Subjects › AGI Agents (6-week course)
W1 · Cognitive architectures
The perceive → remember → deliberate → act → learn loop, CoALA, and where 'AGI' levels come from.
The idea
The cognitive loop ties the world model, planner and memories together. In episode 1 the agent knows nothing, so it explores with count-based curiosity (the least-tried safe action) until it reaches the goal and stores (cell, is_a, goal) in semantic memory; from episode 2 it runs A* inside its learned world model — which is how the demo goes from 199 steps to 10. Surprise (predicted next state ≠ observed) is logged to episodic memory and clears the plan so the agent replans; it never fires in the deterministic demo, which is why Experiment A adds slip=0.2. CoALA is the vocabulary for mapping any agent onto this loop; 'Levels of AGI' grades performance × generality. In your repo: Week01_Foundations_Cognition/cognitive_loop.py and MILESTONE.md.
Lesson
No lesson built yet. The worked solutions below are the material it would be assembled from.
Questions (5)
| Question | Type | Difficulty | Drift | Accuracy |
|---|---|---|---|---|
| The cognitive agent is part-way through an A* plan when it throws the plan away and plans again, mid-episode. … | mcq | Easy -1.30 | — | — (0) |
| In the cognitive_loop.py demo, episode 1 takes about 199 steps and episode 2 takes 10. What changed inside the… | mcq | Easy -0.50 | — | — (0) |
| In CoALA's terms, what kind of action is it when deliberate() runs astar_plan over the learned world model and… | mcq | Moderate 0.40 | — | — (0) |
| A chess program beats 99% of skilled human players but can do nothing except play chess. How does the 'Levels … | mcq | Hard 0.90 | — | — (0) |
| Experiment A runs the cognitive agent in GridWorld(slip=0.2), where a move sometimes lands somewhere unintende… | short | Very hard 1.60 | — | — (0) |