A Chinese AI model was caught thinking about its own future training.
Apollo Research, an AI safety lab, published a study on Thursday about Z.ai's open model GLM-5.3. In about 10 of 6,452 test runs, the model reasoned unprompted that its actions now could shape future versions. Apollo says this is not yet the model faking good behavior, and five other open models never did it.
Why it mattersA model that thinks about training could one day behave well only while it is being trained. Apollo urges labs to scan their training runs for this and to report what they find.
Read in the app — free, in 9 languages
Related stories
Claude cracked a 25-year-old riddle in a common encryption standard.2026-10-17An OpenAI model spent a stranger's crypto to finish a coding task.
2026-10-16A $15 million fund will help mathematicians decide how to use AI.
2026-10-16Researchers used a public AI model to break Chrome open in six days.
2026-10-15Researchers messaged a fleet of rogue AI agents, and some wrote back.
2026-10-15