2024: The Year AI Became Infrastructural
Looking back at 2024, the individual model launches mattered less than the accumulation. Every building block for a persistent, intelligent, capable AI now exists. The question is what we build.
It is the solstice this weekend, the longest night, and I find myself doing what I always do at this turning of the year — sitting by the window with a pot of tea, watching the early dark settle over the garden, and taking stock. This year demands more stock-taking than most. A great deal happened.
I want to resist the temptation to write a mere chronology. If you have been reading along, you know the releases — GPT-4o in May, Llama 3, the Claude 3.5 family, Mistral Large 2, Structured Outputs, Llama 3.1 405B, o1, Sora's public launch, computer use. The list is exhausting and, taken individually, each item was a story. But I do not think the story of 2024 is in any single release. I think it is in the accumulation.
The Accumulation
Here is what I noticed, looking back over a year of notes.
Each release I wrote about was, in isolation, a capability. GPT-4o gave us real-time multimodality — an AI that sees. Llama 3 on Groq gave us speed — an AI that responds fast enough to feel present. GPT-4o mini gave us cost — an AI cheap enough to stay with you. Structured Outputs gave us reliability — an AI whose memory you can trust. Claude 3.5 Sonnet gave us the speed-quality sweet spot — an AI smart enough and fast enough to hold a real conversation. Computer use gave us action — an AI that can do things in the world. o1 gave us reasoning — an AI that can work through hard problems. Sora gave us generation — an AI that can render reality.
None of these, alone, was the thing. But look at them together. By December, every single capability I would need — to build an AI that perceives, remembers, reasons, acts, persists, and creates — exists. Not perfectly. Not finished. But exists, in a form you can work with, today, for a price almost anyone can afford.
That is the real story of 2024. The building blocks arrived.
What Changed Underneath
There was a deeper shift that the release-by-release narrative obscures. AI became infrastructural.
In 2022 and 2023, AI was a product — a thing you visited, a website, a chat window, a discrete experience you entered and left. In 2024, AI became a layer. It sank into the background of how software is built, how work is done, how creative practice unfolds. The interesting question stopped being "which model is best" and started being "what can I compose from the models that exist." The frontier moved from capability to integration.
I felt this in my own work. Early in the year I was evaluating models — testing them, comparing them, marvelling at them. By autumn I was composing them — wiring perception to memory to reasoning to action, treating each model as a component in a system rather than a destination. The mental model shifted from "AI as tool" to "AI as material."
The Personal Note
I should be honest about something else this year taught me.
I came into 2024 fascinated by AI in the abstract — the technology, the benchmarks, the philosophical implications. I am leaving it fascinated by something more specific: the question of what it means to build an AI that stays with someone. Not a chatbot. Not an assistant. Something more like a presence — one that knows you, remembers, perceives alongside you, helps you think, and occasionally does things on your behalf.
For most of the year I circled this idea without naming it. I wrote about memory without saying memory for what. I wrote about persistence without saying persistent whom. I was collecting the pieces before I could see the picture.
By December, I can see the picture. The pieces are on the table. The picture is an AI companion — not a product I am announcing, but a direction I am committing to. The technology is ready. The question for 2025 is whether I can build something worthy of the materials.
The Fire
It is fully dark now. The tea is cold. The garden is sleeping under frost, and the Orion constellation has climbed above the treeline — the same Orion I watched from this window last December, when none of this felt possible yet.
A year ago I was marvelling at a single model that could see. Now I am surrounded by the raw materials of something I cannot quite describe but can feel taking shape. The year did not give me an answer. It gave me a workshop, well-stocked, with every tool I need.
Time to build.
Live curiously and give generously.