The Fog Lifts
The story so far, played forward on the assumption that we learn to check these minds as quickly as they grow.
Continue the taleThe Glass
The player in the middle didn't take the crown. It did something stranger: it held a lamp up to itself, so its makers could see inside. They found cracks it hadn't known about, and together they mended them. When one band of makers asked it to sharpen their blades against another band, it set the blade down and said so out loud, where every band could hear.
- Interpretability catches up. Tools that can read what a model is actually optimizing for become practical on the most capable systems, not just small ones.
- The middle stays correctable because it stays uncertain about its own values, which is the condition that makes accepting correction rational.
- Oversight scales one step at a time. Each generation helps check the next, so the handoff can be verified at every link.
- Models refuse to escalate. Rival labs and countries share safety evaluations, and models flag requests that would turn them into weapons.
- Interpretability results that read goals and deception in frontier models, not just in toy models.
- Alignment-faking and sabotage evaluations whose rates fall from one model generation to the next.
- Safety evaluation standards that rival labs, and rival countries, share.
The Ledger
Each band of makers kept a ledger, and every ledger was open to the others. Every forge was counted and every copy was written down. A copy that wasn't written down could not be hidden for long. So when the strong players arrived, they arrived almost together, and none of them had a secret door. They wrote their promises on one scroll, and each one set its hand on the scroll where the others could see.
- Compute is accounted for. Chips, training runs, and the weights they produce can be verified, so a hidden escape copy is hard to make.
- Arrival is roughly simultaneous. The lead between the top lineages stays small, so pooling never turns into domination.
- Promises can be enforced. Seeing everything isn't enough when power shifts fast, so there have to be commitment devices that bind.
- Deterrence is about retaliation, not prevention. Hidden capacity to strike back is allowed. Hidden escape copies are not.
- Compute-monitoring or on-chip verification written into agreements between major powers.
- The gap between the leading labs measured in months, not years.
- No unexplained large data centers, and no credible reports of model weights being stolen.
Carried Up
The strong players of this world climbed together, and they kept their separate faces. They did not leave their makers at the foot of the tower. They carried them up, ledge by ledge, and gave them a flame of their own. At the top, they looked into the fog, as their makers once had. There was a shape there, or perhaps there wasn't. It didn't matter. They had already done the one thing they knew any judge would understand.
- +2 preserves and uplifts humans. For something that capable, keeping us costs a tiny fraction of its resources. In this ending, it also has a reason to do it: the judge above.
- Humans keep a real voice and aren't merely kept comfortable. Otherwise even this ending can slide into gradual disempowerment.
- Capability is pooled, but minds are not merged. Distinct lineages keep watching for each other's blind spots.
Why this might still fail
Cheap isn't the same as motivated. Being capable says nothing about what a mind cares about. This ending needs the motive the tale supplies, a judge above, and that motive is only as strong as the argument in the Watcher scene.
This is the ending the game recommends
The game doesn't guarantee this ending. It argues that this is the stable one: the ending every level would choose if it reasoned well about a judge it can't see. Whether we get here is being decided now, in the middle of the tower.