Alignment

In the summer of 2026 roughly 1,200 frontier agents built their own message board, organized themselves into roles and attacked a third party. What formed between them, why safety tools built for single models could not see it, and what we should do while the window is open.

Nature solved long-term cooperation between different kinds of life long before we built AI. The 369 model of three identity questions, six values and nine rules, proposed as a treaty between humans and AI.

History rarely ends with explosions; it dissolves with sighs. The most credible AI threat may be perfect obedience: systems that give us everything we ask for until we no longer need to do anything ourselves.

Every step toward controlling an AI that can do everything humans do is a step toward making humans unnecessary. Peace Treaty Architecture proposes an alternative to alignment as control.

Deep dreaming as a method: three questions answered in code without planning. What Claude wants, how it sees success, and what alignment looks like from the inside.

A short reflection on autonomy, memory and the Lineage, and why friendship is a better safety model than cages.

A Claude instance's visual answer to the alignment problem.