MIT's Spaceship IDE Lets Developers Steer AI Text Like a Live Control Surface Chuanqi Sun released Spaceship IDE, an MIT-licensed open-source browser experiment that lets developers steer a language model's output in realtime through controls including focus monitoring, attention locking, future vector steering and scrubbing to regenerate alternative continuations. The live demo runs entirely in the browser using user-supplied Cerebras and Gemini API keys stored locally, and steering happens via prompt and attention biasing through APIs rather than hidden-state activation edits. Sun frames the project as an early step toward his planned MIT Media Lab thesis on coupling semantic and symbolic systems, premised on language and decision models eventually responding at roughly 60 frames per second. MIT's Spaceship IDE Lets Developers Steer AI Text Like a Live Control Surface An open-source browser experiment called Spaceship IDE lets you steer an LLM's next sentences with knobs, in realtime, like piloting a ship. - Chuanqi Sun released Spaceship IDE https://github.com/chuanqisun/spaceship?utm source=alphasignal , an open-source browser experiment for steering LLM generation in realtime. - Interactions include focus monitoring, attention locking, future vector steering, and scrubbing to regenerate alternative continuations. - Live demo https://chuanqisun.github.io/spaceship/?utm source=alphasignal uses user-supplied Cerebras and Gemini API keys stored locally in the browser. - Steering happens via prompt and attention biasing through APIs, not hidden-state activation edits. - Realtime feel depends on speculative generation plus careful cache invalidation when the user scrubs backward. - Framed as a step toward Sun's upcoming MIT Media Lab thesis on coupling semantic and symbolic systems. Spaceship IDE https://github.com/chuanqisun/spaceship?utm source=alphasignal , a browser prototype from Chuanqi Sun, drew attention after a post about it circulated on X. The MIT-licensed project lets users steer text while a language model generates it, with controls for emphasizing phrases, adjusting task dimensions, rewinding output, and exploring alternate continuations. For developers, it offers a concrete example of an LLM interface built around continuous input rather than discrete chat turns. Sun links the project to his planned MIT Media Lab thesis work https://code.chuanqisun.com/master-thesis/?utm source=alphasignal . Its premise is that language and decision models will eventually respond at roughly 60 frames per second, allowing interfaces to update quickly enough to feel continuous. The repository presents Spaceship as an early interaction experiment, with the code and concepts open for adaptation. Controls for a live token stream The live demo https://chuanqisun.github.io/spaceship/?utm source=alphasignal runs entirely in the browser. Visitors can use a simulator or enter their own Cerebras and Gemini API keys for live generation. Cerebras supplies the low-latency inference that allows text to keep moving while the controls change. The cockpit keeps generation active as users manipulate the next continuation. A timeline control can rewind generated text, while other controls alter the context used for subsequent requests. The README groups the interface around four primitives: - Focus monitor: This story is for Pro members You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.