The black box is becoming a thinner moat

Metrics are being solved faster than I expected.

OpenAI says GPT-6 Astra solved 88% of SRE-Bench tasks in one attempt and 99.2% within four attempts. SRE-Bench is a more serious test than the usual coding benchmark. It contains 19 privately built programs, 262 contamination-free binary instances, and 1,572 deterministically graded reverse-engineering tasks. The agents receive binaries without source code and have to work out what the software does.

That matters because the model cannot simply recognise a public repository from its training data. It starts with an opaque artifact and has to recover the structure underneath.

The same pressure is appearing in software engineering. Models are getting better at turning descriptions into working applications, then testing and revising them. Once that becomes cheap and reliable, software itself becomes a weaker source of advantage. The interesting question is no longer only who can build the thing. It is who can build the version that fits a particular person, team or business.

I expect more software to become personal in this sense. A small firm might have its own tools, workflows and agents, built around the way it actually operates. A family might have its own software for planning, learning and managing daily life. The code will still matter, but the advantage may sit in the context around it: the data, the habits and the quality of the adaptation.

I keep coming back to Elon Musk’s argument that hardware might be the enduring layer. Software can increasingly be generated on demand. Physical production still depends on materials, machines, tolerances and logistics.

Even that advantage may weaken.

Mass customisation, 3D printing, body scanning and automated production are already making products more personal. There is a difference between choosing from a fixed menu of options and having a system generate a design from your own measurements, preferences or requirements.

The interesting question is whether manufacturing can become as programmable as software. A person might describe a product, have software design it, and send the result to a production system that makes a single copy.

That is still a long way off. Manufacturing has to deal with safety standards, supply chains and physical failure. A program that is slightly wrong may waste an afternoon. A product that is slightly wrong may injure someone.

But the economic pressure is obvious. As software becomes easier to produce, more attention will move towards the physical constraints that software cannot simply wish away. The next important contest may be between systems that personalise software and systems that can turn those personalised designs into physical objects.