[ACCount36]

Even the very best multimodal LLMs still suffer from a harsh perception bottleneck. They're impressive, but nowhere near as good as human visual cortex.