Rejected for the following reason(s):
This is an automated rejection. No LLM generated, assisted/co-written, or edited work.
Read full explanation
This is a linkpost for https://zenodo.org/communities/verok-scf-dve/records?q=&l=list&p=1&s=10&sort=newest
Internal resistance during the formation of probabilistic reasoning
(The black box may no longer be a black box)
What happens if you can think across domains, but don’t really have the language of those domains?
I’m an architect. I design buildings from blueprints. I’m used to thinking structurally, and in terms of how things are actually built to function. Over time, that merged with a more reflective, almost philosophical way of thinking.
About two years ago, when I started working with AI, the question wasn’t what it could do. It was, what shapes the reasoning process, where does it actually form, and at what point can it still be changed?
Most current approaches try to control AI at the output level, by shaping inputs, injecting external information, or filtering outputs.
But after thousands of long-form interactions and experiments, I found something consistent. By the time an output appears, the trajectory has already formed.
(“trajectory” refers to the evolving path of the model’s internal probability states during inference.)
What VEROK–SCF–DVE proposes is simple: intervention must occur during trajectory formation, not after.
In other words:
“The black box” is not something to explain after the fact, it is something that can be structurally shaped while it is forming.
Open question:
If intervention is needed to alter the trajectory formation process rather than the output level, what is the minimum observational signal indicating that the trajectory is still modifiable?
How can this be measured in current AI systems?