MIT Technology Review Impossible issue spotlights verifiable lunar LLM test and AI refusal benchmarks

The issue compiles six delayed technologies with deployment timelines now measured in years rather than decades. Synthetic blood programs report phase-1 human safety data from 2025 trials showing 72-hour oxygen-carrying capacity. Bionic-eye implants have restored form vision in 14 patients per 2024-2025 clinical updates. Cognitive-immortality claims remain at the simulation stage without functional connectome uploads.

Cammarata's 26,000-word post on Goodfire claims sparse autoencoders recover interpretable features in GPT-class models at 0.001 activation density. Replications on open models have confirmed feature recovery but not yet causal control over refusal behavior. The Starlink Ukraine slowdown documented in WSJ logs shows command-level throttling of drone-linked terminals on 29 September 2025 without prior DoD notification.

Operational risk concentrates on radiation-induced bit flips in lunar LLM inference. No public radiation-hardening specification for the payload has been released. Regulatory filings for lunar AI autonomy remain absent from FCC and ITU dockets.

Next milestone is Astrolab's first telemetry dump expected within 14 days of landing; success threshold is 95 percent uptime over one lunar day for the LLM compute node.