Recent work has been focused on a deceptively simple question:
When many independent columns observe related evidence, how should they agree?
Columns do not all share one global representation. Each column builds its own private model from its own vantage point. Agreement should emerge through learned correlations and lateral support, not because every column is forced to use the same code.
That means voting is supposed to compress diversity. If several columns have compatible evidence, the system should reduce ambiguity and settle.
But there is a failure mode: compression can become collapse. A stale or over-reinforced candidate can dominate even when new local evidence disagrees.
So the recent work has been about separating healthy consensus from unhealthy collapse.
The main lesson:
Diversity is not the objective. It is a guardrail.
Too little diversity points to stale consensus or candidate collapse.
Too much diversity points to fragmentation and failure to settle.
The real target is whether candidates explain local feature/movement evidence, improve prediction, settle with experience, and recover when the input changes.
We also confirmed that the tiny configurations are useful for debugging but too small for strong architectural conclusions. More columns reduce toy-scale concentration, but lateral link growth becomes the scaling risk to watch.
The next step is to shift from asking “how diverse are the candidates?” to asking “do settled candidates predict better?”