# Hard things are becoming trivial

James Edward Ball / 10 September 2026

I have been thinking a lot about how quickly our idea of what is technically difficult is changing.

The latest frontier models are already very good at interacting with tools, business software, and compute environments. I have also been testing this with Fable and Astro. What stands out is that they are increasingly good at doing things, rather than just talking about them.

A simple example: I had a bunch of 2D floor plans of my flat and wanted to generate a geometrically accurate 3D model of the whole thing, including the rooms, walls, windows, and so on. I gave the PDFs to Astro. Using Blender through MCP and a specialist open-source architectural library, it built the model. I could then manipulate it, generate a render, or visualise it directly in a Three.js web viewer.

A few years ago, I would have treated this as a non-trivial software project. Now it is a task. And I think that is where this is going.

The range of outputs is getting much better, but so is the complexity and expressiveness of those outputs. Models are getting better at reasoning not just linguistically or syntactically, but geometrically, spatially, and through tools. They can move between PDFs, code, 3D environments, specialist software, and web applications without us having to build a dedicated workflow for every step.

I expect that within another couple of generations of models, a lot of the things we thought were non-trivial only a few years ago will have become trivial. This is why I am increasingly bearish on building lots of application-specific software.

A lot of bespoke software that looks useful today could become obsolete very quickly if a general-purpose model can simply use the underlying tools and systems directly. We could spend a lot of money building interfaces and workflows that frontier models can reproduce a year or two later.

---

So I think we need to spend more time thinking about what is actually hard when software becomes easy. For me, a lot of that comes down to bits and atoms.

Analysis and post-processing are becoming easy. Compute is another commodity we buy as an enterprise. Software can increasingly be generated or assembled around a problem. The harder things are acquiring new data and interacting with the physical world.

What datasets do we need that do not exist yet? What data do we depend on but do not control? What physical infrastructure do we need access to? What are the things we cannot afford to be locked out of?

For large organisations, I think this means spending more time on shared data infrastructure, physical capabilities, and the underlying systems that future models will depend on. Because the analysis on top of that data is getting easier very quickly.

The data itself is still hard. The sensor is still hard. Getting a new observation of the physical world is still hard.

This is also why I think connectors matter so much. If the model can interact directly with the systems, tools, data, and compute it needs, then in many cases the bespoke application becomes much less important.

So when I talk about thinking around the fringes, this is broadly what I mean. We need to keep asking what is still genuinely difficult as the capability frontier moves.

The answer will change quickly. If we get it wrong, we will spend a lot of money on systems that become functionally useless before they have delivered much value.