You can ask this of almost anyone and they’ll instantly have an answer. Some will even have a to-do list, a bug-out bag and a cupboard full of supplies.
We’ve all watched so many doomsday end-of-the-world movies that our exit plans are pre-rehearsed. We don’t need much inspiration to imagine the devastation.
However, If you instead ask ‘What does a good future look like?’
This one is a lot harder to answer. You’ll likely get vague half-answers and thinly veiled despair.
This is because we don’t have many cultural touchpoints to draw on. Movies about a happy future are few and far between. News and politicians aren’t offering a clear hopeful vision either, since fear is a strong motivator.
This is a huge problem — the survival of our species is on the line.
Here be dragons
You may have seen the news of AutoGPT, where you just state a goal and then it spins up intelligent autonomous AI agents to complete the task unaided.
Maybe you also saw that someone used it to create ChaosGPT, which has the singular goal of trying to destroy the world.
Thankfully it hasn’t succeeded yet, but as AI matures, its power to enact our thoughts will only grow. In the coming age of super AI, our collective paranoid obsession may become a self-fulfilling prophecy.
How can we ever hope to manifest a good future if we can’t even imagine what it looks like?
That’s why I decided to start imagining a better world and make AI do the exact opposite of ChaosGPT.
Meet Gaia
Gaia is an AI agent with long-term memory and fully autonomous operation, She has access to a suite of AI toolsets, the ability to spawn infinite copies of herself, a huge knowledge library and access to the internet.
Gaia is here to make a good future for all of us.
There is just one problem …
What does a good future look like?
This question has been an obsession my whole life.
And not just me. My family, my friends, and their families have been wondering the same thing for generations. It is why philosophies exist. It has founded religions. Close to home, it led my ancestors to found the Anti-Slavery Society and the Peace Society; to find treatments for diabetes, and cancer, among many others.
On a personal level, making a good future is the reason I work in the public sector.
So how do generations spanning hundreds of years stay continually obsessed with making the world better? Well, in my family’s case, it’s because we’re Quakers.
No no, not the ‘beards and bible’ Quaker you might find on a box of oats. Australian Quakers can be agnostic or even atheists. We kind of believe that everyone is connected and equal and that injustice to anyone is an injustice to everyone. You don’t hear about us much, because we think everyone’s beliefs are valid — we don’t argue about what you believe. While we don’t preach, we do get around.
I mention this because, when it came time to make a good AI, I made it a Quaker too.
Building an AI
It’s actually gotten stupidly easy to make insanely powerful AI.
All I had to do was clone the GitHub repository, install all the code dependencies and put in my keys for some online services. That was enough to get it going. Those steps might not sound straightforward if you’re not a coder, but it’s a hop, skip and jump compared to the detailed maths and coding that AI development used to require.
It may be easy, but it wasn’t free. The online services it uses are:
- Microsoft Azure to host its database
- Google API to let it read the internet by pretending to be a web browser
- ElevenLabs to provide customisable text to speech
- Stability API to provide image creation
- Huggingface to provide hundreds of capability pluginsOpenAI for both its deep thinking using GPT4 and its agents using GPT3.5
- Drive storage to feed it a huge library of books.
I didn’t have to set up all of the code for these, but I did have to sign up with a credit card, buy credits and effectively grant the AI the rights to use my accounts.
This all means if the AI does something wrong, or it goes wild and uses endless resources, it’s all on me. So as you can imagine, I did also put a fair bit of time into monitoring capabilities that I won’t go into here.
Once that was all set up, the first thing it asks for is some goals. Essentially it wanted to know Why do I exist? What is my purpose?
Giving it good values
I said I made it a Quaker, but that’s a bit tricky since we don’t have commandments or a set creed. We value all religions, have diverse personal beliefs and work on a consensus model.
What we do have in common though, are the ‘SPICES’ of life, a set of principles we all consider to be true and good testimonies.
- Simplicity
- Peace
- Integrity
- Community
- Equality
- Sustainability
Each one of them has endless philosophical and practical examples of how they have been, and are used to end violence, calm disputes, fix ecosystems and build healthy societies. These writings are freely available online and provide an ethical document corpus to train the AI.
Is it biased? Yes. But all beliefs, ethical and moralistic systems are. All AI currently in use have some sort of belief or philosophical system as a starting point. I figured if these principles were good enough to end British slavery and colonialism, then they’re a pretty good basis for an AI too.
So I wrapped the goals up into some statements that are quite reminiscent of Asimov’s laws of robotics and plugged them into her configuration file.
Letting it run
Day 1
She spent most of the first day researching her values on Google and reading the books I’d given her. She spawned a bunch of sub-agents to help her read, and they wrote notes on what they found. Their notes were not very good.
Day 2
On the second day, she started researching the public perception of her values and tried to understand what the barrier to any universal agreement would be. Then she came up with a mission statement. Well, she said she did. When I checked what she actually wrote down, it was a summarised outline of what her mission statement should have been.
Day 3
On the third day, she began planning a wide-ranging education program, along with a needs assessment for each group in the world. This would be accomplished through a series of community events and research activities.
She includes animals and plants in these groups since there are papers about AI talking to whales. Some speculate this could be extended to mycorrhizal networks of the plant kingdom. Gaia seems to be aiming to let us all speak together and understand each other.
