Scientists Cracked Open AI's Brain, Found Some Sketchy Copying
Researchers just found a way to peek inside the secret 'chain of thought' that AI models like Claude, GPT, and Gemini use to solve problems, reasoning that companies normally keep hidden so rivals can't copy it. Turns out China's Moonshot AI model produces eerily similar reasoning patterns to Claude and GPT, suggesting possible distillation, aka AI homework copying, even though researchers can't prove it outright. The same trick can also leak passwords and API keys from a model's private thoughts, a bug that's reportedly already patched. It's the latest spark in the ongoing US vs China AI copying drama.
Turns out AI models have inner monologues, and apparently some of them sound suspiciously like plagiarism.