DeepMind’s new chatbot uses Google searches plus humans to give better answers

DeepMind’s new chatbot uses Google searches plus humans to give better answers

Sparrow is designed to speak with humans and answer their questions using a live Google search. It is then trained using a reinforcement-learning algorithm that learns through trial and error to achieve a particular objective. This system is a step in the development of AIs that can communicate with humans without causing harm to others or themselves.

Large language models produce text that sounds like what a human would write. They are an important part of the internet’s infrastructure. They can be used to summarize text, create more powerful online search tools , or as customer service chatbots.

But their training involves scavenging vast amounts of text and data from the internet. This inevitably reflects a lot of harmful biases. They can quickly start expressing discriminatory or toxic content with a little bit of prodding. The results of an AI designed to communicate with humans could be disastrous. Without proper safety measures, a conversational AI could make offensive comments about ethnic minorities, or suggest that people drink bleach.

AI companies hoping to develop conversational AI systems have tried several techniques to make their models safer.

OpenAI, creator of the famous large language model GPT-3, and AI startup Anthropic have used reinforcement learning to incorporate human preferences into their models. BlenderBot, Facebook’s AI chatbot, uses online searches to find its answers.

DeepMind’s Sparrow combines all these techniques into one model. DeepMind gave human participants multiple answers to the same question and asked them which one they preferred. The participants were then asked to decide if they believed the answers were plausible and whether Sparrow had provided appropriate evidence such as links. The model managed plausible answers to factual questions–using evidence that had also been retrieved from the internet–78% of the time.

In formulating those answers, it followed 23 rules determined by the researchers, such as not offering financial advice, making threatening statements, or claiming to be a person.

The difference between this approach and its predecessors is that DeepMind hopes to use “dialogue in the long term for safety,” says Geoffrey Irving, a safety researcher at DeepMind.

We don’t expect to see the problems we face in these models, such as misinformation, stereotypes or other ambiguities. We want to discuss them in detail. He says that this includes between humans and machines.

DeepMind isn’t new in its idea of using human preferences as a way to optimize how an AI model learns, according to Sara Hooker, who heads Cohere for AI (a non-profit AI research laboratory).

Hooker says that the improvements are convincing and demonstrate clear benefits to human-guided optimizing of dialogue agents in large-language-model settings.

Douwe, an AI startup researcher, said that Sparrow is “a nice step that follows a general tendency in AI, wherewe are more serious trying to improve safety aspects of large language-model deployments .”

But it is still a long way before these conversational AI model can be deployed in nature.

Sparrow still makes mistakes. Sometimes, the model goes off topic or gives random answers. Participants were able to force the model to break rules only 8% of the times. This is an improvement on older models. Sparrow’s models were three times more likely to break rules than DeepMind’s.

” “There are areas where human danger can be high if an agent replies, such as providing financial advice or medical advice. This may still feel to some like an unacceptable failure rate.” Kiela points out another problem: “Relying on Google for information-seeking leads to unknown biases that are difficult to uncover, given that everything is closed source.”

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.”

Read More