GPT-3 and the Rise of Human-centric Adaptive Software — Part 3
💡 This article is part of a series: Intro, Part 1, Part 2, Part 3
The need for a human-centric design
As I showed in Part 1 and Part 2, GPT-3 ability to transform human language into computer language removes the need for cumbersome, one-for-all user interfaces and allows us to finally design products for humans.
But designing for humans is not as simple as adding a conversational UI to an existing products or building a bot to chat with. It’s certainly not the simplistic “design with the user in mind” that many preach, but few execute.
It’s a paradigm shift that requires a significant change in the way we design products.
I’m going to describe it through three “design rules” I came up with in my journey while building Kyber, a productivity suite on top of messaging:
- Start from the human, not the machine
- Choose the most appropriate human-to-machine interaction
- Think like a human, not a machine
To keep the theory as practical as possible, I’ll present each design rule with an example of a current design solution, how it compares with the human-centric design used in Kyber and the take-away for GPT-3.
Start from the human, not the machine
The status-quo
As noted before, current project management tools are designed based on how computers work internally: forms with input fields that map 1-to-1 the internal task data structure are presented to users to create tasks; tables organized like the internal databases are then used to display the tasks. As a result, these products are utilized as system of records that humans are forced to insert into their natural workflows, making it difficult to adopt them.
In fact, this is not how humans come up with tasks, communicate them or in general work together. Humans work together by talking to each other, by brainstorming on ideas that generate tasks, by assigning those tasks with a simple “Can you do this?”
The human-centric approach
By starting from this human-centric perspective, I ended up designing our project management suite on top of messaging, to turn the important conversations between humans into actionable tasks right when they happen, right in the natural user flow.
Our project management solution becomes a system of engagement, an intelligent layer where the machine understands the human language, automatically extract important messages and parse their entities on behalf of the human to add structure and trigger workflows. And the human can just behaves as a human.
The GPT-3 take-away
GPT-3 ability to understand human language can enable those easier to use, “personalized” interfaces described before — a significant improvement for the human-to-machine interaction. But to take full advantage of GPT-3 innovation, simply adding a conversational UI to your existing product won’t be enough — a completely new design process, one that starts from the human, is required to ultimately build new kind of products.
Choose the most appropriate human-to-machine interaction
The status-quo
Few years ago, in the early days of the chatbot craze, everyone started turning existing products into chatbots: a human chatting with a bot was the design model to follow.
But tens of messages back and forth to order a pizza were not a good idea in the first place; finding, opening and messaging a bot just to ask weather information was not the most efficient solution; having elaborated, human-like conversations with bots was clearly not yet possible from a technology stand-point. As a result, many chatbots failed to gain any traction and the hype quickly vanished.
The human-centric approach
When you assign a task, you don’t ask an intermediary (the bot here) to first create a task and then to tell the assignee about it.
Instead, you just talk directly to the assignee: “Mike, can you do this?”
As such, I decided to follow a model where the machine enhances an existing **human-to-human interaction (a conversation) instead of forcing an unnatural human-bot-human flow.
Tasks are created by the machine as they happen in human conversations, removing the need to switch to external system of records and eliminating the cognitive overload of a classic user interface.
The GPT-3 take-away
GPT-3 can be used to interface humans with machines in many different ways — via conversational input forms, via chatbots, as a generative layer on top of content or as a cognitive layer on top human-to-human conversations. Picking the right interaction in the context of your product is going to be very critical to its success.
Think like a human, not a machine
The status-quo
Today’s workflow builders, i.e. low-code tools that allow automating workflows or building simple applications, are designed, guess what, based on how a computer works. Specifically, how a computer “thinks”: IF (this happen), THEN (do that).
The IF-THEN is a common conditional statement used by machines to execute certain actions based on certain conditions. Even in workflow builders for language-based apps like messaging or voice assistants, designers simply exposed that and similar programming logic to users. Practically, a user is asked to be both engineer in providing all the logical steps and product designer in providing all the details of the workflow. As you can imagine, these products remain hard to use for the vast majority of the users.
But is this how humans actually think when they want to accomplish something?
The human-centric approach
To make our solution even more engaging and collaborative, I decided to offer our users several messaging-based workflows (or “micro-apps”, as I call them) around project management: standup meeting, sprint prospective, sprint retrospective, project mood check-in, etc.
Instead of building each one of them independently, I tried to figure out a more generic model — a micro-app builder — that also users could use to create their own micro-apps, in the simplest, most intuitive way.
By studying how humans think, I realized that to describe a workflow we start with the goal we want to accomplish, not with a conditional statement: “I want to create a task”, “I want to ask a question”, “I want to send a message”.
Therefore, I designed our micro-apps around what I call “conversational primitives” (e.g. task, question, message, etc.), offering for each just a few “parameters” (e.g. recurrence, open vs. closed questions, public vs. anonymous answers, etc.) to configure the associated workflows.
With this approach, our users simply start from their goals (e.g. asking a question), configure few parameters and, in just 2 steps, create a variety micro-apps to cover the most disparate use cases. As a comparison, current workflow builders would require tens of logical steps and details for each step to build the same workflows
To verify the flexibility and efficiency of this new model, I tried to turn random Harvard Business Review tips on how to manage team projects into micro-apps, just by customizing and configuring our basic primitives. I couldn’t find a single case where this model wouldn’t work.
Fun GPT-3 fact: my team provided a JSON representation of each micro-app to allow me to easily create new ones or make changes; we can now use that same JSON and a few micro-app description to train GPT-3 and provide our users with a way to create sophisticated micro-apps just by describing them in plain english.
If you observe, our micro-apps represent the Modules component in the GPT-3 Application framework discussed before and that’s why GPT-3 can be seamlessly integrated with our product.
How such an integration can be so straightforward if our workflow builder was designed way before GPT-3 launched? It’s based on how a human thinks, something that GPT-3 likes quite a bit.
The GPT-3 take-away
GPT-3 understanding of human language bring us closer than ever to the user and its way of thinking. To naturally leverage GPT-3, products will have to be designed based on how the user thinks. As a consequence, these new type of products will be much more powerful and more intuitive to use.
🤝 Building a GPT-3/AI product? If you need any help, happy to work with you.
Originally published at https://ppaolo.substack.com.
