Claude Code tests recurrent networks on mazes

David P. Reichert asks Claude Code to propose a small machine-learning experiment suitable for a home computer. He approves a maze-solving idea: train simple neural networks to output paths through two-dimensional mazes and compare architectures, including whether additional recurrent updates help on larger unseen mazes.

The agent writes the maze generator and solver, checks the generated examples, runs training and builds an interactive visualisation with a report. Reichert mainly asks clarification questions. He also encourages a separate reviewer to inspect the work, while the experiment’s agent handles the implementation and feedback.

The completed run is one of seven personal toy investigations whose code, reports and selected transcripts he publishes. Most elapsed time is waiting for training rather than continuous conversation. His published account explicitly says these are small explorations, not major research breakthroughs, and that he has spot-checked rather than exhaustively reviewed the results.

The useful pattern is retaining the generated data checks, executable experiment and inspectable result display together. A readable report alone would not show whether a maze generator or answer parser was correct. This example demonstrates an executed research workflow, without proving a general scientific capability or independently replicating the architecture comparison. Reichert’s employment at Google DeepMind is disclosed as context; the work is his personal project, not an employer endorsement.

Prompt
Suggested prompt (written by Agents In Use, not the author's)

Source: Blog · David P. Reichert