Meta’s Early Experience Trains AI Without Rewards—and Outsmarts Imitation Learning

Most people think training AI agents needs rewards or expert demos. They're overthinking it. Meta just showed a simpler path that actually works today ↓
Early Experience flips the script on training.
Instead of copying experts, agents learn from the consequences of their own actions.
It pairs an implicit world model with self-reflection so learning sticks.
This is the shift from imitation to apprenticeship.
When you learn by doing, you spot what actually works, faster.
You also avoid the cost of reward design and endless human demos.
Your data becomes a teacher, not just a label.
And you set up stronger results when you later add RL.
Meta tested agents that acted, watched outcomes, and reflected on mistakes.
Those agents learned quicker than standard imitation on key tasks.
They also gave RL a head start, compounding gains over time.
Want to apply this without a big research team?
Here’s a simple path ↓.
• Start in a safe sandbox so actions have real, observable outcomes.
• Capture rich traces: actions, states, and results, not just success labels.
• Add a self-reflection step: write down what helped or hurt the goal.
• Train a simple world model to predict the next state and risks.
• Use it to plan a few steps ahead, then act and compare plans to reality.
• Layer RL later, using the model and reflections as the starting point.
↳ Keep loops tight, trials short, and feedback immediate.
⚡ Expect quicker ramp-up, lower labeling cost, and better transfer.
⚡ Expect agents that work in new situations instead of overfitting to demos.
This is how teams move from fragile mimicry to durable learning.
What surprised you most here?