Open any AI chat tool and ask: 'In 3 bullets, what can you not know about this session?' Then run one checkable task (like reversing a 5-word phrase). You're done when you've noted whether the meta-answer actually helped (Y/N), plus one thing you won't trust from meta-talk alone.
How-to · ORI · ORI.1
Model 'self-awareness' is laggy pattern-match
When a model talks about itself, treat it as incomplete pattern completion from training talk — useful sometimes, not a private inner life.
Worked example
Karpathy points out that a model's 'self-awareness' builds from all the tokens of us talking about them — laggy and incomplete — yet it can still start to 'get' session-level ideas, like compacting context. That's prediction at work, not a soul. Practically: treat meta-talk as a weak signal, and verify it with a checkable task. Don't grant extra privileges just because a model sounded self-aware.