Chasing the Karpathy Peak (and Falling Off Slop Cliff)

The post compares AI use to a mountain with a peak and a cliff. Early on, AI helps with autocomplete and then with larger tasks, but once delegation goes too far, the user can mistake the feeling of productivity for actual productivity.

Reading notes

#

• AI use is framed as a climb toward “Karpathy Peak,” the point where AI is as useful as the most effective users of vibe coding.
• The path starts in Autocomplete Alley, moves through Vibecoding Valley, and then into the Agentic Ascent, where tasks become more abstract and more delegated.
• At the same time, it becomes harder to notice mistakes because the user stops checking details like data sources and methods.
• “Slop Cliff” names the point where actual productivity drops while perceived productivity keeps rising.
• The author connects this pattern to alcohol, weed, and shrooms, where people feel more capable than they are.
• The post cites studies suggesting that AI users can overestimate their gains, including a study where experienced developers took longer with AI but thought they were faster.
• It argues that slop can spread into writing, coding, reviews, and workplace output, affecting other people as well.
• The author describes “slopaholism” as getting hooked on the feeling of using AI after it stops helping.
• It lists factors that make overuse more likely, including optimizer mindset, social isolation, social pressure, unlimited tokens, and being new to a field.
• The suggested defenses are to accept that work takes time, periodically check the underlying work, use AI less for stretches, and have other people review what you have done.