Random bias I noticed with LLMs. When having to choose a random day of the week, they are slightly more likely to choose Tuesday. I really don’t know where this bias comes from.

LLM will become closer to the architecture of the brain and much more dynamic and flexible over time. The main reasons this has not been done already is that such systems are far more difficult to understand and train than feedforward ones but this won't hold true for much longer.

My bet is that the next generation of model will become more and more cyclical and more similar to the brain, with inputs and output more organically connected to the model neurons and increasingly cyclical internal processing.