What's this about?
Ethan Mollick's running theme: today's coding agents (Codex, Claude Code) can generate genuinely novel, playable game demos from a single prompt — he showed a Piranesi-style city-builder of arches and waterways in cyclopean ruins, and a game "about Imminence." His plea to the field: stop cloning the same six example games; the tools can do strange, original things now.
What exactly is "one-shot generation"?
One-shot means the AI produces the whole working artifact — design, code, art direction — from a single prompt, no iteration. Mollick keeps re-running prompts he first tried a year ago as an informal benchmark; where they once failed, they now yield playable, aesthetically distinct games.
Why does it matter?
He ties it back to the 2023 "Sparks of AGI" paper's TikZ unicorn — the early hint that language models had latent visual/spatial creativity. The demos are qualitative evidence that the one-prompt ceiling keeps rising even as the benchmark-skepticism debate rages.