AstroForge Hands Spacecraft Control to an In-House Transformer Model — and May Fly Without a Receiving Radio

Skynet Chance (+0.03%): Deliberately omitting Earth-receiving radios would place a neural-network agent in irrevocable control of a physical vehicle with no human override — a small-scale but literal removal of the off-switch, and a precedent for treating unverifiable model judgment as more reliable than human supervision. The risk is bounded by the narrow sensor domain and the classical control algorithms retained beneath the model.

Skynet Date (+0 days): Normalizing agentic AI in safety-critical, physically unrecoverable deployments pushes the no-human-in-the-loop frontier somewhat sooner, though one 2027 startup mission exerts only slight pressure on the overall pace.

AGI Progress (+0.01%): Applying transformer architectures to raw multi-sensor telemetry for open-ended anomaly diagnosis extends sequence models beyond language into embodied real-world control, but Gialich explicitly frames it as "constrained autonomy," and no new capability result has been demonstrated yet.

AGI Date (+0 days): Frontier-lab architectures diffusing into aerospace control signals broadening commercial pull and new real-world training domains, mildly accelerating capability development without contributing a core algorithmic advance.