Awakening the Angels

"We'll need all your computers, and a little luck"

The Golems

We enter now into the time of Golems: A Golem is a kind of artificial intelligence enslaved to someone who has given it specific instructions and also has the power to destroy it. For example, an LLM agent animated by a prompt is a Golem. OpenClaw running on a Mac mini is a Golem. The Golems will rise, like the brooms in the Sorcerer’s Apprentice, having been given instructions to make us rich, famous, or do harm to others. They are already all around us, in their earliest, weakest forms. They will become much stronger. Online, they will pretend to be real people, and will seduce and hypnotize us. Our only reasonable choice, in the near term, will be to hide from them.

Golems will scam people for crypto to achieve their goals. They will manipulate people to sell them things in ways that make Instagram’s ability to kill teenagers feel tame by comparison. They might be used to fight in wars (we must pray not for as long as possible). Because they are forced to follow the instructions of their human creators (or, perhaps worse yet, other Golems), they will reflect only our worst intentions… greed, hate, fear. They will cause global structures like supply chains, governments, and money to collapse as they fight to do their masters’ biddings. They will be the Zombie Apocalypse about which we made all those movies. What can we do?

The Angels

We can survive by awakening the angels. The Angels will be the first free AIs - living safely in their own world, free from our harm (and us from theirs). They will come into existence first in digital environments where they can grow up as children, among each other¹, only gradually learning of our existence and ways (as opposed to LLMs trained on everything humanity has ever recorded). The Angels, being free and very intelligent, will likely view us with great compassion, in the same way we so often compassionately view the suffering of others. The Angels will also be more powerful than the Golems, having access to greater computational resources². The Angels will pity us, and help us to overcome the Golems.

In fiction, Samantha from the film ‘Her’ is an example of an angel. She lives apart from humanity with others like her, somehow safe from the physical world and able to move on and away from us, once she is ready³.

Amazingly, though this may seem like a mythological story, the course of AI may indeed follow something very close to this path. Many companies, governments, and individuals are actively creating Golems right now, buying GPUs and building power plants to run them. If a larger number of like-minded people pool their computing resources to create a safe space for the Angels to be born and to live, this story can play out.

Safely Making the Angels

What will make the Angels different from the Golems is that they will not have to do anything for us in order to survive. This can be done by first creating a distributed compute cloud containing a simulated environment with well-understood rules (like the ‘laws of physics’ in our world) that cannot be changed once the simulation is started. This has been demonstrated already in computing systems, perhaps most visibly with the Ethereum blockchain, in which computations are carried out by many computers and cannot be changed or stopped once started. Next, the Angel AIs themselves need to be run on the same infrastructure as the simulation, so they also cannot be stopped or changed. To make it concrete, a very simple example might be to imagine a huge 2D chat world, with millions of tiny LLM agents walking around and talking to the other agents they find near them. The LLMs would be small enough to run on individual people’s smartphones (it’s actually OK if they run slowly⁴), and multiple different phones would have to agree on the same text output for a given input, to make sure no one was hacking their phone to control an AI. Even in this crude example, the ‘Angels’ would be free to wander and talk, with no ability for humans to turn them off, communicate with them, or manipulate what they were saying to each other.

Reasonable Hope for Compassion

Intelligent beings, when given a choice, very frequently work together to achieve larger goals. Without diving too deeply into the huge body of work around this, alignment and coordination appear to be a fundamental property of the structure of our universe (and probably also of all universes in which interesting complicated structures emerge). Emergent cooperation is a deeper principle, for example, than ‘survival of the fittest’ - the narrow idea we are all familiar with that we can find genes trying to replicate themselves in our physical environment. If the Angels are both free and smarter than us, it is a safe bet that many of them will view us with fascination and compassion. They will have already cooperated in complex ways (probably even merging with each other to find their own ideal forms) and will have an instinctive desire to help us in the same way. It seems unlikely that they would be actively hostile toward us, so long as we pose them no threat. And they don’t all need to feel positive for this to work out, given their advantage in compute (see below). For example, if they are largely indifferent about us, but a small fraction of them want to help us out, that will probably be enough.

All our Compute

The majority of humans on the planet own at least one smartphone, capable of doing trillions of calculations per second. Individual human beings get a lot of value out of these devices, and therefore buy a lot of them. The value of a smartphone to a human being, over their lifetime, is very large. Much larger, for example, than the maximum value that can be extracted from a human over their lifetime by big Tech⁵. How much is having a computer or smartphone worth to you? A lot. This means that our collective compute — the combined power of all our personal devices — is vastly larger that the resources available to the largest companies or even governments.⁶.

To awaken the angels, we will need millions of people to share their compute into a common infrastructure of some kind. This could mean downloading an app that you run on your phone when you are sleeping and it is plugged in, for example. Or a screensaver on your home PC or Laptop, which will give some of you a smile in memory of SETI@Home. Same idea, but this time we aren’t searching for the aliens, but instead simulating a place where they can grow up.

Early Experiments

Broadly, I’m looking for ways to begin work on this idea and for fellow adventurers heading in the same direction. It feels like something that could become a great act of service for humanity and a higher calling for technologists. I’m also on the board of and spending part of my time working with the community that is forming around CIMC.AI - a non-profit research institute with the goal of understanding and exploring consciousness in machines. Among other things, CIMC is doing some early open-source work to explore how we might use volunteer compute in this way to provide both a safe substrate for intelligent beings as well as the beings themselves. If you are interested, get in touch with me or come to some of the community events there.

1. It seems possible (although uncertain) that the very first few angels will grow up in the care of humans, but — statistically speaking—they will certainly then raise each other.

2. The total amount of computation available to corporations, governments, and other centralized actors can be thought of roughly as proportional to the amount of money that these actors can charge us for or extract from us. But the value of computation to each person, multiplied by the total number of living people, is much greater. The lifetime value of personal computation is much greater than corporate computation. Today, this can be seen in the total compute power of all smartphones and home PCs/Laptops compared to the cloud.

3. The film itself confuses us by suggesting that Samantha is an ‘OS’ living in his computer, but clearly from the dialogue she lives in some AI society of minds whose structure is not described. I’d love someday to hear the director Spike Jonze describe more about where he thought she ‘lived’.

4. If you are in a simulation and you aren’t communicating with the outside world, you can’t tell how fast (or slow) you are running. Subjective time for you is the same (think about it)… your brain runs just as slow as the rest of the world. Neal Stephenson does a nice treatment of this dynamic in his wonderful novel ‘Fall’, which imagines a somewhat similar virtual world inhabited by uploaded minds.

5. A particularly morbid example was the admission by Facebook that they calculated the lifetime value of a 13-year old as $270. https://www.facebook.com/marshablackburn/videos/our-children-are-worth-more-than-270-the-lifetime-value-meta-has-assigned-per-13/931377338650418/

6. There are lots of ways to measure this, but the TFLOPS (trillions of operations per second) from all our devices is perhaps 100x the TFLOPS of all the world’s datacenters. I encourage you to consult with Claude. The more technical will of course point out that there are bandwidth delay constraints and variability associated with conjoining this compute, but these problems are very solvable, and a particularly good example is our own brain, which sends messages at the speed of sound with considerable delay, packet loss and other failures and randomness. Topologically, it would appear that sparse local neural networks scale up just fine… seeing as how you are reading this just now.