A new class of throwaway software just fell into coffee-break pricing: hyper-custom, on-demand interactive worlds generated from a paragraph of text, the kind of work no human would sit down to do.
Karpathy's Lord of the Rings build is the cleanest case. He gave Anthropic's Opus 5 the opening of Tolkien, a roughly one-million-token budget, and about two hours. Opus 5 returned about 5,500 lines of three.js code that procedurally placed polygon assets in three-dimensional space and animated them in a browser. His reaction, "Sure, why not, it's ~free," names the cost-and-stamina asymmetry now defining what large language models can produce on demand.
The naive read is release news. The mechanism is sharper. A frontier model has effectively infinite patience and roughly free marginal cost; a human does not. That gap unlocks the "ephemeral GTA of X" Karpathy coins: the spectator-mode NPC walkthrough, the character-POV variant, the procedurally rendered cutscene for any scene a user names. The same week, a Hacker News practitioner built a generalized DeLorean hero animation for any film on the same pattern, calling the moment the start of "throwaway software."
The ceiling is just as durable. Karpathy flags that Opus 5 could not watch its own output: it took slow screenshots, audited them, and patched its own jank by hand. The model that wrote a 5,500-line world still cannot natively see or play it. Stamina is cheap. Perception of the model's own generation is the next bottleneck, and every "built from a paragraph" demo will hit the same split until that moves.