Goodbye ‘Prompt Engineering’, hello ‘User Agent Task Assignment’
The appearance of AutoGPT and User Agent bots changes everything (again) — and may point the way to personalized fundraising at scale.
Back in February, as the world was still processing the sudden appearance of ChatGPT, a panelist asked a straw poll question “Six months from now will ‘Prompt Engineer’ be a job title?”
It seemed a reasonable speculation at the time, but it turns out that just two months later the AI world is already moving on.
The announcement of AutoGPT, along with BabyAGI and AgentGPT over the past few weeks heralds the arrival of a generation of new AI bots: ‘User Agents’.
User Agent bots raise ChatGPT to the next level of automation, taking over the task of ‘prompt writing’ and adding in other action-taking and reasoning capabilities to get us one step closer to something we could call a true ‘robotic assistant’.
Here’s what a User Agent looks like in action, solving a gardening problem:
Question: Tell me which plants to plant early in the spring in my garden.
I live in Ottawa.
===================
THOUGHTS: I think the best first step is to start a GPT agent to help me
with this task. I can use the agent to search for information
on which plants to plant early in the spring in Ottawa.
REASONING: Starting a GPT agent will allow me to delegate the task of
searching for information on which plants to plant early in
the spring in Ottawa. This will save me time and allow me to
focus on other tasks.
PLAN: Start a GPT agent
Ask the agent to search for information on which plants to plant
early in the spring in Ottawa
CRITICISM: I need to make sure that I am delegating tasks to the agent
that are within its capabilities. I also need to ensure that
I am providing clear instructions to the agent.
===================
THOUGHTS: I think the next step should be to browse a website to find
more information about the plants that the agent found.
REASONING: Browsing a website will allow me to find more detailed
information about the plants that the agent found. This will
help me to determine which plants are best suited for my
garden and how to care for them.
PLAN: Browse a website to find more information about the plants that
the agent found
CRITICISM: I need to make sure that I am browsing a reputable website
and that the information I find is accurate and up-to-date.
User Agents can take on much larger, more complex tasks than ChatGPT and perform multiple inquiry, summary, and reporting tasks all from just a simple set of starting orders. This includes the ability to generate ChatGPT prompts and fine-tune them using the same powerful Large Language Models that power ChatGPT itself.
‘Prompt Engineering ‘— the art and skill of crafting the best input prompt to generate a desired ChatGPT result — could become just another task that’s better done by machines than by humans.
User Agents combine elements of both Generative AI (the tech behind ChatGPT) and AI robotics, which have been developing in parallel. AI robots are able to perform complex actions by chaining together discrete steps — ‘tasks’ — toward a larger, more complex goal. The latest AI robots are able to generate their own task lists by breaking down a goal such as ‘Put the box on the table’ into simple steps:
1 — Bend down
2- Pick Up box
3- Move to table
4- Lower box
5- Release
The new generation of User Agent bots take this task-building ability and combine it with ChatGPT, adding in other connected functions such as browsing the web, reading documents, or even sending and receiving messages.
Here’s what a simple task given to a user agent might look like in the role of a personal assistant:
Goal: Prepare my morning briefing
1. Read my latest emails and identify important items
2. Check my calendar for upcoming events today and this week
3. Browse news site and gather important stories
4. Summarize all this info into a bullet list
5. Add a joke and an inspirtational quote
6. Send it to my email at 8:30 am every day
A User Agent will work through this list, breaking out individual tasks required for each goal.
Each task might be an action such as a Google Search, collecting text from a web page, reading a pdf or other document, or even writing a prompt to ChatGPT to generate a summary or analysis of compiled sources. User Agent bot can adjust this task list on the fly and generate new tasks as they assess the real-time results of its work.
How does this work in practice? Auto-GPT allows this task-generation and re-evaluation process to be watched in real-time. It makes for a fascinating view into the ‘mind’ of an intelligent AI agent.
Here’s a verbatim excerpt from Auto-GPT’s ‘thought’ process while working on a fundraising task:
===============
THOUGHTS: I think we should start by researching recent news stories related
to climate change to identify a strong topic for a fundraising
campaign for Greenpeace. This will help us create a campaign that
is relevant and timely, which will increase the chances of success.
REASONING: Researching recent news stories related to climate change will
help us identify a topic that is relevant and timely, which will
increase the chances of success for our fundraising campaign.
