ConvNets are a decent model of how the ventral pathway of the human visual cortex works. But LLMs don't seem to be a good model of how humans process language. There longer-term prediction taking place in the brain. Awesome work by the Brain-AI group at FAIR-Paris.