AI and Structure

AI: capability vs structure

I read an interesting article today by Pratik Karki, a young Nepali entrepreneur on the international AI stage. He is the founder of Anthromind.

ai vs structure

He starts with a prediction of Geoffrey Hinton who said a decade ago that radiologists will be replaced. But today, he says, radiologists are more in demand than ever. Was Hinton wrong? Yes and No. Yes, he was right about technology in that technology can do the job. And No, because regulations require that someone accountable needs to sign-off on the image interpretation, which creates a human expertise bottleneck.

ai vs expertise

This brings us to our second point on the capabilities of AI on knowledge elicitation. Pratik puts it eloquently:

“Scientific computing is brutally hard for a human and comparatively easy to capture, because you can check whether an answer is right. Legal interpretation is easy to state and hard to capture, because there is no answer key. The real split is whether a verifiable ground truth exists.”

The takeaway: structure and capability

There are two factors that will affect the way we use AI: structure and capability.

hyper-scaling: storage and tools

Another interesting article I just read talks about how higher education has been duped by corporate computing systems into a dependency of sorts that impacts “knowledge production practices and scholarly infrastructure.”

Two cases are highlighted. First is the Google cloud computing services (including Google workspace apps, etc.) that started with unlimited strorage but has since capped storage and introduced tiered service levels. Second is the generative AI tool included with Microsoft 365, better known as Copilot, which initially started out as an optional enhancement, but without an easy way to disable if at all.

Although I admit I am not totally against AI usage, it is undeniable that as a researcher in an R1 academic institution one does feel the weight of corporate greed.