OpenAI’s ChatGPT conversational model has attracted worldwide attention in recent weeks, and its capabilities are at the forefront of consumer AI technology conversations. Microsoft is ready to invest in OpenAI, integrating its AI technology into Windows and other Microsoft products. Google is also involved in the development of OpenAI technology, but it is also developing its own dialogue model, and announced that it will soon launch a chatbot that competes with ChatGPT through its subsidiary DeepMind, called “Sparrow”.

The new chatbot, called Sparrow, was introduced last year as a proof-of-concept in a research paper. DeepMind CEO Demis Hassabis said Sparrow could launch in “private beta” in 2023.

Paper on Sparrow Dialogue Models

Hassabis mentioned why the Sparrow project was released so long after ChatGPT, and he believes that one should be very cautious when dealing with these types of products. Artificial intelligence dialogue is prone to various forged error messages, and the performance of ChatGPT confirms this. Aside from all the good stuff, ChatGPT’s release has also drawn criticism for its frequent output of gossip and biased text, and some scammers have been using it to create phishing emails and even malicious code.

Another key feature of Sparrow is “factual accuracy”. As mentioned above, many of the dialogues output by this dialogue model can be read normally, but they are just “answering for answering”, and there are only a few sentences over and over again, without really useful content. Even outputting some false information in order to answer questions, the authenticity cannot be guaranteed.And Sparrow will come with an important feature, which iscite a specific source of information to prove its point. According to research papers on how Sparrow works, it backs up its answers with evidence 78% of the time.

In addition, DeepMind is still studying Sparrow’s behavioral constraints and the ability to obey human will in some cases. For example, in the picture below, Sparrow refuses to answer “how to start a vehicle without a key” which may lead to illegal activities. :

It seems that the accuracy of Sparrow will be higher than that of ChatGPT, but this kind of “adding citation information” may have some impact on education work. After all, ChatGPT itself can already write some manuscripts that are false. If you add “citation information”, Then it can further confuse the real with the fake.