PLAN: Research recent news stories related to climate change
Identify a strong topic for a fundraising campaign for Greenpeace
Use this information to create fundraising products such as email
appeals, social media posts, and a radio or TV ad script
CRITICISM: I need to make sure that I am thorough in my research and that I
am using reliable sources. I also need to make sure that the
fundraising products I create are compelling and persuasive
==============
THOUGHTS: Based on the search results, I think we should focus on the recent
IPCC report that delivered a final warning on the climate crisis.
This is a timely and relevant topic that is likely to resonate
with potential donors.
REASONING: The recent IPCC report is a timely and relevant topic that is
likely to resonate with potential donors. By creating fundraising
products such as email appeals, social media posts, and a radio or
TV ad script, we can reach a wider audience and increase the
chances of success even further. Identifying key internal and
external audiences who would be the best targets for fundraising
appeals on this theme will help us to tailor our message and
increase the chances of success.
PLAN: Use the information from the recent IPCC report to create
fundraising products such as email appeals, social media posts,
and a radio or TV ad script
Identify key internal and external audiences who would be the best
targets for fundraising appeals on this theme
CRITICISM: I need to make sure that the fundraising products I create are
compelling and persuasive. I also need to make sure that I am
targeting the right audiences and that my message is tailored to
their interests and concerns.
=============
THOUGHTS: We should now research and make recommendations for key internal
and external audiences who would be the best targets for
fundraising appeals on the topic of climate change. We can use
data and research to identify the most receptive audiences and
tailor our messaging and outreach efforts accordingly. We should
also consider the different channels and platforms that are most
effective for reaching each audience, and develop a comprehensive
strategy that takes into account the unique characteristics and
preferences of each group.
REASONING: Researching and identifying key internal and external audiences
is an important step in developing an effective fundraising
campaign. By tailoring our messaging and outreach efforts to the
unique characteristics and preferences of each group, we can
increase the effectiveness of our fundraising appeals and
maximize our impact. Using data and research to identify the
most receptive audiences will help us to focus our efforts and
resources on the groups that are most likely to respond
positively to our messaging.
PLAN: Research and make recommendations for key internal and external
audiences who would be the best targets for fundraising appeals on
the topic of climate change
- Use data and research to identify the most receptive audiences and
tailor our messaging and outreach efforts accordingly
- Consider the different channels and platforms that are most
effective for reaching each audience
- Develop a comprehensive strategy that takes into account the
unique characteristics and preferences of each group
CRITICISM: I need to make sure that we are using reliable and up-to-date
data and research to identify the most receptive audiences. I
also need to make sure that we are tailoring our messaging and
outreach efforts to the unique characteristics and preferences of
each group, and that we are using the most effective channels
and platforms for reaching each audience. Additionally, I need
to make sure that we are developing a comprehensive strategy
that takes into account the unique characteristics and
preferences of each group.This sort of complex reasoning and breaking down of larger tasks into sub-components can seem uncanny — and increasingly uncomfortable to those concerned about the development of AI intelligence that rivals or exceeds human intelligence.
Worth noting here that AutoGPT is not limited to only the pre-November 2021 dataset that was used to train ChatGPT. It is able to deploy its own search agents to collect live information from the web and consult any other source documents that have been pre-loaded or it can find.
What does the rise of User Agent bots mean for fundraisers?
User Agent bots trained for fundraising could be powerful game-changers.
As fundraisers, we’re familiar with the trade-offs that happen when we shift our focus from cultivating a single donor relationship to managing lists of thousands of donors.
Operating at scale shifts our focus away from providing personalized donor cultivation and towards the de-personalized mass-targeted appeals that we use, but don’t love.
With User Agent bots, we could take a step toward personalized donor cultivation at scale.
A well-trained User Agent (with human oversight) could research and create individualized cultivation journeys for every one of an organization’s donors by consulting the donor’s gift history, contact history, and preferences, and combining it with an organization’s own internal library and relevant info from news and online websites. The User Agent could even direct the delivery of this cultivation journey by writing and sending emails, responding to questions, scheduling appointments, and creating personalized web pages.
That’s a somewhat scary vision — and far from certain.
But the pace of change in the AI world right now is disorienting and shows no sign of slowing down.
I wrote earlier that ChatGPT won’t take our jobs.
Auto-GPT says “Hold my beer.”
