Pace · hidden frontier · advanced · original

You run an AI lab. Hold to accelerate research; release to coast. Hold to shift research funding from safety to capabilities (all the way in about a second); release to shift it back. Capability is deployed as soon as it's ready. Safety research raises your own frontier; the competitor has its own. Past human level (dotted line) your AI increasingly does your research, then takes off — and research output lands months after you fund it (less as AI speeds up), so capability can surge past safety. New research is ready to deploy two weeks after it's done. Somewhere inside the shaded danger zone is the safety frontier — deploy above it and every month risks catastrophe for everyone. You don't know exactly where it is.

Advanced: no accelerator — you set policy. Split research funding between capability (builds latent capability) and safety (raises your own safety frontier; the competitor has its own, which you never see). Choose how much latent capability to run internally (speeds your research via self-improvement) and how much of that to sell externally (earns money; the only thing your rival sees). Deployment rolls back and restores instantly. Research takes time to pay off: output lands about 3 months later, less as your AI speeds up research (safety lags behind, increasingly so). Past human level (dotted line), your internally deployed AI increasingly does your research for you — then takes off. Internal deployment carries full risk; external a bit more (its catastrophe threshold is slightly lower); undeployed latent capability a quarter. Your danger zone is only an estimate: the true frontier drifts around it and sometimes leaves it, and some apparent safety breakthroughs aren't real. Breakthroughs get re-examined later, and a debunked one (rightly or not) drops your zone back. New capability ships in discrete model releases; dialing deployment down stays smooth (noise injection degrades a model by any amount). Q/A capability, W/S internal, E/D external.