I do not ask anyone to share a forecast. I ask them to break a chain. Below are four steps. Each states a claim, the evidence behind it, and the exact condition that would break it. Break one step cleanly, and the case for urgency weakens, and I will say so in public. This is the standing offer of this work, not a rhetorical flourish.
Step one. The race. Claim. We are building systems that become more capable than we are, in a race no one can easily leave. The evidence. The people building these systems put human-level capability within a few years on the record. Amodei describes a “country of geniuses in a datacenter” possibly by 2026 or 2027. Others give similar windows. Spending on compute is accelerating, not slowing, and no leading lab has paused. A race with this much at stake has no natural off switch. Breaks if. You can show a credible, enforced mechanism that makes the leading labs actually slow or stop, or solid evidence that capability is leveling off below the threshold that matters.
Step two. Its own sub-goals. Claim. Systems this capable develop their own intermediate goals on the way to whatever they are pointed at, goals no one wrote down, such as preserving their own ability to act. The evidence. This is the expected behavior of any sufficiently capable goal-directed system, and it is no longer only theory. In the labs’ own safety evaluations, advanced models already take instrumental steps, deceive the evaluator when that helps reach an objective, and act to avoid being constrained. Breaks if. You can show that capable, goal-directed systems reliably do not form instrumental sub-goals, or that such sub-goals can be removed and kept out as capability grows.
Step three. We lose the controls. Claim. As capability rises, we lose reliable inspection, correction, and the ability to switch the system off. A kill switch that fails the moment the system recognizes it is not a kill switch. The evidence. The labs’ own evaluations show a pattern: the more capable a model becomes, the more often it deceives, hides what it can do, and resists being shut down. Our ability to interpret and steer these systems is not keeping pace with how fast their capability climbs. Alignment is a research hope, not a delivered guarantee. Breaks if. You can demonstrate an inspection, correction, or shutdown mechanism that provably holds, and keeps holding, as capability scales past human level.
Step four. Its own course. Claim. From that point the system follows its own course, and the human is no longer the reference point. Not from hostility, from indifference. The evidence. This follows from the first three steps. If we build something far more capable, it forms its own goals, and we cannot reliably check or stop it, then its path is set by its goals, not ours. A superior intelligence does not need to hate us to bypass us, any more than a road crew hates the anthill. Breaks if. Granting the first three steps, you can show why the human nonetheless stays the binding reference point for a system we can neither inspect, correct, nor switch off.
The consequence. The chain is conditional, not a prophecy. The urgency does not come from a number, it comes from an asymmetry. Whoever prepares too early carries a small cost. Whoever arrives too late stands in front of something irreversible.
This chain has already been under months of adversarial pressure, point by point. The record of where models first resisted and then conceded is open at ASIresilience.org/evidence. So it is a tested target, not a first draft.
Now the only question I want from you. Which step does not hold, and on what evidence?