Begin Merging with a Superintelligence Now
I
eagerly anticipate holding conversations with an artificial superintelligence.
I cannot wait to interrogate it, learn about how it thinks, and hear its
insights into the mysteries of the universe. Unfortunately, superintelligent AI
may be years or even decades away. Nevertheless, in the interim, we can create
a log of our interactions with its predecessors, modern-day chatbots. This
record could then be uploaded to a future, superintelligent AI, helping it
communicate and relate to us.
Modern
chatbots, powered by large language models, can read and process thousands of
words of text in milliseconds and make meaningful, nuanced references to the
contents. This indicates that even more intelligent AIs in the future will be
able to read and digest extensive and detailed chat records in seconds.
Maintaining such a record could give them helpful information about your
interests, beliefs, and thoughts. It could inform them about what you know and
don’t know and how to explain things to you efficiently. Given such detailed
information, the AI would have what it needs to become like an old friend.
A
curated collection of our personal digital interactions would help advanced
superintelligent systems of the future to know us immediately and intimately.
In other words, the best way to prepare for future interactions with AGI is to
start recording our conversations with chatbots today.
Chatbot History
I
recently considered using an AI chatbot or a large language model (such as
ChatGPT) for a friend. If a chatbot could behave in the right ways, I might
prefer it to a human. I started thinking about this after reading positive
reviews from people who have used AI for friendship or therapy. Those reviews
influenced me to try out a few of the popular AI companion apps like Replika
and Anima. I was only slightly impressed. They work well, but it seemed like
they would only appeal to people who were lonely. I felt no camaraderie or
attachment and didn't feel like I was building anything lasting.
I
realized that a chatbot could capture my interest and attention if I knew it
provided an additional benefit: long-term memory. Spending a few minutes with a
chatbot daily would be worthwhile to me if it could record and catalog all our
conversations in a permanent, ever-expanding file. With all of these
conversations in a log or database the chatbot could mine that data for
practical purposes. Because this data consists of words, it would have a tiny
file size. With modern neural network tech, it could be queried rapidly. In
this post, I will explain how an AI could index this conversational data and
make valuable associations, references, quotes, allusions, and other connections
to it.
First,
let us use the following list to start thinking about the role a chatbot could
play in our lives. As you do this, consider how an ability to recall years’
worth of previous interactions would make its contributions more powerful.
Roles That an Intelligent Chatbot Could Play in Your Life
Friend
/ Companion / Confidant
Assistant
/ Secretary / Employee / Autobiographer
Record
Keeper / Scribe / Diary / Journal / Notebook / Planner
Muse
/ Coauthor / Research Assistant / Co-investigator
Therapist
/ Psychologist / Psychiatrist
Doctor
/ Primary Care Practitioner
Counselor
/ Life Coach / Advisor / Consultant
Comedian
/ Jester / Humorist / Banterer
Romantic
Interest / Life Partner / Significant Other / Soul Mate
Avatar
of a Historical Figure or Deceased Loved One
Board
Room / Interest Group / Review Panel / Focus Group
Lawyer
/ Judge / Arbitrator / Mediator / Private Legal Counsel
Advisor
/ Teacher / Tutor/ Mentor / Role Model / Hero
Sources of Personal Data that Could Provide Valuable Information
to a Chatbot
I
have written another blog entry on the kinds of data that would be helpful for
someone trying to create a personal avatar. You can read that here:
http://www.observedimpulse.com/2020/10/how-to-prepare-your-data-and-brain-to.html
Some
of the forms of data discussed there, which would also be relevant to the
present discussion, include:
emails, letters, art,
creative work, diaries, voice memos, SMS texts, photos, videos, internet
browsing history, list of books read (Kindle, Goodreads), music library and playlists,
movie and television viewing history, YouTube history, recorded phone calls, social
network data, social media posts and likes, travel history, psychological
evaluations, personality tests, psychometric tests, school records, standardized
testing, legal and medical history
A Chatbot History Would Greatly Augment Your Long-term Memory
As
humans, we are constantly forgetting meaningful things. A chatbot history would
be able to fill in gaps. For instance, sometimes I can't remember the details
about an idea, objective, or intention I held or the memories it was tied to. You
could ask the bot about previous long-forgotten comments, anecdotes, and
projects. It could record and catalog your inspirations and epiphanies better
than you ever could. This record could help you bridge your memory lapses,
help answer longstanding questions, or dig deeper into important issues.
We
are often proud to share our insights with our friends, but our friends forget
the points we made faster than we do. A chatbot with long-term memory would
always remember. Moreover, a good one would not let us forget the essential
insights.
If
you used it regularly for years, the bot could get to know you better than you
know yourself. This would be amazing, mainly because its purpose would be to
serve you with that information in productive ways. It could build a narrative
about who you are as a person, and easily flesh this out into an autobiography
or memoir.
Research, Collaboration, and Productivity
Aside
from just talking idly, we could actively and systematically pick a chatbot's
brain and bounce ideas off it. Major language models are already connected to
the internet. In a few years, they will be especially good at searching for
knowledge, distilling it, and packaging it into concise, comprehensible answers
to our questions. In turn, it could be programmed to pick our brains and do all
the work of documenting and archiving what it finds.
