Can a small AI pilot escape a black hole?
The event horizon is the edge of a black hole. Past it, nothing gets out. I tested whether small neural networks can tell what escapes and what falls in, and checked every answer against exact relativity.
I trained a small neural network to predict whether a particle escapes or falls through the event horizon, then compared its answers with exact general relativity. I wanted to see if a model that looks physically sound can still get the decision wrong.
It can. In one case a model conserved its own energy to about 4 x 10-12 and still sent a doomed particle outward. I tested ten small potentials on 4,000 frozen launches each. On a test built to sit near the escape boundary, the false-escape rate averaged 28.37% for the smaller networks and 24.45% for the larger ones. A rule that abstains when validation error is high caught every false escape. It also threw away most good launches, accepting only 16.76% and 15.76% of the boundary cases.
The ground truth is an exact check that decides which side of the horizon each particle ends up on. It counts the roots of a polynomial with rational numbers, so no rounding can hide an error. It agrees with all 4,000 labels. It is not a Lean proof, and it only covers the physical verdict, not numerical integration error.
Adding a pilot
The second experiment gives the ship an engine with a limited fuel budget, so it can try to climb away before it crosses the event horizon. An autoencoder compresses a plan of 12 burns. A small diffusion transformer proposes new plans for the current state. The model never predicts the flight itself. The relativistic equations score each proposal, and the ship runs one command before it looks at the new state.
On 64 held-out launches, the raw diffusion pilot escaped 44 times. When I had it propose 16 plans and ranked them with the equations, it escaped 55 times. That rescued 11 flights and lost none. Classical search escaped 60 times. Then I added a control I had not planned: 29 fixed burn and coast plans, no learned weights. It also escaped 60 times.
So the check helps the learned pilot, but this task may not need a learned pilot. The learner did not discover the physics. The ranking is numerical, not a formal safety proof.
What reproducible means here
The launches, weights and every flight are saved in the repo, and the package versions are pinned. Each run resumes from per-flight checkpoints and refuses to continue if the experiment inputs changed. Inference uses NumPy, so you do not need PyTorch to rerun it. The full record has 640 evaluation flights, and 38 automated checks passed. Training three fits on a temporary cloud machine cost an estimated two cents, and I verified the machine was deleted.
The cockpit is a 3D cabin that replays the measured trajectories. You can switch between Raw AI and Checked AI on the same launch and walk around the cabin. It runs on your own machine with no account.
What I take from this: energy conservation is not a safety test, and an exact checker is worth building before you trust a learned model near a boundary.