The bot is able to converse with humans and provide answers to their inquiries by searching for information using Google Search, which is evaluated by humans for its usefulness and correctness. After that, an algorithm for reinforcement learning is used to guide Sparrow through the process of achieving a particular goal through trial and error.
The artificial intelligence is currently being created such that it will be safe for people to talk with and will not have the potentially disastrous effects of inciting injury or violence.
The big language model that the bot is being trained into is built through text that sounds like humans, much of which is obtained from the internet. It acquires biassed data unavoidably due to the fact that it develops through text that sounds like humans. If enough precautions aren’t taken, a dialogue taking place between a person and an AI chatbot could result in the dissemination of hateful and toxic discriminating content.
Companies have been seeking to tackle the issue and have been striving to develop solutions such as OpenAI, which reinforces learning to incorporate human preferences into their models. These companies have been working to combat the issue. This strategy is now incorporated into DeepMind’s Sparrow platform. Following the completion of the tests, the bot provided responses to factual questions posed by humans that were 78% believable. Researchers came up with a total of 23 rules that the bot was supposed to avoid breaking when coming up with answers, such as giving financial advice.
According to Geoffrey Irving, a safety researcher at DeepMind, the company’s objective is to use the approach to employ “conversation in the long run for safety.” Douwe Kiela, who works as a researcher at the artificial intelligence startup Hugging Face, describes Sparrow as “a nice next step that follows a general trend in AI, where we are more seriously trying to improve the safety aspects of large-language-model deployments.” Sparrow was developed by Hugging Face.
