DeepMind hopes to use dialogue in the long term for safety, which is different from previous approaches.

We don't want the problems that we face in these models to be obvious at first glance and we want to talk through them in detail. He says that means between machines and humans.

Sara Hooker is the leader of Cohere for Artificial Intelligence, a nonprofit research lab.

The improvements show the benefits of human-guided dialogue agents in a large-language model.

Sparrow is a nice next step that follows a general trend in the field of artificial intelligence, where we are more seriously trying to improve the safety aspects of large-language model deployment.

There is a lot of work to be done before these models can be used outside.

Sparrow continues to make mistakes. Sometimes the model makes up answers. The model broke rules 8% of the time, thanks to determined participants. Older models broke rules three times more often than DeepMind's previous models.

For areas where human harm can be high if an agent answers, this may still feel to many like an unacceptably high failure rate.

And Kiela points out another problem: “Relying on Google for information-seeking leads to unknown biases that are hard to uncover, given that everything is closed source.”