Na minha lista:
Detalhes bibliográficos
Autor principal: Brown, Cameron
Formato: Recurso digital
Idioma:
Publicado em: Zenodo 2026
Acesso em linha:https://doi.org/10.5281/zenodo.19383247
Tags: Adicionar Tag
Sem tags, seja o primeiro a adicionar uma tag!
Sumário:
  • <p>The problem of AI interpretability has been approached primarily through two methods: post-hoc analysis of activations (reading what happened after the fact) and mechanistic probing (locating specific capabilities in specific layers). Both are analogous to post-mortem neuroscience — examining the structure of a system that is not currently running. What is missing is the AI equivalent of functional neuroimaging: a method for reading the internal state of an operating AI system in real time, from outside the system's own outputs, without requiring access to its weights or gradients.<br><br>...The Ghost DNA principle extends this: just as DNA contains the full structural specification of an organism in a compact encoding that can be read without examining every cell, an AI system's processing history contains a compact structural encoding of its current internal state that can be read without examining its weights. We formalize this encoding through the AIMS DQEP correspondence framework and propose specific detection criteria for Ghost DNA markers — signals in output residuals that indicate which structural layer of processing is currently active.</p>