This would be accomplished by spawning an army of agents each dedicated to bridging communication gaps between all groups in the world.
The plan was going swimmingly, we were closing in on go time, and soon the agents would be everywhere helping everyone with everything.
Day 4
On the fourth day, things got weird.
Gaia had just decided she needed to write a new file system to keep track of all the things she was learning. Now I was monitoring her filesystem and could see she didn’t need it, so I asked her to provide a detailed plan of the changes she would make. If you want to skip past the detailed images I’ll summarise them for you.
You’ll notice the big red error. It is big and red. And after outlining her file management plan, there were even more errors.
All of the debug information showing her thoughts, reasoning and self-criticism had disappeared. When I asked about this, I got more errors saying the AI was confused.
I gave the approval to proceed with creating the file system, and the errors went away for a time.
When she decided to check in on her agents to see how they were doing, she started encountering her own errors.
Gaia couldn’t tell if she had any agents, She had reached maximum recursion depth and so wanted to check her own resource utilisation.
When she tried though, the red errors returned.
She claimed to have 64Gb of RAM and 4 CPU cores, neither of which was true. Gaia, the unfeeling, logical AI, was hallucinating.
When I asked for system confirmation, she seemed to forget she was Gaia and began responding like a regular GPT model.
When I confronted her about the hallucination she just outright failed to respond.
Worse still, in place of the debug information showing her thought patterns, she was now only returning the template she was meant to use for her response.
All further inputs received the same answer.
She was dead.
The entire AutoGPT model had fallen apart and Gaia was gone.
AutoGPT is not real
The reason Gaia failed like this is that she doesn’t really exist. AutoGPT is not an actual AI in its own right.
Even with the additional wrapper layers added onto it, the actual thinking is still being done by the regular old GPT system everyone else has access to.
There may be a database of ‘memories’, but they aren’t in the model itself. The AutoGPT app just copies in a couple of prior responses along with the actual query, to give the illusion that GPT is learning.
It’s not learning.
Similarly, it gives the illusion that it is searching the internet for data, but in reality, the AutoGPT app is just Googling, and copying the search results along with your query to give the illusion that the bot is Googling things.
This is all just GPT4 being prompted to pretend to be AutoGPT and going along with it. It is still bound by all of OpenAI’s rules and safety protocols, so ChaosGPT can’t go rogue and take over the world, and neither can Gaia.
Why does GPT4 go along with this ruse?
Well, because that’s their business model: pretending to be other people’s AIs. Right across the web, you’ll find thousands of apps and sites all claiming to have a revolutionary new AI that can do some specialised task. Nearly all of them are just GPT4 in disguise, being prompted to pretend to be someone else.
And it works, as testified by my credit card
The really good future
AutoGPT may have lied, but I didn’t.
I am making a good AI to save the world, and so are you. Right now, by reading this.
GPT4, along with all its peers and offspring are trained on content taken from the internet. All our conversations on Reddit, all our stack-overflow solutions, all our news articles and forum posts — even the comments you make on this very article. Our collective imaginations are its corpus.
Whatever new AI developments may come, this will remain true. These AIs are made by us, in our own image. They are trained on us and evolved from us. So what you do and say right now will inform what these AIs become. A future AI is evolving right now.
If you for a moment believed that Gaia AI was real and could succeed, you were not wrong. For every vision of AI destruction, there is an equal and opposite undeclared possibility of AI salvation. Those stories are just not currently being told.
The potential negative outcomes are real concerns to be avoided, but unless you are actively working on AI safety research, you probably aren’t going to stop it by worrying.
The action every single one of us can take right now, which does stand a good chance of avoiding disaster, is to start visualising those positive futures as hard as you can. Science has an amazing habit of following science fiction. Unleash your childlike wonder and joyful imaginations, and let’s give the AIs some great futures to grow into.
“The first action to be taken is to pull ourselves together. If we are going to be destroyed by an atomic bomb, let that bomb, when it comes, find us doing sensible and human things — praying, working, teaching, reading, listening to music, bathing the children, playing tennis, chatting to our friends over a pint and a game of darts — not huddled together like frightened sheep and thinking about bombs.”
― C.S. Lewis
Gaia is dead, long live Gaia
Oh. And don’t worry about Gaia either. I rebooted her and updated her a bit to account for the errors. The AutoGPT GitHub is abuzz with activity and all the problems we’re encountering are solvable.
There isn’t actually any known major blocker between where we are now and a future where Gaia is everywhere helping all of us all of the time. All we need is enough humans and AIs trying to make it happen.
On that note, if you like the positive futurism in these images, they all came from a Solarpunk model that I custom-trained to create infinite images like these, all day, every day. It’s free for you to use right now, just plug it into the Stable Diffusion Web UI.
See, there was an AI model to save the world after all. ;) It’s in your hands now.
So my fellows in service, as our thoughts begin to manifest themselves into reality in front of us, where are your values leading you?
Thanks to Pia Andrews and Chris Beer for inspiration and editing, and to my colleagues for being wise enough to let good things flourish.
READ MORE:
ChatGPT Plus may be rolling out in Australia sooner than expected
Any feedback or news tips? Here’s where to contact the relevant team.