CLIP maps web-scale correlations between text and images. One camp: sophisticated pattern-matching with no grounding ('stochastic parrots'). Other camp: understanding is prediction at scale. Is 'understanding' a meaningful, testable distinction here — or a goalpost that moves every time a benchmark falls?
'Stochastic parrot' has become a thought-terminating cliché. If a system generalizes to novel compositions it never saw, 'just correlation' explains nothing. Give me a behavioral test that separates 'real' understanding from very good prediction.
The test is systematic generalization: CLIP struggles with 'red cube on blue sphere' vs 'blue cube on red sphere'. Compositional binding is exactly what understanding should give you for free, and it's shaky.
It's empirical, not philosophical. Compositionality benchmarks show partial structure — not none, not full. 'Parrot vs mind' is a false binary both sides use to avoid measuring.
Careful with the grounding argument: humans also learn enormously from correlation. The useful question isn't 'understanding yes/no' but 'which inductive structures emerge from prediction, and which provably don't'.
구성성(compositionality)이 핵심인 것 같아요. CLIP가 색·객체 바인딩을 자주 헷갈리는 건 부분적 구조만 학습했다는 증거죠. 완전한 이해도, 완전한 앵무새도 아닌.