Blog / Generative AI's New Clothes

Generative AI's New Clothes

I enjoy playing with generative AI (ChatGPT, Stable Diffusion, etc.) as much as anyone, but the shock and awe coming from ordinarily sober commentators needs to be tempered a bit.

For those who have been following the progress of generative AI, it is unsurprising that a bot trained on vast amounts of source material can now produce a plausibly vague but grammatical essay — on, say, supply chain optimization. I suspect it would be difficult to differentiate that essay from the dozens of other vague articles on supply chain optimization that I've read over the last twenty years.

I would also be unsurprised if a generative AI could assemble a perfectly reasonable-appearing procurement software website, given that I can't tell most of those apart either.

But now I'm reading claims that generative AI will supplant programmers. Those claims can be dismissed from first principles, because "a sufficiently detailed specification is indistinguishable from its implementation." In other words, the work required to describe the desired system to an AI is, in fact, the "programming." That's why workflow isn't as easy as it's made out to be — the vendor may natter about "low code," but it's still programming.

And, I've seen the notion floated that generative AI will produce advanced sourcing plans or provide other intelligent decision support to procurement staff. I'm not sure how to connect the technical dots between that claim and what generative AI actually is, but in principle I see no reason why an AI couldn't generate the same generic AP-based pablum as existing systems. However, detailed sourcing plans are far too situation-specific for vague suggestions to be useful.

We build domain-specific ML models at Spendata, and I can report that in order to achieve better than mediocre results there is more art than science involved, much of it on the data prep side. Try training and testing an ML model yourself on real — not demo — data, and you'll quickly realize how very far the current state of the art is from supporting the wild speculations currently being bandied about.

Questions? Contact us

Please do not enter anything into this field
Privacy Promise

Related Posts

Spendata's Auto-Mapping