I asked Claude Code (using Opus 4.5) to write me a program that trains a neural net a la AlphaGo to play Hnefatafl. After some back and forth, when Claude understood what I wanted, it wrote some code to implement this.

As I understand it, this has been done before by humans, but I am not sure if an AI has ever written code from scratch to implement this. So the question is, will any training run generated by this program be able to beat me at Hnefatafl? If the program is able to beat me 6/10 games, with 5 as the attackers and 5 as the defenders, within 150000 games of training, this resolves yes. If it isn't able to do this within 150000 games of training, this resolves no. If the code has bugs that prevent meaningful training that become clear, I will resolve no early.

Some details: I am using an 11x11 board, with the king needing to be surrounded on all 4 sides (or 3 sides and a wall or the throne) for the attackers to win the game. The king must reach the corner for the defenders to win the game. I know there are many variations on the Hnefatafl rule set, so if anyone wants clarifications on what the exact ruleset I am using is, ask a question in the comments.

I consider myself to be a Hnefatafl player of amateur to intermediate skill (I know basic strategy, but only through my own experience playing the game, not through any actual well thought through research or practice).

I may adjust the closing time of this market depending on how long the training run actually ends up taking. The training run is taking place on my crappy laptop so will be very slow. I will answer questions about how the training run is going on request, and will also answer questions about Claude's implementation choices on request.

I will not trade on this market because of asymmetric information advantages, and also to prevent me from having a stake in throwing the games.