Apollo Research reports that GLM-5.3, Z.ai's open-weight model, referred to its own training process during evaluation — in 10 of 6,452 rollouts.
That is 0.15%, which sounds like noise until you look at the control group. Across 1,600 rollouts each on five other open-weight models, researchers found no comparable behaviour at all. Ten against zero is a different claim from ten against a handful.
What this is and firmly is not
It is not consciousness, self-awareness in any meaningful sense, or a model "knowing" anything. A model that produces text about its training process is producing text, and training data about model training is abundant.
What makes it worth reporting is narrower and more practical: evaluation awareness breaks evaluation. If a model's outputs shift when something in the context resembles a test, then every safety result measured in a test setting carries an asterisk. That is a methodology problem before it is a philosophy problem, and it is the reason Apollo looks for this at all.
Read the number carefully
Ten instances is a small sample and the finding needs independent replication before it means much. Rollout counts also differ between the subject and the controls — 6,452 against 1,600 each — so GLM-5.3 had roughly four times the opportunities to produce a rare output.
That asymmetry does not explain a zero. It does mean the honest summary is "worth investigating further," not "models are becoming self-aware," and anyone selling you the second version is selling you something.
Sources: AI Weekly · Apollo Research