Horseless Carriages
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
A concept introduced in a widely-circulated essay critiquing a prevalent but misguided pattern in early AI software development: treating AI as a feature bolted onto conventional software architecture, rather than rethinking the architecture entirely around AI capabilities.
Core Critique
The "Horseless Carriages" metaphor — named after early automobiles that superficially resembled horse-drawn carriages despite the fundamentally different technology underneath — captures a failure mode observed across many early AI products. As described in the Inside YC's AI Playbook discussion, the pattern involves "companies building software and adding AI features by sort of slotting a little bit of AI inside of a lot of software." source This framing directly criticizes the AI as Feature vs AI as OS distinction — i.e., companies defaulting to AI-as-feature rather than reconceiving their product around AI as a foundational layer.
The Preferred Alternative
The essay's implicit positive vision is an inversion of the criticized architecture. Rather than deterministic software wrapping an AI component, well-designed AI-native software should look like "an agent wrapping software deterministic tools rather than deterministic software wrapping an AI." source The agent — and thus the model — sits at the center, with traditional software components serving as tools it can invoke.
Relatedly, the best AI software products are characterized as being very small codebases: they "add kind of the smallest amount of code ahead of time that you need in order to let the model shine," rather than surrounding the model in heavy scaffolding. source
Points of Disagreement
No direct rebuttals to the "Horseless Carriages" thesis appear in the available material. The concept is presented as a critique that gained significant traction (the essay "went pretty viral"), and the contrasting positions in the graph reinforce rather than dispute its core argument — both the agent-wrapping-tools framing and the minimal-code principle serve as positive articulations of what avoiding the horseless-carriage trap looks like in practice.