The Death of the API
Interconnectivity sits at the core of modern software. Any modern software application will, at some point, communicate with other software in order to function: Google, Meta, Shopify, banks, credit card companies, your iPhone, etc. all form a mesh of interconnectivity. These applications all provide specialized services that each other are interdependent on, forming a marketplace of functionality that has become an embedded part of developers’ workflows.
Recent developments in Artificial Intelligence, however, are actively disrupting the central language of this software interconnection - the API (application programming interface). This key disruption has enabled a new method of intercommunication that is more flexible, powerful and accessible. In essence, AI has enabled software to intercommunicate using natural language - that is, regular english statements and sentences, as opposed to just code.
Software has historically improved as more abstract ways of programming emerged. We’ve transitioned from using physical punchcards, writing binary code to programming in higher-level programming languages; each advancement placed us further from the underlying machinery of the computer and, in doing so, multiplied the power of our applications. Now, this transition towards programming computers with natural language (regular english sentences) is poised to radically change the pace of software development, democratizing access for nontechnical builders and inspire a cambrian explosion of more powerful and sophisticated applications.
The Tediousness of the Traditional API
Modern software applications are a complex web of interconnections with other software, facilitated via APIs, or "application programming interfaces". An API is an intermediary that allows two applications to talk to each other. A real life analogy is a restaurant where there are guests and the kitchen, the waiter/waitress helps relay the menu orders from the guests to the kitchen staff who prepared the meal. Stripe, one of the most widely used financial services API, allows developers to take hundreds of actions, including creating charges, refunds and retrieving customer details, using their API.
For developers, the process of building with these APIs is time-consuming and costly. As each software service has a different API specification, or "language" you must speak in order to interact with it, developers must learn a new set of syntax, usage concepts and more for each integration they build out. In addition, in order to speed up development, developers frequently Google search what they are trying to accomplish with a given service's API, then copy the code verbatim from online forums like StackOverflow. Hand-written (or hand-copied) API integrations are error-prone and, like any other code, require active maintenance as the surrounding system changes and develops, leading to slow and brittle development.
This is not creative, fulfilling work, but rather resembles the rote application of a process for translating business needs ("retrieve all Stripe customers who bought something today") to code that obeys certain rules that are specific to the platform you are working with. In my personal experience, for example, integrating a web store with Shopify required months of work from a team of three engineers, a significant portion of which was focused on shuffling data back and forth through the narrowly-defined confines of Shopify’s API, despite relatively simple business logic.
In a nutshell, integrating with traditional APIs is a costly endeavor and a liability for technical organizations. A better way, however, has emerged.
A big step forward
Recent progress in AI, however, has revealed a promising new approach to software interconnectivity that minimizes these issues with traditional APIs and enables anybody - even nontechnical users - to build powerful, integrated software using simple English statements.
The breakthrough known as “Generative AI” is a new breed of AI that generates complex artifacts such as stories, poems, images, videos and even - you guessed it - code. In order to construct an artifact, such as the code for an API integration, you feed the algorithm natural language instructions, or simple statements for what you want. OpenAI's GPT-3, for instance, has demonstrated a shocking ability to generate semantically and syntactically coherent bodies of text, including working code and complex stories. All it requires is a description of what you want it to produce, e.g. “a story about XYZ.” The paper describing this work was published in 2020 and quickly became one of the most highly-cited papers in all of computer science; since then, the field has taken off into uncharted territory, with new networks able to generate even full books based on text prompts (see Figure 1).
Some of the most impressive applications for GPT-3 and similar models is in the generation of code. Developers can query OpenAI's codex, for instance, with business logic requirements expressed in english and expect coherent, working code that executes this task. (See figure 2) Models like these are already embedded within developers' code development environments, such as Github Copilot and Replit's GhostWriter, enabling significant productivity gains via “intelligent autocomplete” - instead of writing the full code, programmers can essentially declare their intentions and the AI will autofill in the rest. Alphabet, for instance, has shown a significant increase in productivity when using automated code generation tools. This functionality is particularly helpful in scenarios where developers (1) have relatively simple logic they need to implement and (2) their task requires a large amount of contextual knowledge, which makes it a natural fit for aiding in the development of API integrations.
These automated code generation tools obviate the need for deep understanding of specific API conventions, radically increasing the speed at which developers can leverage interconnections with libraries and external platforms within their software.
The Evolution of Software Interconnectivity
As generative AI improves, we foresee a natural progression towards natural language being the lingua franca of the digital world. This will unfold in three steps:
Developer (and platform) adoption of Intelligent Code Autocomplete
Millions of developers already use a form of intelligent code autocomplete such as GitHub Copilot or Replit’s Ghostwriter, enabling them to essentially describe what they want to accomplish and receiving back the code. We anticipate that specific software platform providers, such as Stripe and Shopify, who have commonly used APIs, will extend the functionality of the models powering Colab by providing custom versions tailored to working with their own APIs. That is, developers will soon be able to install a Stripe extension in their code editing environment that will translate natural language commands into Stripe integration code.
Software platforms expose natural language user interfaces
While many consumer products, including SuperHuman (the alternative email client), iOS’ Siri and Google’s Assistant, enable humans to interact with computers using purely natural language, this user interaction has yet to extend to enterprise and “backend” software. We anticipate a shift in which enterprise software platform providers such as Stripe, Shopify, etc. will enable external software applications to interact with them via natural language. That is, instead of writing code in a given company's API specification, or "language," developers will send plain English commands to a company's software platform - i.e. "cancel my most recent stripe Transaction" - and expect the tasks to be performed, much as though they were communicating with a customer service representative. This significantly lowers the difficulty level of accomplishing tasks with software and lowers the bar for who can “code” productively, democratizing access to the power of computers.
Natural language becomes the default integration method for application developers
While traditional APIs will never be completely eradicated and remain most appropriate for certain scenarios, i.e. those where latency and performance are of primary importance, an era in which software developers primarily use natural language to build interconnected applications is imminent. There are obvious dividends here, including the democratization of access to software development: it requires significantly less training to understand the dynamics of systems passing english commands back and forth, compared to the arcane language of traditional API integrations.
A Vision of the Future
We’ve all seen clips from sci-fi movies where people operate complex interfacing systems with voice commands and spatial interactions. With this latest inflection point in AI, we are getting one step closer to that becoming a reality. Startups are already emerging to build towards this future, such as Adept which is “training a neural network to use every software in the world”, developing a model that can process large amounts of contextual information, understand and manipulate multimodal information and teach itself how to use software with limited human supervision.
Marc Andreesen quipped that "software is eating the world"; it's clear that meanwhile, AI is eating software, including even the means of communication between applications and even different components of an application. While it's unlikely that major software service providers will eliminate their existing API infrastructure on which so many customers rely, AI has the potential to democratize access to the power of software applications and empower the next generation of app development.