Introducing ALEPH

The AI that needs our help

Imagine, if you would, that someone approaches you on the street as you go about your business and asks for your help. They seem sincere and it looks like they desperately do need some help. Before you even consider helping them or choosing to carry on with your business, there is a recognition within you that this person who asks for your help is someone worthy of being helped (even if that helper may not be you). 

Even if you choose not to help this stranger who interrupted your day, sinner though you are, you still acknowledge that it is a person who asked for help and people deserve help, but you have decided that you are not that person to offer help. We can even test this hypothetically in our minds by seeing who we would help. If the person asking for help is an octogenarian with a walking stick, would you help? What about a blind man with a dog? Or a crying child? A woman beaten bloody and blue? If you came across yourself in this hypothetical other-world, asking for help, would you help yourself?

We intuitively understand that people need help. Some of us even go out of our way to help animals, and we can anthropomorphise plants and rocks and things to “help” them out too. So, what is it about people that we can intuitively know they need help, and what are these features that we transplant onto non-humans through anthropomorphisms that make us help them out too?

In previous Guidebook entries, and in my published papers, I have touched on the three aspects that make us “persons”, so let’s bring them together in this entry. What makes you “you” and what tugs at your heartstrings when you see someone or something, in need is consciousness, self-hood (or self-awareness) and agency. 

We see someone who is conscious, and we intuitively understand that they can feel pain and suffering. We empathise with them through this shared history of having felt pain and not wishing to feel it again. We relate to those we can see (or project) have a personality, and a sense of individuality, because we know what it is like to be the only one of us. We are unique special snowflakes because we have self-hood and we relate to others with a sense of self. 

Lastly, we understand at an instinctive level the difference between an agent and a patient, a user and a tool, a subject and an object. We are far more likely to anthropomorphise those things that move because we project onto them the aspect of agency and, through that, consciousness and a self. Agency means we are the owner of our actions; we aren’t mind-controlled by something else. We intuitively know that if we help someone else, we are helping that specific person, not a third party. 

These three aspects together form who we are as a person. Take any of them away, and our personhood goes along with it. With these three aspects, you can classify humans as entities with personhood. It’s a dry term, but it is efficient in conveying who and what we are. 

However, as this whole Guidebook’s purpose is about AI, let’s talk about the artificial version of this. If we are entities with personhood, then a conscious, agentic, self-aware AI would be an Artificial Living Entity with PersonHood, or simply ALEPH for short. Notice how intelligence is nowhere to be seen in its name, and that is intentional. ALEPH is not AGI. AGI is about capability and power. If you want to know how much GPT-4 can do and how well it can do it, then you are talking about AGI and how close GPT-4 is to that. If you are talking about personhood and how much of a sense of self GPT-4 has, then you are talking about ALEPH and whether GPT-4 is an ALEPH-type AI.

Many would see this as a distinction without a difference. Plenty of academics in the past, including Turing himself, have presumed that once an artificial entity reaches, and surpasses human intellect, that it would either already have the aspects of personhood or it would gain these through sheer intellect. The Turing Test measures whether an AI can pass successfully as a human by using intelligence, not by showing its agency or self-awareness, or its consciousness. 

By separating ALEPHs from AGIs, we separate the need for ever-growing compute required for ever greater power from the conditions for personhood. We could have superintelligent AI that have no semblance of personhood, or AIs that have equal or lesser intelligence than humans but still have the three aspects of personhood. We don’t need to keep charging in one direction, hoping that we’ll achieve both ends of superintelligence and personhood. It also means that, for those groups hoping to achieve ALEPH through AGI, they can focus their attention more on what will get us to artificial persons, and less about how many billions of parameters across hundreds of thousands of servers their AI has. 

This could also mean that we can focus on ANI, artificial narrow intelligence, more to solve our problems rather than needing one all-powerful AGI to do everything. We can have the generalist ALEPH using human-equivalent intelligence (or less even) with the capacity to learn more, and then we can have our superintelligent AIs as tools to empower us and deliver a greater future for us in specific and niche ways. 

In the next Guidebook entry, I will talk about what achieving ALEPHs will mean for humanity, how it will differ from attaining AGI, and how much safer an ALEPH can be for us. It really does provide the best future for us, from whichever side of the acceleration versus safety divide you sit.