Most software still worth building before artificial general intelligence arrives will, I think, be what I call a domain-specific harness (DSH). That is a specialized application that connects a model to one domain’s tools, data, workflows, permissions, and users. If AGI becomes superintelligence, even those harnesses stop being built as separate things. The system itself would create, change, and run whatever interfaces, integrations, controls, and security boundaries a task needs. Those layers would still exist in effect. But nobody would design and code them the way we write software today.
Until then, the tools I want to write are small and very easy to maintain. They help both people and models. And they help us reach a responsible, aligned AGI sooner rather than later.
The division of labor I am aiming for is simple. Human intuition steers the exploration. The AI, and soon the AGI, reasons through the data, writes the analysis, codes, and designs the experiments.
Everything here has a sell-by date. These claims describe what models could do when I wrote them. Their ability has been growing faster than my notes during the current AI boom. Treat each claim as provisional. If AGI arrives, and more so if it becomes superintelligence, most of this becomes out of date. Even if it never arrives, the details will still age quickly.
Current as of September 2026.