Inside the AI-as-a-Service Phenomenon
And a peek into AWS’s latest playbook
In his 2017 shareholder letter, Jeff Bezos hinted at a new wave of AWS value-add products — AI-powered APIs.
“Amazon Lex (what’s inside Alexa), Amazon Polly, and Amazon Rekognition remove the heavy lifting from natural language understanding, speech generation, and image analysis. They can be accessed with simple API calls — no machine learning expertise required. Watch this space. Much more to come.” — Jeff Bezos
In the past, AWS has been prudent in shaping and keeping up with trends, be it hosted DBaaS (database management as a service) or even hosted messaging and application services. And there’s every reason to assume that this new wave of products will prove just as successful.
Even though the concept of AI-as-a-Service is buzzword laden — AI and SaaS are, after all, the driving trends of the day — it’s worth looking past the hype at the underlying factors which could make this trend prove seminal.
Factor #1: Tech loves abstractions
If I have seen further, it’s by standing on the shoulders of giants.
This adage is no truer than in the context of software. The most basic software applications of today bring to the table a level of complexity that outshines that of applications of even a few years ago. This is down to the fact that software projects rely heavily on libraries and APIs that preclude the need to reinvent the wheel.
While AI developers have access to useful pre-trained networks and simplified frameworks, APIs that automate tasks end-to-end have yet to become mainstream. That is, the Twilio equivalents of AI are yet to emerge.
If your app requires voice-to-text support, until recently, you had little choice but to pencil a few extra months into your product development timeline, while your data-scientists and developers put together a scalable voice module. Now, with choices like Amazon Polly, voice integration is a matter of a simple API integration. Isn’t that something!
Factor #2: Machine learning is actually difficult
A few free courses on Udacity is all it takes a good developer to come to the conclusion that AI is easy to do. The body of standard problems and training datasets out there makes it seem, at first glance, that the problems are easy to solve.
The reality is that domain-specific solutions require a lot of meticulous dataset curation and iterative hyper-parameter tuning.
What’s more, the kind of numbers that look good at first glance on an academic paper don’t cut it when it comes to your company’s business needs. You might be impressed at first by an academic paper that describes and algorithm which matches SKUs across retail sites with an accuracy of 90%. But were I to tell you that this translates to 1MM faulty records when scaled up to a corpus of 100MM products, you’d probably do a double take.
Customers demand quality, and this requires your data-scientists to work on eking out every last percentage point of improvement. Businesses that understand this reality are very likely to shell out for the SaaS subscriptions required to get access to top-of-the-market deep-learning models, and allow their teams to focus on what it takes to differentiate their organizations.
Factor #3: Demand for AI
When your competitor launches a feature that takes their offering to the next level, you have two choices — keep up or watch the train pass you by. And present-day AI enables precisely the kind of improvements that can change the playing field overnight.
Given this, it’s likely that even slight industry shake-ups will give rise to demand where none may have existed previously. Enter on cue — packaged APIs that help businesses catch-up with their competitors.
Factor #4: High cost of building models
A majority of the cost in building machine learning models is concentrated in the training phase (think of this as algorithm creation) rather than the deployment phase (what it takes to execute a prepared algorithm). And this cost is non-trivial — data-scientists are expensive to hire, and the infrastructure costs that come with significant projects are sizeable.
The economics of this setup lends itself perfectly to an API-centric approach. A single player absorbs the capital cost of training models, and makes it available at a margin on deployment cost to a range of customers downstream, delivering cost benefits to all involved!
Factor #5: Training data isn’t democratized
The beauty of the deep learning revolution is that the knowledge and code that drives it is largely open-source. This gives anybody the opportunity to train their networks using the latest greatest techniques.
Unfortunately, the high quality training data that deep learning so depends on (which we’ve written about previously), is not as readily accessible. Beyond generalized datasets released for academic purposes, most valuable domain-specific datasets lie behind the firewall, i.e., are privately “owned”. This implies that, some companies have to ability to build better models than others, regardless of factors like team quality and availability of funding.
Companies that find that they don’t have the datasets needed to match their ambitions will look elsewhere — to companies that own/build/lease the datasets needed to build these models.
Factor #6: A plethora of repeatable tasks
“Vertical SaaS is a SaaS company that focuses on customers in a specific sector such as healthcare or media. Horizontal SaaS is a SaaS company that will sell to anyone in any sector”. — Blossom Street Ventures
AI APIs can be viewed through the same lens that SaaS companies often are — horizontal vs vertical. Google’s NLP API and Amazon Lex are examples of horizontal APIs, since their utility isn’t restricted to any single vertical. Semantics3’s own Categorization API and Feature Extraction API are examples of vertical APIs, since they are singularly focused on the needs of Ecommerce companies.
While we’ve started to see a wave of Horizontal AI API enter the market, we’ve barely seen the tip of the iceberg when it comes to Vertical AI APIs. Think of the possibilities that lie ahead — medical APIs for healthcare institutions to plug in to, Ecommerce APIs for retailers and brands to utilize, routing APIs for logistical companies to capitalize on…
In each of these cases, developers have the opportunity to tune their models to the needs of a very specific group of customers, making it all the more likely that out-of-the-box solutions will slot in perfectly with customer expectations.
The scope of opportunities on offer via Vertical APIs alone makes it likely that AI-as-a-Service will prove to be a telling trend.
The beacons for the AI API wave are most certainly flashing. It now remains to be seen how this impacts companies’ tech strategies, which vendors win through, and critically, how this helps speed up the democratization of machine learning.
Want to learn more about Semantics3’s AI APIs for Categorization, Matching or Feature Enhancement? Drop us a note!
Published at: April 18, 2017