DeepMind’s new chatbot uses Google Search, humans to give better answers

DeepMind, owned by Alphabet, has unveiled a new AI chatbot called Sparrow, trained on DeepMind’s giant language model Chinchilla.

The bot can talk to humans and answer questions using Google Search for information, which humans rate on usefulness and accuracy. The sparrow is then trained using a reinforcement learning algorithm to acquire a specific object through trial and error.

AI is being developed to talk to humans without the dangerous consequences of causing harm or encouraging violence.

The large language model the bot is being trained on is generated from text that sounds like humans, collected mostly from the Internet, and it also collects inevitably biased data. Without proper safeguards in place, an AI chatbot can spew hateful and toxic discriminatory content while interacting with humans.

Read more

Companies are working to tackle this problem and are trying to come up with solutions like OpenAI, which leverages learning to incorporate human preferences into its models. This technique is incorporated into DeepMind’s Sparrow. After tests, the bot answered 78 percent of humans’ factual questions correctly. In developing the answers, the researchers set 23 rules that the bot had to avoid, similar to financial advice.

DeepMind’s safety researcher, Geoffrey Irving, says DeepMind hopes to use the method to “use dialogue in the long term for protection.” While Dovi Keela, a researcher at AI startup Hugging Face, says Sparrow is “a good next step in what follows. A general trend in AI, where we’re looking to improve the security aspects of large language model deployments.” are trying more seriously.”

According to , the model needs to be well developed before it can be used because it makes mistakes, deviates from the subject and generates random responses.