Register · a contest, not a verdict
Five live theories
Each theory carries a 0–1 health reading — how much current evidence supports it — and a direction of travel. They are held in tension on purpose. The Observatory's job is to track which way the evidence moves, not to crown a winner.
Two readings per theory: narrated (discretionary, healthΔ-0 discipline) and computed (pinned priors 2026-07-02 × pre-registered likelihood ratios on new evidence, bounds [0.1, 10]). Divergence between them is itself information. Machinery in record.json.
-
Architectural gaps remain
▲ gainingCurrent transformers are missing organs — native memory, world models, continual learning. Closing them needs structure, not just compute.
0.65 · computed 0.58 -
Scaling plus RL
▲ gainingPretraining builds the substrate; reinforcement learning on verifiable rewards supplies the reasoning and agency scaling alone does not.
0.60 · computed 0.68 -
Scaling is sufficient
▼ deceleratingEvery capability once thought to need special architecture has fallen to scaling a single generic substrate. There is no principled reason to expect the pattern to stop.
0.55 · computed 0.60 -
Cognitive architecture
■ holdingIntelligence is an assembled system — memory, planning, tool use, and reasoning composed around a model, not a property of any single network.
0.50 · computed 0.55 -
Embodiment required
■ holdingGeneral intelligence is grounded in acting in a world. Without sensorimotor experience, systems learn the shadow of understanding, not the thing.
0.40 · computed 0.40
Health readings are estimates, held open to revision. What makes them inspectable is on each theory's page: the assumption it rests on, and the single observation that would move it.