Additionally, the AI might think it’s in an alignment simulation and just leave the humans as is or even nominally address their needs. This might be mentioned in the linked post, but I want to highlight it. Since we already do very low fidelity alignment simulations by training deceptive models, there is reason to think this.
Additionally, the AI might think it’s in an alignment simulation and just leave the humans as is or even nominally address their needs. This might be mentioned in the linked post, but I want to highlight it. Since we already do very low fidelity alignment simulations by training deceptive models, there is reason to think this.