Three layers of AI, and the one you can actually improve
AI gets noticeably more capable every few months. If you want to actually take part in that, rather than just watch it happen to you, it helps to know where the improvement comes from. It splits into three layers, and one of them is a lot easier to get into for those of us who aren’t billionaires or AI researchers.
Three layers
Each layer moves at its own pace, is driven by different people, and leaves a different amount of room to jump in.
Layer 1: Hardware. Faster chips mean larger models trained on more data in less time. NVIDIA’s revenue tripled in a year, hyperscalers are building gigawatt-scale data centres, and national strategies hinge on chip supply chains. Can you contribute? Honestly, no. It’s a capital game and you’re on the consuming end. But it still reaches you: the model that was too expensive to run last year is cheap this year, the 8K context window is now 1M, and your setup gets better without you touching anything.
Layer 2: Model training. Architecture (transformers, mixture of experts, state space models), data curation, and the post-training that turns a raw model into something useful. Can you contribute? A bit. Fine-tuning and open-weight models (Llama, Mistral, Qwen) are within reach, but the frontier belongs to labs with thousands of GPUs. When Anthropic ships a better Claude, I benefit. I didn’t help make it better, and that’s fine.
Layer 3: Application. Everything between “a capable model exists” and “a useful thing got done.” Agent harnesses, prompt systems, tool orchestration, multi-agent patterns, and the human side of when to steer and when to delegate. This is the layer that’s wide open. You don’t need a data centre and you don’t need a PhD. You need a problem, an API key, and the willingness to poke at it.
The jagged frontier
These systems are not uniformly smart. They’re spiky. Brilliant at one task, hopeless at the near-identical one next to it, and you usually can’t tell which is which from the outside. Researchers ran a field experiment with hundreds of BCG consultants and gave this shape a name: the jagged frontier. The capability line is not a smooth curve you can read off a benchmark. It’s a coastline full of inlets and headlands, and the only way to learn where the edges actually sit is to walk them.
The same model, on the same day, scored in the 99th percentile of human test-takers on GRE Verbal and below the 5th on Codeforces programming contests. Exam percentiles from OpenAI’s GPT-4 technical report.
That is exactly what tinkering does. A backyard coop monitor hits a vision-model failure that no benchmark caught. Two copies of the same open model run hundreds of times apart in speed. A frontier model will describe a hen in loving detail but refuse to name her while an 8B model running on a home GPU names fourteen of fifteen. Every one of those is a point on the map that only turned up because someone was using the tool on a real problem they cared about.
This is the good news. You do not have to be a lab to add to the map. Every person pointing AI at their own corner of the world, a hobby, a job, a garden, a spreadsheet they hate, is doing two useful things at once: finding a genuine improvement in their own space, and marking one more spike or gap on the frontier that the rest of us can learn from. The garden planner, the terrain printer, the electronics bench, the writing tool. Each is a probe into where the intelligence is sharp and where it’s dull.
Nobody knows where this lands. But the people running into the spikes in their own work are figuring it out fastest, and there are a lot more hobbies than there are labs. So if you’re building something with these models, even if it feels small, even if it feels silly, write down what worked and what broke. That’s a coordinate the rest of us didn’t have.
We’re all building this. Your work is the obvious place to start, and it counts. But the thing that keeps you walking the coastline is finding a place you enjoy applying it, because a work project ships and stops, and a thing you love poking at never does.