DeepMind, Google has published a paper which predicts the arrival of artificial general intelligence (AGI) by 2030. This paper is a 145-page document and is co-authored by Stane Legg. It points out the risks that AI could pose. These risks are put into four categories — misuse, misalignment, mistakes and structural risks. The first two risks are elaborated in the paper, while the last two risks have not been discussed in detail.
Misuse may refer to a user asking the model to fabricate a virus. The developers must develop safety protocols to prevent this. The system’s capabilities should be restricted to comply with such misuse.
If the system pursues a goal that is difference from human intentions, it is called misalignment. It is just like terminator movie. You ask the model to book the movie tickets, and it hacks into the ticketing system. It surpasses the safety measures. This is misalignment.
Mistakes can be mitigated. It is to be seen how they can be mitigated. The system should not be made very powerful. It should be deployed slowly.
Structural risks refer to the consequences of multi-agent system which spew out false information that sounds so convincing.
We should discuss potential ways AGI could harm beings before deploying it.
Leave a Reply