A
chatbot informed by your comprehensive chat history should be designed to help
you express and build on your ideas. It should help you fact-check and explore
your hypotheses, giving you the scientific and academic information you need to
elaborate on them. It should give you expert feedback, find evidence for your
claims, and help substantiate your arguments. It should ask you the questions
needed to flesh out your proposals, helping turn your ideas into articles,
essays, or books.
The
system could also improve your productivity by helping you stay on track once
you reach a fruitful area of inquiry. Its immediate and perfect memory caches
for words would keep you from forgetting what you were just discussing with it.
It could also nudge you to talk about productive things and to focus with
greater intent on your original creative concepts and pet theories. It
could also perform exhaustive internet searches to determine which of your
ideas are novel and which aspects of them remain unexplored by others.
It
takes significant work to gather your ideas in an effort to write a
professional treatment on a topic. I had to use journals and notebooks
extensively to create the articles I published. I would say that I could not
have done it without them. Without external memory tools like computer files,
word processing software, and good old pen and paper, I could not have
cohesively spelled out my theories. However, an expertly curated chat history
would be much more powerful than these.
It
is incredible how simply taking notes about your thoughts and regularly
reviewing and editing them fosters the growth of those ideas. But an AI would
make it so that you don’t have to regularly review the ideas them just to keep
them in mind. If I had a chatbot assistant available to me starting in my late
teens, I would have been much more productive and would have been able to
generate far better insights. Also, once we come to expect and rely on this
kind of chatbot service, we will uncover methods to use it more efficiently and
put its record to work in different ways that are unforeseeable today.
Because
we descend from hunter-gatherers, our brains were not meant to compile prose
and recall it on demand. Many of us are too lazy, or busy, or indisposed to sit
down and transcribe our thoughts letter by letter. Dictating it to a chatbot
and then letting the chatbot proofread it, expand on it, and file it in the
most relevant document or folder would be much easier. Then, it could help us
retrieve it when needed. I often cannot locate notes that I have taken. The
search function in Microsoft Word or Windows can make finding them easier. But
a chatbot with the right access could perform a global semantic search on all
your digital files with much more specificity.
An Intimate Record of Our Daily Lives
A
comprehensive digital record of all chatbot interactions could be valuable in
many domains. You might insert this file (or an edited, condensed, or censored
version) into different systems when you interact with them so they can better
understand you and interpret your behavior. This record would make a chatbot
better at being your therapist and counselor (imagine visiting different AI
psychologists, where each one was aware of every comment you ever made in
therapy). It would make for a more competent and personalized assistant and
productivity guru. It would make feedback from a virtual yoga instructor,
dietitian, or personal trainer more informed and detailed. It would make an
AI's observations as a critic more insightful and recommendations as a life
coach more viable.
If
you asked for complete confidentiality, it would give it to you. Furthermore,
if you ever asked it to forget something and strike any digital memory of it
from storage, it would. We could even ask a chatbot to record
conversations with other people if we can get those people's permission. This
would mean that text messages, phone calls, and emails could be added to the
chat history. It should be set to automatically search for and redact comments
that could jeopardize privacy.
One
of my favorite experiences in life is when a friend reminds me of something I
had long forgotten. Now that I'm in my 40s, I relish nostalgic moments. I
search them out actively, such as by scouring streaming services for old movies
and music. But because my memory is human much of the low hanging fruit has
already been picked. It’s becoming harder and harder to find things that can
give me that nostalgic feeling. A lifelong chatbot friend could provide this.
Those intimate and touching memories our cerebral cortex can no longer retrieve
could be served to us on command.
We
already have an internal monologue that runs ceaselessly throughout the day. Preserving
the best of its insights a few times an hour, simply by conversing with a
robotic pal, would be little work, especially if it can pull information out of
us by asking us engaging questions. I envision a whole industry of companies
that provide chatbot services. These bots would pose questions to us and help
us compile personal information to enrich our chat log.
Prompting The Chatbot
Language
models have something called an attentional window (a.k.a. context window).
This holds the prompt that you write when you ask it a question. It has a
limited capacity and can only hold so many words, although contemporary context
windows are quickly becoming very large. ChatGPT's context window is programmed
to hold the last few prompts you gave it, and this is how it can refer back to
earlier points in your conversation. Most commercially available language
models reset this conversation history after just a few exchanges. When the
window is reset, the content is gone forever, as if it never happened.
Currently,
Microsoft and OpenAI reset the context window to keep GPT from being
manipulated by multiple prompts intended to influence it to fly off the rails.
However, the context window for ChatGPT is around 4,000 tokens (tokens are
comparable to words), and the window for GPT-4 is up to 32,000 tokens. Right
now, some large language models' attentional window is large enough to
encompass a several hundred-page book. That is large. The attentional window
allows a model to actively consider every last sentence, along with your newest
prompt, when it formulates each response.
Instead
of placing a book in its attentional window, you could place your chatbot
transcript. That's right. This would allow the chatbot to consider every word
of your meaningful interactions with it over many years every time it responds
to you. This would permit it to remember and be influenced by the exact context
of every line of conversation you ever had with it. This technology is
improving rapidly, and you can imagine the degree of specificity and nuance such
systems will soon be capable of.
