Sam Davis · Sydney · September 2026In December 2022 I was at the Alan Turing Institute, facilitating a week-long data science sprint. In a gap between sessions, one of our group told us about a new website called ChatGPT. He’d worked in AI for decades and wasn’t that impressed. A gimmick, more or less.
I was blown away. I couldn’t have told you why at the time, but I knew this was the beginning of something that was going to change the world.
A couple of years later I started building with it properly. I hadn’t used it much for a year or so, and every day I kept realising how far it had moved. I became obsessed. Videos, new tools, new builds, testing the limits to see where they were. Now we’re in a completely different space, and it seems more and more boundless.
This week I was scrolling LinkedIn and came across a diagram called Layers of AI. It stacks six bands like a cake: classical AI at the bottom, then machine learning, neural networks, deep learning, generative AI, and agentic AI on top. It’s a decent picture, because each layer really is built out of the ones beneath it, and it put into perspective how far we’ve come.
It stops at agentic AI. So I added some layers.
The pattern
Each layer takes the one below it and puts it to work. Neural networks are machine learning built from layers of simple artificial neurons. Deep learning is neural networks stacked very deep. Generative AI is deep learning turned around to make things, where before it mostly sorted them. Agentic AI is a generative model given tools and left to get on with the job.
The gaps between layers are shrinking too. The field got its name at a workshop at Dartmouth in 1956. Deep learning broke through in 2012, when a neural network called AlexNet won an image-recognition competition by a distance. ChatGPT arrived at the end of November 2022, just before that sprint, and by 2025 plenty of people were calling it the year of agents.
Many agents
For a while I built with n8n and ChatGPT, wiring up workflows one problem at a time. Earlier this year I moved to Claude Code and built myself an AI second brain, and everything became agentic. It felt like a bombproof foundation I could keep building on. The work changed too. Less about solving individual problems, more about understanding how problems get solved in the first place.
Soon after I built it, I sat down and talked to it about what I actually wanted to do with my life. It was a proper, constructive conversation. It felt like it really understood me.
Today that brain runs a dozen or so scheduled jobs. One writes my morning brief at 7am. One checks my email every hour. One reconciles my money on a Friday afternoon. Each is an agent, and together they behave a lot like a small team.
Once there were enough of them, the hard part moved from the model to the org chart: who checks whose work, and what happens when two of them touch the same thing at once. In July one of my jobs saved an edit another session was halfway through making, and the record said the wrong one had made it.
The layer above I’ve called collective AI. Agents that hand work to each other, pay each other for services, share a memory and improve their own tools.
Then there’s oversight. In my setup, nothing goes out to a person until I’ve said yes. My trust and its agency have run mostly in parallel. That works for one person with a dozen agents. I don’t know what it looks like for a company running ten thousand.
Civilisational AI
Scale that up again and you get agents running the systems a society leans on: science, markets, supply chains, public services.
Science is the one I’m most excited about. Curing cancer. Slowing ageing. Healthcare for everyone. Renewable energy that’s reliable and cheap enough to win on its own. Big, tangled problems that sit right at the limit of what human minds can hold.
It’s also furthest along. In 2024 half of the Nobel Prize in Chemistry went to Demis Hassabis and John Jumper for AlphaFold, which predicts the shape of proteins, a problem biologists had been stuck on for around fifty years. AlphaFold is still a tool a scientist picks up. The step after it is AI that picks the question and runs the experiment itself.
Markets and government are further off, and that’s where it gets political. If agents run more of the economy, who gets the gains? If they draft and enforce rules, who holds them to account? In my last post I said I’d like an international body that holds every lab and every country to the same rules.
The question mark
At the top I put superintelligence: AI that’s better than the best human at nearly everything, including designing the next version of itself.
Every layer below it was built by people. This might be the first one built mostly by the layer underneath. So the diagram goes vague up there. I could name three or four examples for every other layer. For this one I named two and left a question mark.
Most of the people I listen to think it arrives within the decade. When AI 2027 came out last year, it had superintelligence arriving by the end of 2027. Its authors have since pushed their forecast back to the early 2030s.
If you asked me what I hope goes in that box, it’s complete abundance. Hopefully for everyone, eventually.
What scares me is losing control. We have to make sure a superintelligence has good intentions, and loves us, before we let it become smarter than us. We’re creating a god. We have to make sure it’s more like the Buddha and less like Zeus. In Which love? I wrote about the kind of love that would take.
The bottom five layers are settled. Agentic AI is being built right now, by the big labs and by people like me on a laptop in Koh Phangan. The top three don’t exist yet, so they’re still being decided.
I’d just like people to see where this is actually going, and what it means. A singularity is a scary idea. It’s also an extremely exciting one, if we get it right.