What makes a good AI co-pilot?
And why Clippy and Siri failed at being one.
While in a previous blog post I already covered, why the future is assisted, in today’s post I’ll focus on what AI Assistants (=co-pilots) look like today, what distinguishes them from smart chatbots and not-so-smart speech assistants, as well as a checklist that of five properties that separates great from not-so-good co-pilots.
Definitions: What is an AI co-pilot?
In my view:
An AI co-pilot is a specialized software that you can interact with while you are working on an array of tasks and that assists you in accomplishing them by:
the provision of context-aware information that you need and/or
the ability to execute tasks for you that are
simple but time consuming or
require computational abilities you don’t have
Some history on co-pilots
The idea of co-pilots that assist us in everyday life is not new. The vision for a digital assistant in the workplace had been around at least since the 80s; such as in this vision ad by Apple:
An also in pop culture, assisted working became an ideal to be desired, such as J.A.R.V.I.S., who helps Iron Man design his suit, defeat his enemies and much more.
However, the implementations of co-pilots in actual software has been terrible for the past decades:
While pretty much every person older than 30 years remembers clippy as one of its first major commercial implementations of an assistant, few of us shed a tear when it disappeared.
Later Voice Assistants, such as Apple’s Siri and Google Assistant and and Microsoft’s Cortana never lived up to the expectations; being mostly used to set a timer for its users.
Why did co-pilots fail in the past?
So why should anything be different this time?
There are 3 key reasons why co-pilots failed in the past and where their visions differed substantially from what users could experience:
Not understanding users’ intentions
Not saving significant time for a given task
Never offering anything the users could not have done themselves
Solving these flawed aspects by using Large Language Models (LLMs) are the first 3 of 5 aspects that have changed the game around AI assistants:
What makes a great co-pilot?
1. It understands the user’s intent
There is nothing more frustrating that trying talking to a cold and soulless machine when it does not even “get” what you are trying to do. With the release of ChatGPT this barrier has been removed. While the art of writing good questions for AI (“prompt engineering”) is still a skill worth developing, a good assistant does not rely on specific keywords or sentence structures to derive meaning from the users’ instructions.
For this it needs to…
…understand multiple languages / colloquialisms / domain-specific language etc.
…have memory of your conversation to understand requests and additions in context
…be able to ask concise questions back when user requests are unclear
2. It saves significant time on mundane tasks
While the tasks themselves can be generic or completely domain specific, reducing implementation time by a factor of 10x makes all the difference:
Here it does not matter that the results might not be perfect or could also have been achieved by a junior person that took the notes; the fact that you have that result instantly at your disposal makes all the difference:
A good AI co-pilot feels like having a capable intern who assists you instantly.
It will make mistakes; its work won’t be necessarily on par with yours but it produces drafts for you instantly and with a bit of guidance lets you achieve your goals much faster.
3. It can do things you are not able to alone
While having an assistant that does things 100x faster than a human enables you to achieve super human results yourself; a great co-pilot goes beyond pure efficiency gains:
A great AI co-pilot makes you not only more efficient, but more effective.
While e.g. spell-checking, summarizing, pre-formulating drafts, etc. save you time and allow you to perform your tasks much faster, time savings are only the tip of the iceberg of what assisted work will look like in the future:
The real power of an AI assistant is when it helps you with complex tasks, such as:
planning
forecasting the implications of actions
decision making based on data
managing dependencies
While the underlying best technological approaches that will drive these abilities are still in their infancy (for the technically inclined, check-out e.g. Retrieval Augmented Generation (RAG) and Autonomous Agents), the best existing co-pilots today (08/2023) already start guiding users towards their results.
As humans we are very limited in our abilities to retain and recall information and to hold to many concepts in our head simultaneously. This is why assisted decision making that analyzes inputs, suggests alternatives and simulates outcomes for the user, allows them to take better decisions. This ultimately enables the creation (depending on the applied domain) better code, policies, medical treatments or products:
4. It lives where users work
While only the future will tell whether we will ultimately all have a single personal AI assistant that will follow us around all day or whether depending on the context we will interact with different assistants; it is clear that the co-pilots we will interact with will assist us inside the tools we use daily.
This allows the co-pilot to leverage previous work, get relevant context for each interaction, have domain-specific abilities and ultimately help with the task at hand.
This also means that many interactions with the AI assistant aren’t necessarily through voice commands or full requests like “improve this code”. Since the co-pilot is able to understand your intentions, its actions come in multiple forms, such as auto-complete, highlighting, smart buttons, popups and other forms.
Some of the best AI assistance features many of us use daily are not even visible in the form of an “AI bot”, but rather have become part of very subtle user-experiences. Many of us hit “tab” multiple times per day when writing emails with gmail forgetting that merely 5 years ago this functionality seemed so mindboggling, it drew laughs and cheers from a major crowd:
5. It makes you enjoy your work more, not less
Great co-pilots make you feel in charge and with access to super powers.
As new generations of workers and experts in each field will take for granted the ability to collaborate with an AI assistant, many companies will fail on the way to provide value to those users as they will turn out to be nothing more than a fancy gadget.
However, the assistants that ultimativley will shape our daily lives will be the ones that allow us to focus on the enjoyable parts of our work: being creative, making complex trade offs and design decisions,
In Conclusion
Large Language Models that are used in tools such as ChatGPT have brought us the technology that early assistants from the 2000s were missing to become great co-pilots by translating user intent into something the computer can act upon and giving us a tool to save time on mudane tasks. This technology however enables the true potential to be delivered by specialist solutions that make us not only more efficient but more effective, by providing co-pilots that live where we work, give us superpowers and make us enjoy work more; not less.
Personal plug:
Given my enthusiasm for great co-pilots, you won’t be surprised to learn we are currently gathering feedback for our Assisted Engineering co-pilot for hardware engineers. Give it a try!