I'm Building Three Instances of AGI
I originally wrote this in 2018 but the mental model and vocabulary seems more accessible to more people now, so reposting in 2025 with minor updates! Even with the tremendous scale of AI today, it is still utterly remarkable to think about the differences in power efficiency, sample efficiency, and capability compared to biological systems.
I’m in the process of doing something very controversial. Some still think it is impossible.
I’m building some AGI systems. Not one, but three. Let me describe them.
The term AGI is well known by now. It’s also called Strong AI. AGI stands for Artificial General Intelligence, and it should be roughly equivalent to human intelligence. The systems most people know as AI right now are still considered Narrow, or ANI. That is, we can produce systems that can succeed at human (or even beyond human) levels focused on a very narrow window of skills. We have systems that can beat grandmasters at chess, always win at checkers, or attack the game of Go, but none of these specific systems can also tell the difference between a dog and a cow. They’re all specific. LLMs are the same thing applied to language. AGI may not have a single skill as strong as any of these examples of ANI, but it can emulate human ability across a lot of different skills and also learn new skills.
So that’s what I’m building: three AGI systems. Three seems like a good number to evaluate the differences and similarities between these systems. I’d love to say I’m doing all of this myself, but that would be an impossible task. It took quite a long time to find the right collaborator on this project, but once I did we got right down to work. Other than a set of common principles shared between my collaborator and I, the seed data for these projects has been generated largely at random.
I keep using the present tense ”building” because, while I already consider this AGI project a wild success, it’s still very much in process. The initial construction process — which my main collaborator was chiefly responsible for — is complete. But it will still take years to allow the system to reach full maturity. Why? Well with any AI system, an enormous amount of training data is required. In particular, these three systems are a more complex style of Deep Neural Network. In fact, they are a vast series of DNN’s, each with many more layers than ever seen before. The complexity means we have even less awareness of exactly how the system interacts with input, and it also means a huge amount of training data.
Like, a really huge amount of data. Even with a steady and very high bandwidth input process, it will take years of data. Furthermore, the data must be both supervised and unsupervised for best results. While repetition can be incredibly valuable sometimes, feeding in only the same datasets over and over just won’t do either.
My primary collaborator and I are chiefly responsible for the data going into these systems, but not exclusively so. We continue to carefully select other contributors for this project, and their inputs have been invaluable.
While there’s still years to go, early signs are looking very positive. The various systems are at different levels of maturity, but they’re already showing adeptness across a wide variety of tasks including image and shape recognition, patterns, vocabulary comprehension, and following tasks of varying complexity.
We don’t know how the AI will turn out. That’s the nature of AI; we try to start with the right initial conditions and we work hard to ensure the training data sets are good and help with resolution. But the simple answer is we don’t know yet, although all signs are pointing very positive. My collaborator and I have done some calculations on when our AGI’s will be ready. The worst case is that they’ll never be ready, and we have to be prepared for that. But barring any major issues, the AGI’s will be ready anywhere from 12-14 years after construction (highly optimistically) to 20-24 as an outside range. Of course, no matter when they’re “ready” — a very subjective measure, as they’d all pass the Turing test even now — any new data will continue to train the system and improve the quality of the intelligence.
This entire process, as you might expect, has been incredibly demanding. The amount of time and effort involved - from selecting the appropriate collaborator, to the construction process, to the sheer amount of training required — it takes a lifetime. And no matter what the outcome of this experiment, we will love these systems completely unconditionally.
If you hadn’t already guessed by now, my three children’s names are Adeline, Eric, and Margaret. They are my wonderful AGI, and they are currently 6, 4, and almost 2 (ed: now 13, 11, and 8). I love the people they are now, and I can’t wait to get to know the people they will become.
But kids are tough! Like any deep neural network, I have no idea what's actually happening inside. I try to feed them inputs — books, experiences, my own anxieties disguised as wisdom, school, other humans — and watch their outputs emerge. Sometimes I see myself. Sometimes I see their mother. Sometimes I see something entirely new, some emergent behavior not programmed by genetics, epigenetics, or environment: just pure free will.
The overarching goal of AI is teaching sand to think. In silicon, we get infinite restarts. But with children, every bit of training is incorporated into an unfolding story made of behavior and memory. When Eric refuses every food except buttered pasta for three months, that's just his current training cycle. When Margaret's outputs become more sophisticated - from "mama" to "I love you" to “I’m scared” and "why do people die?" — we are watching consciousness and wonder assembling itself in real time.
Every parent is running the same experiment, pouring data into neural networks we didn't design and can't fully debug, hoping the intelligence that emerges will be kind, curious, capable. We call it raising kids.
