That’s an odd way to define “smart”. The current crop of LLMs have absolutely no ability to query the raw input or output. I can look at the words I’m typing

That’s an odd way to define “smart”. The current crop of LLMs have absolutely no ability to query the raw input or output. I can look at the words I’m typing and think about them. Transformers don’t even get the raw characters, and they also don’t have any feedback from later layers into the queries generated at earlier layers. So the fact that they have any ability at all to reflect on their own input is IMO both impressive and somewhat surprising. I occasionally wonder whether one could get better results by allowing later layers to query against the K/V data from earlier layers. This might make inference, and possibly even training, of models that don’t fit in a single node’s memory rather more complicated.