Slopacolypse

The prediction

Andrej Karpathy's term for a flood of AI-generated, low-quality content he predicted would define 2026 across code repositories, academic preprints, newsletters, and social platforms alike. As agent coding and AI content generation become cheap and easy, the volume of output across every one of those surfaces rises sharply, but quality does not rise with it. AI makes production easy without improving taste, rigor, or selection, and the result is a flood of plausible-looking but shallow or wrong material that gets progressively harder to filter for whatever is actually good.

Geoffrey Woo's corollary for investors

Geoffrey Woo draws the implication forward into a specific prediction about outcomes rather than just content quality. As AI drags the median down toward slop, the resulting distribution does not flatten, it bends into a steeper power law: the floor that counts as passable content rises for everyone at the same time the ceiling that counts as genuinely original work becomes scarcer and more valuable, so the premium on the non-slop tail rises together with the volume of slop beneath it.¹ He operationalizes the belief as a personal filter rather than leaving it as commentary. Anything that reads as AI-generated boilerplate gets discarded the moment he recognizes it, and his stated instruction to anyone trying to reach him is not to send him AI slop, but to write him something that is not already in a model's training set.²

Why the corollary matters more than the prediction

The prediction alone is a claim about content volume that anyone publishing online can verify for themselves. Woo's corollary is the more useful half, because it converts a complaint about the internet's quality into an investment and hiring thesis: in a world where production is nearly free, the scarce and rewarded skill is judgment, curation, and whatever cannot be produced by prompting a model with everyone else's prompt. A filter that discards anything with the shape of AI boilerplate is a crude instrument, and it will misfire against genuine work that happens to resemble the pattern it is trained to reject, but it is also a coherent bet that as the baseline rises, distinguishing signal from the flood becomes the actual competitive advantage rather than a secondary concern.

What remains open

Whether the flood is a permanent equilibrium or a temporary transition is unresolved. Markets for quality content could plausibly self-correct as filtering tools improve on the other side of the same AI wave that created the flood, in which case the steeper power law Woo describes would compress again once better curation tools become as cheap as the generation tools that created the problem.