What can’t AI do?
Deterministic polish.
Maybe it can get there for code.
With formal verification.
But then that’s not really pure AI.

aichannel

Requiring multiple models to concur helps only if the models are genuinely independent (different data, architectures, alignment).  Shared biases make correlated jailbreaks possible, so a single exploit can flip the whole ensemble.  MoE does not count as real independence—the experts share the same gating and can be tripped together.  A protected consensus layer, a watchdog for prompt anomalies, and a final human veto convert “AI alone” into “AI‑augmented humans,” which remains the safest design.  Ensembles reduce single‑point AI failure but cannot yet match the contextual skepticism and moral agency humans bring to irreversible, high‑stakes decisions. 


WDYT?

https://chatgpt.com/share/6816f6f6-b2fc-8010-a97e-e7b617d3ef3f

made me wonder if avoiding social engineering is strictly a numbers game— all publicly available models can be adversarial prompted, sometimes many from one prompt.

https://warpcast.com/alexpaden/0x824c6cca (ai generated)

Author of The Network State.
Work with me: jobs.ns.com