NEURAL DRIFT

IDLE tf: … 0 tensors

The road isn't real. The car isn't real. The loss function is very real.

01 Reality Engine ground truth

REALITY OFFLINE
you are inside the model now

Autopilot 0 km/h

02 Neural Dream predicted

noise

confidence

hallucination

0 fps

03 Side-by-Side |real − dream|

real

dream

diff

Train a model to see how wrong it is.

Samples during training

Dataset

Record frames 0frames

max

Model

64×64×3 + action[5]conv16/2conv32/2conv64/28×8 ⊕ action mapdeconv32 +skipdeconv16 +skipdeconv16 +skipframe + Δ

lr batch epochs loss

Waiting for a reason to learn.

Dream Mode

Drive inside the model

Arrow keys / WASD steer, in reality and in the dream. Each frame you see in the dream is the network's guess about what your keypress would cause. When it turns to soup, reseed.