The Reframe: Why Worlds Are the AI Curriculum, Not the Product
Keywords: Artificial General Intelligence, AGI, Developmental AI, Active Inference, Cognitive Science, Curriculum Learning, Machine Learning, Research Methodology, Project Halo
If you build a custom neural network that becomes exceptionally good at playing Chess, you have not built a general intelligence. You have built a Chess engine. If you build an agent that navigates a specific 2D maze with high efficiency, you have not built AGI. You have built a maze-solver.
For decades, AI research has fallen into the trap of over-instrumenting the toy. Builders create a virtual world, tweak a model's hyperparameters until it solves that specific world, and claim progress. But the model's competence is tied entirely to the specific coordinates, visual cues, and state variables of that single game.
It cannot leave. It cannot adapt.
After our struggles with the moving orb in Volume 2, we had a moment of brutal, honest reckoning. We had spent weeks trying to get a recurrent network to track a bouncing red dot, adding BPTT steps and stacking frames, only to realize we were circling a primitive. We were trying to squeeze AGI-shaped behavior from a 12x12 gridworld toy.
We realized the world itself was not the goal. Worlds are curriculum, not the product.
We rewrote our core strategy. We drafted ARCHITECTURE.md and reframed the project around a developmental ladder.
The Developmental Ladder
We defined a sequence of 10 cognitive "organs" that a developmental intelligence engine must acquire to achieve generality. The rungs were designed in a strict dependency order:
[1. Self-Preservation Drive] ──> [2. Relational Perception] ──> [3. Explicit Memory]
│
[6. Cross-World Transfer] <── [5. Abstraction] <── [4. GENERATIVE WORLD MODEL]
│
[7. Skill Composition] ──> [8. Active Experimentation] ──> [9. Self-Model of Learning]
│
[10. Open-Ended Curriculum]
We resolved to follow a strict, self-imposed discipline to prevent self-deception: 1. One organ at a time: Never stack multiple changes and guess which one did the work. 2. Ablation before belief: Every new organ must be tested against an ablated version (with the organ turned off) and a structure-destroying control. 3. Resurrection is the gold standard: The ultimate proof of a new organ is that it makes a previously impossible task learnable, rather than making an easy task slightly faster. 4. No organ ships before its headroom is measured in its absence: We must prove the world requires the capability before we build the mind to solve it.
Defining the Exit Criteria
We frozen the v1 codebase and established the v0 $\rightarrow$ v1 exit criteria. Before we could focus on prediction or abstraction, we had to prove that: * Dual drives (curiosity and caution) were present and measured. * Bounded memory with value-aware eviction worked. * A reusable metrics/experiment harness existed. * Results were robust across multiple random layouts and seeds.
Only when all 5 criteria were met did we graduate. This ensured that every rung we built was standing on a solid, verified foundation.
What We Left Behind
By reframing the project, we made a conscious decision to stop chasing "AGI-in-a-toy." We accepted that our creature would start near zero competence and only ever master the tiny environments we constructed.
The value was not in the capability of the creature itself, but in the methodology: creating a glass-walled, fully inspectable agent where we could watch prediction error collapse, mutate components, and measure the outcome one variable at a time.
With this new discipline in place, we turned to the first major validation of the ladder: cross-world transfer.
In the next chapter, we look at Episode 7: how we proved that a learned causal belief could cross from one world family to another, and how we mathematically decomposed what carried.