chuanqisun/spaceship
AI EngineeringAn experimental browser IDE for steering LLM text generation in near real time with attention and future-vector controls.
What it is
Spaceship IDE is a conceptual prototype for interacting with a near-realtime language model. It explores whether an LLM can participate in a fast feedback loop for shaping ideas, rather than returning a single completed answer. The interface presents four experiments: a focus monitor that visualizes model attention, an attention lock that biases generation toward chosen phrases, future vector steering that tunes decomposed task components independently, and realtime regeneration that lets you scrub through alternative continuations.
It is aimed at people interested in LLM interaction design, creative tooling, and human-AI collaboration, rather than at production application development. The project comes from the author's work toward an MIT Media Lab thesis on coupling semantic and symbolic systems for creativity and agency (project background).
Try it
Open the hosted simulator at chuanqisun.github.io/spaceship. The simulator can run without keys for exploration; to go live, supply Cerebras and Gemini API keys. The keys are stored in the browser's local storage, and the README explicitly warns that live use is at your own risk. The same page links to the source repository for local inspection (repository).
For a quick visual introduction before opening the app, the repository includes a screencast (demo video) and clips of each experiment: focus monitor, attention lock, future steering, and realtime regeneration. There are also an Instagram reel and a YouTube walkthrough.
Requirements and limits
Live operation depends on external Cerebras and Gemini API access and keys; the simulator alone does not produce live model output. The README does not document a build or local development workflow, so the hosted demo is the supported entry point described here. Realtime behavior relies on aggressive speculative generation and cache invalidation, which means the experience is tied to the latency and behavior of the chosen providers. This is a demonstration of possible interactions, not a general-purpose IDE or a stable API for building products.
The project is MIT-licensed. API keys remain client-side, but sending requests to third-party providers is inherent to live use.
HotRepo's take
The project appeared in HotRepo's new-project ranking on 2026-11-16 with 104 stars gained in a window of roughly five and a half hours, alongside a Show HN thread titled "Near-Realtime LLM Steering," which recorded 8 points and 2 comments. The discussion frames the work as a conceptual prototype motivated by the rise of fast decision models and specialized inference hardware; the author's comment asks whether LLMs can sit inside a realtime feedback loop for shaping ideas (HN discussion). That framing matches the repository's own four experiments and makes the project interesting as an interaction-design probe rather than as a ready-to-use tool.
Written by AI from the project's README, updated . The project's own docs are the source of truth.