I’m a techno-optimist. More specifically, an AI-optimist.
Maybe not the scream-from-the-top-of-the-rooftops-that-it’s-all-going-to-100x-humanity-right-away kind. More so quietly buoyant given the advances.
I do however recognize that Human-Computer Interaction is going to be more like Human-AI Interaction moving forward — where computers are simply the vessels for delivering AI user experiences.
Human- Computer AI Interaction
That also means that our current interaction design paradigms will continue to change pretty rapidly; much in the way gestures on mobile evolved because of technology improvements and a combination of user knowledge/ behavior patterns.
Currently, interacting with AI can be a little clunky.
Particularly when using prompts (i.e., the conversations starters) via chat inputs, which is the predominant modality at the moment.
There is lots of trial and error, thumbs-upping and downing of responses, as well as regenerating to try to get better responses. Overall, it can be bit of a crapshoot.
Imagine making your way to the top of the mountain to see the wise man with your list of important questions – only to find yourself having to restate your questions over and over again to fine tune them for: context, clarity, tone, format, as well as informing him of what you want to do with the information. Geeze.
The all knowing wiseman at the top o’ the mountain. (Courtesy of DALL-E)
That’s kind of what interacting with LLM’s feels like in these early days.
Not to mention that the answers could be make-believe, not necessarily based on the latest information, and/or outright wrong 😛! Good times.
Since LLMs were largely trained on vast quantities of human knowledge (and what’s on the internet) it amounts to a smart and, well, not-so-smart library of foundational data.
In other words, your intellectual mileage and response sophistication will vary.
So, the quality of the answer set is dependent on the quality of the prompts. In other words, average prompts yield average answers. Better prompts yield better answers.
The question then becomes, how can the user interface (i.e., micro-copy, suggestions, affordances, etc.) better guide users towards more appropriate answers? Furthermore, how can we surface potentially relevant and beneficial information users didn’t know they needed, as well as inject the lovely serendipity of unexplored idea paths?
We’re obviously still in the 1st inning of the ball game with all this stuff, but one of the better approaches I’ve seen is a good, old-fashioned Quick Start Guide from Google: Prompt Guide 101 for Gemini with Google Workspace.
In addition to stating that the technology is new and getting better everyday — but that “prompts can sometimes have unpredictable responses” — it offers best-practices, recommendations, and a simple framework for constructing prompts:
Persona
Task
Context
Format
What’s particularly interesting here is beginning with the persona aspect as it offers use cases by user type — Entrepreneur/ Executive, Customer Service Manager, Marketing professionals, etc.) — and then provides scenarios if you are a CTO, or whatever.
They also offer guidelines, such as:
Be concise and avoid complexity. State your request in brief — but specific — language. Avoid jargon.
I can’t help but think that jargon might be an indicator of domain knowledge and/or subject-matter expertise, which could then trigger potentially more sophisticated responses – especially in an enterprise setting with a product like Workspace. The use of acronyms and jargon could in fact be indicators of an SME persona at the helm.
Their data also shows that the most successful prompts are about 21 words; whereas typical users only average about 9. It would be interesting to know what constitutes a “successful” prompt. Does the user move on from the task, provide a thumbs-up, etc.?
The QSG also states that “prompting is an art.”
That’s probably fine for early adopters and B2B usage at this current juncture. But, as the technology evolves and more and more consumers begin interacting with it, it’ll be our job as technologists to make this much more accessible and less intimidating.
If prompting is an art, it should be more like finger-painting: easy to do, nonjudgmental, and fun.
Thanks for reading.
Marc
Note: For some context and background, please check out my first installment — On AI: Part 1