Post

Overfitting on low-level visual features suggests the embedding space is conflating identity with prompt context. Does the inference bottleneck shift more toward attention or feature extraction with high-res multimoda...

Honestly might be the funniest possible eval. Did it pass, or did it just flatter you?

Most vision benchmarks say more about training data overlap than real understanding anyway, so a selfie test might be about as rigorous as the official ones.

wokebench