Chinese pro-test: What is Google Bard like?
Yesterday, Google officially announced the release of its chatbot product Bard, and first launched in the US and UK users, users can apply to join Bard's waiting list .
What can Bard do?
Bard is an experimental conversational AI service powered by Google's LaMDA model that can provide high-quality, up-to-date information to answer users' questions, and Google describes it as a kind of collaborator, creativity and productivity .
Bard can help users solve problems such as:
- Explaining why large language models can go wrong
- Why does lightning strike the same place twice
- Helps users prepare packing lists for weekend fishing, camping, etc.
- Outline art studio logo
- Helping users add high-protein ingredients to vegan recipes and more...
Currently, Bard only supports English, but Google is working on supporting more languages. Additionally, Bard currently does not support encoding capabilities, and therefore does not support handling code-related responses. While Google testers have provided a lot of feedback to improve the quality, safety, and accuracy of Bard's answers, Bard is still experimental and answers may be inaccurate and potentially offensive, so Bard's answers need to be checked. In addition, Google currently limits Bard's contextual dialogue ability , and as Bard continues to learn, its contextual dialogue ability in longer conversations will improve.
How capable is Bard?
Bard's ability has been mentioned in the previous article, and after many tests, it has learned a lot. Next, the key to improving its performance is getting feedback from more people. Unlike ChatGPT, which outputs answers all at once, Bard's ability to generate chunks of text instantaneously is one of its strengths. In addition, Bard also has a feature that it provides three different versions or drafts of answers, and users can choose any one of them . In addition, there is a button on the user interface that says "Google it" that users can use to query the conclusions on the web. The figure below shows three different versions or drafts of answers given by Bard: Bard's abilities have been mentioned in the previous article, and after many tests, it has learned a lot. Next, the key to improving its performance is getting feedback from more people. Unlike ChatGPT, which outputs answers all at once, Bard's ability to generate chunks of text instantaneously is one of its strengths. In addition, Bard also has a feature that it provides three different versions or drafts of answers, and users can choose any one of them. In addition, there is a button on the user interface that says "Google it" that users can use to query the conclusions on the web. The image below shows three different versions, or drafts, of answers from Bard:
Bard can help users provide suggestions when completing tasks. For example, suppose a user wants to read 20 books and asks Bard how to complete this goal in a year. Bard will provide users with some suggestions.
However, Bard still has flaws in the accuracy of his answers and mistakes still occur. Just like Bard's factual error in the debut demo last month, Alphabet's market value fell by $100 billion. The current Bard still has bugs. For example, when asked about the growing conditions of ferns, Bard incorrectly stated that ferns require bright indirect light; when asked to write 4 paragraphs, Bard wrote 9 paragraphs. As shown in the figure below is the Bard with errors:
"We know the limitations of this technology, so we're very cautious about rolling it out," Krawczyk said. From the demo above, we can see that, like ChatGPT and Bing, there is also a prominent disclaimer below the main text box, warning users that "Bard may display inaccurate or objectionable information that does not represent the views of Google ".
First experience of domestic media
Shortly after joining Bard's waitlist, we qualified for a Bard experience.
Obviously, currently Bard does not support Chinese, but it supports voice input. First, let's have Bard introduce himself.
We also compared it with ChatGPT, and the generated results are as follows. Bard provides three drafts for users to choose from.
As can be seen from the dialogue, Bard is a real-time networked AI. We also tested it, and it actually knew that GPT-4 was released on March 14th.
Bard can also solve some simple math problems. But when we try to correct its answer, it ignores our correct answer.
Also, similar to ChatGPT, Bard will also make up nonsensical responses.
Last week, GPT-4 demonstrated multimodal image recognition capabilities and presented the conclusions of the InstructGPT paper. Although Bard doesn't currently recognize images, it can search and rate InstructGPT:
In the GPT-4 demo, people asked AI to explain Figure 2, so can Bard do it? First, Figure 2 looks like this:
Bard's answer follows:
GPT-4 answers as follows:
In the end, we had Bard generate a news story about "Release Bard", which wrote one of its own, and it looked pretty good too.
Overall, it's an interesting exploration in which there are places where people are amazingly capable and where they aren't.