Published thinking.

Andrej Karpathy coined "jagged intelligence" for a strange property of today's models: the same system that can pass a radiology board exam cannot reliably tell you that 9.9 is bigger than 9.11. Human competence is round — strength in one area predicts strength in the ones beside it. Machine competence is jagged: superhuman peaks and absurd valleys sitting side by side in the same tool.

That mismatch is exactly what fools clinicians. We calibrate trust through credentials, so a model that clears the licensing exam earns the same deference — yet passing the exam says nothing about whether it is safe on your worklist. Dr Tham's reframing is to stop asking "how smart is this AI?" and start asking "where is this AI smart?" — then test it on your own cases, not the vendor's benchmark.

"The same AI that can pass a radiology board exam cannot reliably tell you that 9.9 is bigger than 9.11."

Singapore's updated AI in Healthcare Guidelines draw a line many frameworks blur: they define three distinct stakeholder groups across the AI lifecycle — Developers, Deployers, and Users — and attach specific obligations to each.

That distinction matters more than it sounds. When an AI tool underperforms in a clinical setting, the first question is rarely technical — it is "whose job was this?" A framework that forces ownership to be named before deployment, not litigated after, addresses the failure mode Dr Tham sees most often in practice.

"Most deployment friction comes from exactly this: unclear ownership."

Budget 2026 signalled Singapore's intent to formalise AI governance at a national level. The instinct is sound — but governance that is designed top-down tends to discover its blind spots in production.

This piece looks at what inclusive governance actually requires from the ground up: practitioner voices in the room where rules are made, professional bodies like RADII translating between regulators and the clinicians who live with the consequences, and enough room left for the startups building the tools to keep moving.

After 16 years in private radiology, three models come into focus: restructured hospitals, small independent centres, and large private groups. Each trades something for scale — autonomy, reach, or resilience — and each demands something different from the radiologists inside it.

The thread that runs through all three: as automation takes over more of the workflow, the judgment of the person in the seat becomes more consequential, not less.

"The more we automate, the more crucial the driver in the seat becomes."