Karpathy's Scale and Solving Horseless Carriages (www.kush.pw)

Pete Koomen’s critique—exemplified by Gmail’s assistant taking longer than manual drafting—frames a clear design rule: the only metric that should drive AI-native products is TTC (time to completion). Drawing on Karpathy’s “scale” from manual tools to full agentic autonomy, the goal for apps is to move rightward along that axis so human time approaches zero. That reframes success as minimizing user effort, not feature count, and demands redesigning software around automation and agency rather than tacking on AI bells and whistles. Technically this means building vertical, hyper-personalized agents that connect to a user’s real-world tools and workflows (connectors to legacy systems, context engineering, and model-driven orchestration), automating recurring decisions even if verification gates remain. The Gary-the-subcontractor example shows how classifiers, context retrieval, and workflow execution can collapse TTC for job triage, quoting, billing, and responses. The trade-off: narrower audiences but much greater utility and defensibility; generic platforms risk commoditization as model tooling lowers build cost. Practically, startups should prioritize deep user modeling, reliable action APIs, and iteration on agent reliability—because as models, context management, and orchestration improve, the winners will be those who move fastest along Karpathy’s scale.