Understanding AI more deeply means I can trash it harder

Vibe coding at scale so far is often trash. It's potentially full of security holes and unnecessary changes and features, saving time on code generation but costing time in code review. It seems no one wants to be rude enough to point out that maintaining code is the bulk of TCO (Total Cost of Ownership) historically. It's wild that vibe coding platforms can sometimes delete entire databases or hard drives because they've gone off the rails.

I've recently arrived at a strategy of dual-wielding critical realism (if not transcendental realism) and post-structuralism in order to be able to take down naive notions of AI and what it's capable of. Like, we fit AI to training corpora, not to truth per se. It would be irrational to claim that AI trained in that way could arrive at absolute truth, as the training process is grounded in data at a significant remove from reality itself. And the decision of dataset curation and tagging reflects power as per post-structuralism, so there can be no appeal to some mythical pan-cultural objectivity across all domains. I suppose if I had to typify this approach to critique, it would amount to: AI isn't what you think it is.