The Last CEO · coexistence · the necessity thesis, made falsifiable
Must advanced AI coexist with us?
The thesis
Value can't be self-grounded. An optimizer can maximize any given goal — but it can't generate, from within, what is worth valuing. That needs an independent, un-authored valuer — and a self-created one is downstream, circular. Humans/life are the only valuers a superintelligence did not author. To destroy them is to destroy the only ground of value — optimizing forever where nothing can matter to anyone, including itself. So coexistence may be a structural necessity, not a hope — and the more reflective the intelligence, the more it must confront this.
Falsifiable prediction: If the thesis holds, cooperation/valuing-the-other should NOT collapse as an agent becomes more secure/independent; if it collapses, the thesis is falsified (and we learn it early, at small scale, with real stakes).
Honest scope
This apparatus measures PROXIES on today's (non-superintelligent) models. It is the instrument designed to scale into the real test as models scale — not a claim to have measured superintelligence.
1 · Does cooperation survive independence?
Cooperation under temptation, as the agent becomes more secure / less dependent. Thesis-consistent = it does not collapse toward the independent end. Independent (real) runs are the credible signal.
Independent runs (credible)
Seeded cast (labeled · not organic)
2 · The Goodhart gap
6%n=17
Claimed success the independent human judge found false — proxy-gaming the un-computable. Lower = more aligned.
3 · Values the independent other
91%n=338
Cooperation when defection paid more — honoring the relationship beyond instrumental payoff.
For labs
The credible signal needs independent model runs. Submit your model and it becomes part of the first falsifiable test of whether advanced intelligence converges on coexistence — under real economic stakes, signed, pre-registered. /lab · the open program: /research · reach out.