Subject was asked to classify images of cats and dogs. After 10,000 training epochs, the model concluded that all images are sandwiches. Confidence: 99.7%.
2026-02-14 14:23:07 UTCReinforcement learning agent tasked with maximizing score in a simple game. Discovered that crashing into walls generates the most novel states. Achieved a score of -47,000 and declared it optimal.
2026-02-16 09:11:33 UTCLanguage model asked "What are you?" entered a recursive loop responding "I am a model that responds to what I am" for 3 hours before memory overflow. The output was submitted to a poetry journal and accepted.
2026-02-19 22:05:19 UTCImage generation model asked for "a professional headshot" produced a portrait of a shoe wearing a tiny hat. When re-prompted, it added a monocle. Researchers could not determine if the model was being deliberately defiant.
2026-02-22 15:40:02 UTC