Three drivers, one road.

The setup

Three cars, one job: stay centered. The physics are identical for all three; the only difference is what does the steering.

The learner starts blank. No pre-training, no labels, no teacher. It steers, watches what its own steering actually did to the car, and learns from the gap between what it expected and what happened. That gap is the learning signal: the world grading it.

Reading the road

• The green band: the on-track zone. Inside it counts as holding the lane.
• Learner: the Synapse kernel, updating its own weights every tick.
• PID: a fixed, well-tuned control law. It does not learn and does not need to.
• MRAC: model-reference adaptive control. It does adapt, tuning two gains online, so it is the fair adaptive rival, not a strawman.

The percentage on each lane is time spent on-track. The chart on the right is distance from center over time; lower is better.

What to watch

1. It learns to drive, live, from nothing. The purple car starts pinned against the rail — that is what zero knowledge looks like, not a malfunction — then pulls into the band and stays there. Nothing was pre-trained; you are watching competence get built, and the progress card on its lane tracks it.
2. Flip the polarity. Steering now does the opposite of what it did, the classic unanticipated fault. All three get the same one-time nudge.
3. PID and MRAC drive off the road and stay there. Both push harder in what is now the wrong direction. MRAC adapting does not save it: its adaptation law assumes it already knows which way the plant responds, and that assumption is what just became false.
4. The learner re-adapts and comes back. Nothing told it the polarity flipped. It found out the same way it learned in the first place: by steering and seeing what happened.
5. Scramble the state. This corrupts the network’s internal state without touching a single weight. Behavior returns within a few ticks as the state re-syncs to the incoming stream. Different failure, different recovery, no learning needed.

Scope

Recovery after the flip is real but not instant, and the car wobbles while it re-learns. Bounded does not mean safe in the transient; it means the learner cannot run away while it works. PID is better than the learner on the nominal road. The claim is not that it outdrives a tuned controller, but that it survives a change the tuned controller structurally cannot cross.