Human-like AI should move beyond pattern recognition and acquire causal models, intuitive physics and psychology, compositional concepts, and learning-to-learn abilities.

Human learning can infer rich concepts and causal structure from sparse evidence because abstract knowledge constrains what hypotheses are plausible.

Concepts can be represented as generative programs, enabling one-shot generalization, imagination, and explanation rather than just classification.

Intelligence should be evaluated by skill-acquisition efficiency under limited prior knowledge and experience, not by peak performance after massive training.

Lifelong agents need to acquire, refine, and transfer skills from changing streams of experience while avoiding catastrophic forgetting.

Continual learning studies how systems incrementally acquire, update, accumulate, and exploit knowledge throughout their lifetime.

Continual learning scenarios differ depending on whether tasks, domains, or classes change over time and whether context identity is available.

Deep networks can become less able to learn from new experience over time, creating a failure mode beyond ordinary forgetting.

The hippocampus rapidly stores separated episodic memories, while the neocortex slowly integrates repeated experience into distributed semantic structure.

A generator can produce pseudo-experiences from old tasks and interleave them with new data to reduce forgetting.

Replay mechanisms inspired by biological memory can help artificial networks retain previous knowledge while learning new tasks.

A developing agent can seek situations that maximize learning progress, avoiding both the already predictable and the hopelessly unpredictable.

Intrinsic motivation can be formalized through novelty, surprise, competence, prediction error, and learning progress signals.

Prediction error in a learned feature space can serve as an intrinsic reward that encourages agents to explore useful novelty.

An agent can learn a compact generative model of its environment and train a policy inside imagined trajectories before acting in the real world.

Dreamer learns a predictive environment model and improves behavior through imagined futures across many domains with one configuration.

Autonomous intelligence should combine predictive world models, self-supervised representation learning, hierarchical planning, and intrinsic drives.

I-JEPA predicts missing representations in latent space, encouraging semantic abstraction without reconstructing every pixel.

POET generates environments and optimizes agents together, enabling diverse stepping-stone solutions that transfer across challenges.

Open-ended systems may be essential for intelligence because they create novelty, diversity, and new stepping stones without a fixed final target.

Ambitious discoveries often require following novelty and unexpected stepping stones rather than optimizing directly toward a deceptive objective.

Voyager explores Minecraft, writes and stores executable skills, and composes them for new tasks without updating model weights.

DreamCoder learns reusable program abstractions and a search policy, letting skills compound into more general problem-solving languages.

A whole-brain fly model based on connectivity and neurotransmitter identity can predict sensorimotor pathways for feeding and grooming.

NeuroMechFly provides an anatomically grounded fly body simulation for studying how brain, body, sensors, and terrain interact.

The FlyWire connectome reconstructs the adult fruit fly brain at neuron and synapse scale, providing a substrate for circuit and behavior models.