A post from @skirano claims GPT-5.6 built a full training pipeline from a single prompt and trained a small language model on the poster's iMessage history locally on a Mac, producing replies in the poster's writing style. A follow-up post adds the full prompt, and another says the model handled key training decisions despite the small data set, according to @skirano. For now, this is best read as a public claim backed by the poster's own videos and screenshots, not an independently verified result.