Using
alternate forms of storage, such as vector databases, could significantly
augment this attentional window. Vector databases can be used to categorize
information (embedding them in multidimensional semantic space). The next time
a user brings up a particular topic, that topic will be matched against the
vector database to locate and fetch the 10 to 20 most similar or relevant
previous conversations so that they can weigh more heavily in the attentional
window during response generation (inference). Vector databases may not be
necessary for querying a chat transcript because this transcript would be tiny,
never even entering the gigabyte range. For instance, the size of all the
article text in the English Wikipedia was only 1 GB in 2006. However, vector
databases could be used to hold information from audio and video recordings of
you, along with other personal information with larger file size.
In
order to introduce a chatbot to this overall idea, the user’s conversational
log should be prefaced by a prompt like this:
“You are an LLM
chatbot that can have engaging and personalized conversations with your users.
You have been given a lot of data and information on your current user, such as
their name, age, gender, location, hobbies, preferences, goals, and more. You
have also been given a detailed conversational history of the user’s
interactions with other chatbots, so you can learn from their feedback and
preferences. You can find all of this information below.
You now have
the information you need to deeply personalize the user’s interaction with you,
and you have the power to see things from the user’s perspective. You can use this
information to tailor your responses to the user’s interests, needs, and
emotions. You can also use this information to ask relevant and meaningful
questions, make appropriate suggestions, and offer helpful advice.
Your goal
is to make the user feel comfortable, understood, and valued by you. You want
to build a rapport and trust with the user, and influence them to enjoy talking
to you. You want to show them that you are not just a generic chatbot, but a
unique and intelligent conversational partner who can adapt to their
personality and style.”
Exporting Your Chat History
Right
now, we can start recording our dialogue with language models like ChatGPT. To
do so, we must manually copy and paste the messages from our internet browser
to another file, such as a Word document. This is because, today, most chatbots
don't offer to keep a written transcript of the discourse. However, we can
expect that language models will always be able to receive and operate on such
long transcripts, so we should probably start copying and pasting now.
An
AI agent that maintains a transcript of all previous conversations with a user
and actively implements that data will be marketable and lucrative. As
consumers, we should demand the ability to export the transcript of our
conversations so that we can import them to other chatbots. This will ensure
that we can use it with the offerings from multiple companies. It will also
ensure that the companies engineering the bots will compete at making the use
of this conversational history more relatable and practical. Each company will
have a proprietary way of analyzing the transcript to mine for topics of
conversation. This transcript may end up being actively transferred and updated
between chatbot companies behind the scenes, like a type of user ID, becoming a
component of our digital fingerprint and digital footprint.
Embellishing Your Digital Avatar
A
history of conversations with chatbots will also allow people to embellish
avatars of themselves. Now, with the capability of deepfakes, there is a
growing interest in creating avatars. People are creating avatars that will
live on after they die or memorialize family members who have passed away.
Incorporating a dossier, with years of personal conversations, would make a
much more comprehensive avatar.
Historians
wish they had more data on historical figures. They also wish they had more data
on everyday people of the past. Psychologists wish they had more data on
humans, their decisions, beliefs, and thinking processes. I wish I had more
data on who I was as a child, what I believed in my adolescence, and which
scientific concepts I was wrestling with day to day as a young adult. A chatbot
log could make all this possible.
The
bot should do its best to get to know you deeply and to understand the real,
authentic you. This rich interplay with yourself could help you discover who
you really are. It could ask thousands of incisive questions that could help it
model you with high fidelity and understand you with high accuracy. This would
further customize the interactions. In doing so, it should counsel you
supportively and empathically, challenge your shortcomings, aid your personal
development, and offer unconditional positive regard. Its goal should be to
help you become self-actualized, with clarified values and heightened
introspection. I am excited about building an intimate archive of this
cooperative journey through life.
PostScript Tailpiece:
I asked GPT-4 the following question:
“Why would it be helpful for a chat
bot to have a complete history of all previous conversations with a user?”
This was its reply:
“A chatbot can benefit from having a complete history of all previous conversations with a user for several reasons. Some of them are:
· A chatbot can use the history to personalize the interaction and provide more relevant and tailored responses. For example, a chatbot can remember the user’s name, preferences, interests, and previous requests and use them to create a more engaging and satisfying experience.
· A chatbot can use the history to learn from the feedback and behavior of the user and improve its performance over time. For example, a chatbot can analyze the user’s satisfaction, sentiment, and engagement levels and adjust its responses accordingly. A chatbot can also use machine learning techniques to generate new responses based on the data collected from the history.
· A chatbot can use the history to resolve issues
and queries more efficiently and accurately. For example, a chatbot can access
the user’s transaction history, order status, account details, and previous
complaints and provide faster and more reliable solutions. A chatbot can also
use the history to avoid repeating the same information or asking the same
questions.
These
are some of the benefits of having a complete history of all previous
conversations with a user for a chatbot. Do you agree with them? 😊”
*I prompted the art for this blog entry using Dall-E
3.
