How Code Generation Fixed My AI Training

video1.0<iframe src="https://www.loom.com/embed/9d30b771cc5644fa9e48688c31e0e1c3" frameborder="0" width="1920" height="1440" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>14401920Loomhttps://www.loom.com14401920https://cdn.loom.com/os/loom-media-content/loom/ec535bad-4599-41d9-a6f4-9dde24150689/a8edeef5-5639-4f65-af49-57f1c5d27c3b/sessions-thumbnails-9d30b771cc5644fa9e48688c31e0e1c3-bc5f2fa6b3ddacd5.gif/binary?authToken=eyJraWQiOiJtaWNyb3MtZ3JvdXAvdGRwLW9zLWFyY2hldHlwZS9yZnFxa3FhcHB2bGJ2bDVmIiwiYWxnIjoiUlMyNTYifQ.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.fqMU4BoxX9xBmqEuw7ShsyRoPGx4wgJTt9VcBvsqeGEWDdEASy77jqekOWKV1hxQwWT2c4rrAkofE9AYklUeIYVUbV1KB69_xoAxw44sCJQw_zcuFrhMZF66iY-lYW-PIZgmkSrd1qrjqMPDS3l5hAshU5bOAIUg8pTCd2UWPZZAFi498PzKdJ3P1C04xKZI3VUIpiio9AlSIyLf3VQu16zAFlhUz9j_L4tvFwV4d5Flp0JsRkCamluAy6HG5XonzadXqQdcny2wA9HRYwvVNX_tbwzMRehApYzdlbzfddwPiOz5hURhiIrSyJP53NRL4zOnOen9M-FnfIE0CqQZbg&Expires=1793646241&Signature=Yp%7ETyHz-wCvgIt31dct5ass6gN%7EUMd%7EusILUvsa6fKO6DAPTqWIXf5xoY%7EqYiCuezO8V6G7c5PihemgxRVZvnrr1Ph7guQcDpf-yG3PasfWNMmeum3N0CD0puQ4qeRMZM-jsRAliM-h3%7E4WwBEO57JrYKmeYw8SVPXpEgmED19SrR9Rv9l-Zj8%7EJTeTDHYhdEIxyVa%7ETpysBerRpTJfG3uE8rOH3ZAbmmFfibtgNzUJqX0otXD9OGZAs8yt23uYqNtREklF-8ZIined1wDH2vXS636IMK5GqhEmG8NXxih4YekJgmeIjraW3MQvjp4Hrqz4dXefOJZS7Tz4sM1Ux4A__&Key-Pair-Id=KQOSYIR44AIC0339.441How Code Generation Fixed My AI TrainingThis Loom shares Artem’s real debugging story about a solitaire neural network originally written with quad-code. He describes a public solitaire agent that, with lookahead and search, found wins in 55.8% of games, then became less effective after he built a game engine and visualization to inspect the network’s move behavior. The issue was reward shaping in the game engine, where the reward logic produced negative rewards and improper loop incentives, which he then fixed by balancing rewards, for example plus five points for empty columns and minus five when refilled, and aligning training to avoid internal bonuses. He also explains adding a rule and tests to prevent this behavior from recurring, as part of his use of GlotCode beyond just coding.