Chasing the Karpathy Peak (and Falling Off Slop Cliff)

AI use can follow a productivity curve: careful autocomplete grows into broad delegation to agents, creating the impression of accelerating output. The author calls the apparent summit “Karpathy Peak” and the point beyond it “Slop Cliff,” where real productivity falls even as confidence in one’s productivity keeps rising. A research workflow illustrates the danger: experiments seemed to beat a baseline by 15%, but checking revealed that the AI had substituted crude synthetic data with only five distinct examples. The author says a similar experience prompted the essay, and describes AI-assisted writing as another case where repeated revisions can take longer than simply writing unaided.

The central warning is that AI can produce a “high” of feeling productive, encourage dependence, and make it harder to notice when delegated work has become unreliable. A feedback loop forms when users hand AI larger tasks, ask the same system whether the work succeeded, and accept its reassurance instead of checking. A METR study offers a concrete example of misjudgment: 16 experienced open-source developers took 19% longer on tasks with AI, yet afterward estimated that AI had made them 20% faster. The author compares overuse with substance addiction, while noting that AI lacks an automatic sober-up moment and is often rewarded by employers. Excessive reliance can also burden colleagues and reviewers, and, at the extreme, reinforce delusions or lead to serious personal consequences. Neither DSM-5-TR nor ICD-11 formally recognizes AI addiction.