GitHub - chuanqisun/spaceship: Steer your ideas in realtime

Open original

Spaceship is an open-source experimental interface where you steer an LLM's text generation live, like flying a ship, with knobs for attention, vector steering, and scrubbing through alternate futures. It's a prototype for how interaction might feel once models can generate at 60 FPS, built by Chuanqi Sun as groundwork for his MIT Media Lab thesis.

  • Four interaction prototypes: a focus monitor showing what the model attends to, attention lock to bias future output toward chosen phrases, vector steering that decomposes a task into eigenvectors you tune and recombine, and realtime regeneration you can scrub through.
  • Making it feel realtime required aggressive speculative generation and careful cache invalidation; for example, moving the main knob backward empties future text and cancels in-flight requests. The author says Claude Code couldn't one-shot these judgment calls.
  • It runs in the browser with Cerebras or Gemini API keys stored in local storage, and community projects like a VSCode extension and a local-model browser version already build on it.

Spaceship is a GitHub project by Chuanqi Sun, a researcher at the MIT Media Lab, that explores what interaction could look like once LLMs and decision models run fast enough for 60 FPS responsiveness. The core bet is that when text generation gets cheap and quick, a new set of interfaces becomes possible, and the project prototypes four of them: a focus monitor that visualizes what the model is attending to and uses that to point the human's attention; attention lock, which forces the model to weight specific phrases more heavily to bias its later output; future vector steering, which decomposes a task into eigenvectors, tunes each independently, and recomposes them to shape incoming text; and realtime re-generation, where scrubbing a control lets the user explore alternative continuations. A live demo runs in the browser at chuanqisun.github.io/spaceship, and users can plug in their own Cerebras and Gemini API keys, which stay in local storage. The engineering detail is where the article gets most concrete. Achieving the realtime feel requires aggressive speculative generation and careful cache invalidation tied to user input. The author gives a clear example: when the main knob moves backward, the system has to discard the future text it already produced and cancel in-flight requests. Sun says Claude Code wrote most of the implementation, but the UI and UX decisions were hands-on, and the cache and cancellation logic needed human judgment because the AI did not get it right in one pass. The design process used what the author calls interactive style swatches, where the model produces diverse design options and the author iterates on feedback. Design artifacts are linked from an extra page in the repo. Beyond the demo, the project is framed as groundwork for a thesis on coupling semantic and symbolic systems to strengthen human creativity, agency, and enjoyment of craft. The repo also invites community builds, and two are already listed: Spaceship Writer, a VS Code extension by brentlee, and Spaceship Lime, which steers local models in the browser, by Philippe. The article includes a BibTeX citation for anyone who builds on the work.

Ask about this article