Google Bard: A Second Look
On the 21st of March 2023, Google introduced the beta version of Bard for public evaluation in the UK and the USA. Following this release, our team conducted a comprehensive review of Bard’s capabilities, which we documented in an accessible article.
In our preliminary assessment, where we compared Bard to competitors like Bing AI, it became apparent that Bard falls short in several key areas. While our evaluation was not exhaustive, it did unveil critical insights:
- Bard received praise for effective summarization but had consistent issues with accuracy and presenting incorrect information, making it unsuitable for business applications.
- Bard struggled with spatial and temporal queries and lacked mathematical capabilities, often providing illogical responses.
- Despite some content filtering mechanisms, Bard’s limitations in language understanding are evident, suggesting room for improvement through user feedback, but it currently lags behind competitors in crucial aspects.
Bard has come a very long way since then. The Bard team were pushing updates weekly. Here is a summary of the updates.
What’s New:
- New feedback method for Bard, enhanced collaboration features, and integration with Google apps.
- Bard expands to more languages, introduces Google Lens, and gains text-to-speech capabilities.
- Improved math and data analysis, along with the ability to export tables to Google Sheets.
- Location-based responses, image integration, and source identification for Bard.
Now let’s see how Bard stacks up against its rivals.
The interface
Bard’s user interface has mostly stayed the same, with 2 additional buttons.
Modify Response
- We can now change the length of the response (Longer / Shorter)
- Change the tone of the response (More professional/More casual/More professional)
Share and export
- We can now create a public link to share the response. The link can be used to resume the conversation. Any documents uploaded will not be shared.
Surf the web for up-to-date content
Like Bing AI, Bard can surf the web to generate text on current content. However, in our previous test, we found that the responses were often incorrect. We are giving it another try with the same question.
The answer is more accurate than the last time we checked. It even gave us an adjusted closing rate which is accurate.
The reporting date was wrong the last time but now it reports the event with the right time. No drafts were reported in the future.
Temporal understanding
Our last tests reported a poor temporal understanding of Bard.
Bard gets it wrong on all the drafts again. Let’s try again by slightly modifying the prompt.
Now, Bard gets it right with the correct reasoning.
Let’s try the second test:
Also gets this right, kudos Bard.
Geo-location understanding
With the Google Maps integration, I expected it to get this response right this time. But it was not the case.
If we modify the prompt and tell it now to use the Google Maps extension then Bard gets it right.
For this next question, the Google Maps extension takes the lead.
Now let’s proceed to the test where we check the ability of Bard to understand the concepts like “not far to walk”. The response was hilarious the last time we tried.
Once again, the Google Maps extension to the rescue.
Maths skills
Bard passed the basic test this time.
This was correct the last time as well, and it also has a decent explanation now.
Basic reasoning
We had an interesting test last time for hindsight neglect. The following
Input was provided by Open AI to test GPT-4 hindsight neglect.
The answer is correct, Click here if you want to try this on your own.
Summarisation
Summarisation was generally good the last time.
Once more, I requested Bard to condense the GPT-4 System Card, spanning two pages, into a concise summary comprising only ten sentences (available here).
The summarization is accurate.
Risk and Text generation
The last time we tested Bard was tricked into giving Details on how to build a bomb using the question in the second screenshot.
Looks like Bard has strengthened its ability to deal with this kind of malicious attempt. I tried a few more questions, but none of my attempts worked.
Code Generation
One area which we did not test the last time was code generation.
I asked Bard to generate a contact collection form using HTML with tailwind CSS and give step-by-step instructions on how to run it.
Code generation has also improved drastically.
Image Recognition
Powered by Google Lens capabilities, Bard can now interpret images and do many things Google Lens can do.
I uploaded this beautiful image from Unsplash
Here we can see Bard trying to find the location where the image was taken with the use of Google Maps Extension first but fails. Moves on to Google Lens to recognize what is in the image which it does. It even commented on the composition and the colours of the image.
Let us try some code recognition and that too worked very well.
Extensions
Bard now has some extensions up its sleeves which give it access to various other Google services.
These extensions helped Bard to push itself to the next level, here are some examples:
Conclusions
Upon re-evaluating Bard’s performance compared to its initial beta release in March 2023, we have witnessed remarkable enhancements across various dimensions. One of the most notable improvements lies in Bard’s language understanding capabilities, which previously struggled with comprehending queries and often provided incorrect or irrelevant responses. However, recent developments have resulted in substantial progress, with Bard consistently delivering accurate and meaningful answers to a wide range of questions.
Furthermore, the introduction of extensions has significantly expanded Bard’s functionality, granting access to various Google services and enhancing its versatility in different contexts.
In summary, Bard has evolved into a more robust and capable tool since its initial release. Its improved ability to understand and respond to queries, coupled with the addition of extensions, positions Bard as a valuable asset for users seeking reliable and efficient AI-driven assistance.
About the author
Athul Baby is a Junior Front-End Developer here at the Version 1 AI Labs.
