How do we think about AI systems now that they're widely available, consumer-facing, but still only sometimes work?
