptime and response time akin to AWS (and likely is powered by AWS or one of the big 3 to be honest)

ML APIs - check and check. All the power of the “latest, greatest, and most performant architecture” with the "guaranteed uptime and response time of AWS” that can be “immediately tested to get a sense of the project’s viability” and “rapidly scaled” upon establishing a proof point. All possible behind a simple API structure (or even a really simple UI backed by that API).

When the above vision is achieved, businesses would be quick to adopt. Consider the required capital expenditures or human talent just to keep up with one of the fastest moving industries in the world (just compare what a computer could recognize or translate 10 years ago vs today). Instead of many months and/or years implementing an algorithm (gathering the data, annotating it, understanding how to implement the architecture, testing it) and then productionalizing it, a company could get off the ground in days instead.

And if the math doesn’t check out at scale, and it costs too much to continue outsourcing the ML to an API, corporations can then build it themselves. But that’s the beauty - corporations can quickly test their assumptions early on and consider different routes (build vs buy) at different levels of scale as opposed to front loading all decisions to the outset.  

This also opens up interesting revenue models:

  • What if Tesla’s autonomous driving software was available to any car company with a compatible fleet accessible to the internet (assume 5G is also already widely distributed to send data)? It’s not too far of a stretch given their open sourcing of patents, APIs would enable them to more safely distribute the software, and coordination of fleets would lead to even more efficiency over individual actors. Maybe Tesla’s already elevated market cap could reach skies only SpaceX has seen...
  • What if OpenAI’s Five or DeepMind’s AlphaGo (after more generalization) was available to any gaming company or used to create opponents for professional teams to train against? Some of the world’s most popular games like League of Legends are already in the cloud and thus face relatively few technical challenges to productionalize.
  • What if the best language models could then be used to power chat bots, customer service agents, text analysis, and other language tasks? (to be fair, this is already starting to happen today as we'll see below)

Today and Tomorrow’s Players

Today, research by the best and the brightest often encounters expensive, years-long, uphill battles before product launch. This process normally prevents a lab from independently succeeding without corporate investment or even being a corporate lab, especially in our current environment where industries are increasingly concentrated.  

ML APIs offer a meaningful possibility to flip this relationship in regards to AI and research labs. The best example here is OpenAI, who now offers access to their best-in-class language models powered by GPT-3 (model and code never open sourced) through an easy to use API. See the business model below and at the following link:

Image

These labs will still normally be limited by their engineering talent, so a key requirement will be hiring the right talent or identifying the right partner to convert the research. In the case of OpenAI, I suspect this was a big reason for their $1 billion investment from Microsoft.

Speaking of Microsoft, the tech companies and their labs are also ripe to capture value given their proximity to innovation and muscle to put it into production. Consider the face recognition services offered by Microsft, IBM, and Amazon before closing due to the concerns on their accuracy. Or that they all already offer APIs for general tasks such as Google’s below in the Computer Vision space:

Lastly, I see a role for independents who may not be responsible for the breakthroughs themselves, but are experts at implementation. In other words, these folks spike on engineering and have sufficient ML knowledge to transition research into the real world. HuggingFace, a growing name for language models, is one of the best existing examples of this category.

Resemble AI, a pure play for generating text-to-speech services, is another great example of how specialists are unlocking new use cases like automated media narration in video games or documentaries.

Predictions and Ramifications

Predictions catalyze a conversation, facilitate learning, and calibrate us to the world. Below are a few predictions and/or considerations for ML API’s impact.

(Of course, we’re still in the early innings, so everything can change depending on the actions of big tech, researchers, and the community)

  • $ will accelerate to a combination of labs and independent companies that can distinguish themselves with the best APIs (HuggingFace) or best technologies (OpenAI). Capitalizing on recurring revenue and high margins, these folks will benefit from high valuation multiples.
  • Expect the open sourcing of tools to support this movement akin to Kubernetes for micro services. Research labs, independents, and even the cloud computing companies may all cohesively lead the charge here given their aligned stake in $ capture.
  • Labs will no longer be as tethered to large blue-chip companies. We’re already seeing this with the (mostly) independent OpenAI, and perhaps we’ll start seeing this with other well known players like the Max Planck Institute
  • 2nd order innovation in ML (meaning products built on ML as opposed to algorithmic innovations) is about to take off as ML APIs democratize the technology even more for the masses. Much like AWS unlocked new companies and business models, expect the same for ML APIs.
  • Demand for consultants and in-house ML will shift. This doesn’t mean that demand for consultants and in-house ML will evaporate. It could, but that’s unlikely. More likely is that demand shifts FROM development of models TO leveraging these APIs and/or curating amongst them.
  • Demand for professional data scientists will be somewhat alleviated. There will still be more demand than supply but suspect there'll be a shift of resourcing $ from ML researchers to ML practitioners and engineers

Other Considerations

The industry has traditionally made its mark and achieved exponential progress the last two decades through its dedication to openness. The upcoming cultural shift away from this openness will have significant repercussions of which the most serious will include the challenges of the community and consumers to test for accuracy, robustness, and diversity when models and data are hidden behind opaque APIs.

Progress begets more problems.