Charlot Lab · Energy First Architecture · phase 4 spike (internal)
Phases 1–3 gave EFA a sparse latent that remembers, models-and-plans, and learns without forgetting. Phase 4 is the world model's real payoff, straight from the world-models literature (Ha & Schmidhuber → Dreamer): once you can imagine the world, you can learn a policy inside your own imagination and pay almost no real experience for it. Here EFA saw 4,000 real steps of random play — just enough to learn the world — then trained a goal-reaching policy entirely inside that learned latent model, with zero further real steps, its own latent energy as the only reward. It then acts in the real world below. A conventional model-free learner had to take ~190,000 real steps to do worse. That gap is "intelligence per sample" — the twin of EFA's intelligence per watt.