Google chose the name 'Bard' as a marketing strategy, as there are no known algorithms with that name. However, we know that LaMDA is the driving force behind this innovative chatbot.
Learn everything that has been discovered so far about Bard and review interesting research that can give us insight into the algorithms that drive this hardware.
Stay tuned to learn more about what Google Bard has to offer and how it can change the game in the field of artificial intelligence.
Understanding Google Bard
Furthermore, the google bard facilitates the exploration of subjects, summarizing information found on the internet and offering links to websites with additional information. Throughout this article, we will dive into the details of this innovative chatbot and learn more about what Google Bard can do for you.
The Google Bard saga: overcoming challenges and reinventing yourself in the world of artificial intelligence
ChatGPT has gained industry attention for being a innovative technology, with the potential to transform the search industry and shift power dynamics, affecting both Google search and the profitable search advertising segment. Faced with this scenario, Google realized it needed to act quickly.
On December 21, 2022, the New York Times reported that Google activated a "red alert" to define an agile response to the threat posed to its business model. After 47 days of efforts focused on the new strategy, Google announced, on February 6, 2023, the launch of the long-awaited Bard, demonstrating its ability to adapt and innovate in the face of market challenges.
The Google Bard Stumble: What Went Wrong with the AI Chatbot Launch?
The inaccuracy in Google's artificial intelligence meant that Bard's presentation did not have the expected impact, raising doubts about the tool's effectiveness and reliability. This incident showed that even with significant advances in AI, there is still room for improvement and development.
As a result of this episode, Google's shares suffered a significant drop, losing one hundred billion dollars in market value in a single day. This setback reflects a loss of confidence in Google's ability to meet the challenges of era of artificial intelligence approaching.
Google Bard: How does this new AI tool work?
Two important features of LaMDA training are safety and foundation. Security is achieved by tuning the model with data annotated by external collaborators. The justification occurs when the LaMDA consults external sources of knowledge, such as information retrieval systems.
The research article highlights that this allows the model to generate answers based on known sources, rather than just plausible answers. Google used three metrics to evaluate LaMDA outputs:
- Common sense (if an answer makes sense)
- Specificity (how specific or generic the answer is)
- Interest (whether the answers are insightful or arouse curiosity).
These metrics were judged by third-party evaluators and the data obtained was used to improve the system. The research concludes that third-party review and the ability to verify information with a search engine are useful techniques to improve the model.
Through the use of third-party annotated data and external APIs, Google Bard promises improve the security and substantiation of the responses generated by the model. This paves the way for significant advances in the application of language models to dialogues and other practical applications.
Bard: the evolution of Google search with artificial intelligence
Despite this, the lack of clarity created the perception that Bard would be integrated into the search, which was never the case. According to the disclosure made by Google, the idea is that Bard is a feature that helps filter complex information and diverse perspectives into easy-to-understand formats, whether to seek different points of view or delve deeper into related topics.
It is clear, therefore, that Bard is not the search itself, but rather a functionality within it, which is not intended to replace search, but to make it more efficient and easier to use for users.
What is a search function?
On its "How Search Works" page, Google explains that these features are designed to provide accurate, relevant information in the format most useful for your query, which could be a web page or real-world information like a map or stock at a local store.
However, in an internal Google meeting reported by CNBC, employees questioned the use of Bard in search. Some have argued that large language models such as ChatGPT and Bard are not reliable sources of factual information.
The response from Jack Krawczyk, product lead at Bard, was clear:
“I just want to make it very clear: Bard is not quest.”
Google's Vice President of Search Engineering, Elizabeth Reid, reiterated that Bard is a separate search tool. In short, Bard is not a new version of Google search. It is a functionality that can be used to improve the user experience when searching for information.
Bard: An Interactive Way to Explore Topics
Learning about a topic can take a lot of effort to figure out what you really need to know, and people often want to explore a wide range of opinions or perspectives. That's where Bard comes in.
Recently announced by Google, Bard is a interactive tool designed to help users investigate knowledge about topics. Unlike conventional search, which provides links to answers, Bard helps users explore and delve deeper into a specific topic.
Through Bard, users can access a wide range of perspectives and opinions on a topic, making the search for information much more interactive and engaging. With Bard, learning about a topic can be a more enriching and satisfying experience.
Although Bard is not a new version of Google search, it will certainly be a welcome addition for those seeking more complete and in-depth information on a topic. With Bard, knowledge exploration will be interactive and engaging.
How Bard and LaMDA handle factual errors in language models
However, this approach fails in areas where facts are constantly changing, what researchers call the "temporal generalization problem." Up-to-date and accurate information cannot be trained on a static language model.
To solve this problem, LaMDA, and possibly Bard, developed a tool called Toolset (TS). This set of tools includes an information retrieval system, a calculator, and a translator.
The information retrieval system, which is basically a search engine, can return snippets of content from the open web with their respective URLs.
TS tries an input string across all its tools and produces a final list of strings by concatenating the output lists of each tool in the following order: calculator, translator, and information retrieval system.
However, if a tool cannot parse the input, it returns an empty list of results and does not contribute to the final output list. This approach allows the Bard provide fresh, high-quality responses combining knowledge of the world with the power, intelligence and creativity of Google's language models.
How AI research is influencing conversational Q&A systems like Bard
The following algorithms are relevant for artificial intelligence-based question and answer systems. One of the LaMDA authors worked on a project that aims to create training data for a conversational information retrieval system.
The goal was to create a system that could read web pages and predict which questions would be answered by a given passage. For example, an excerpt from a trusted Wikipedia page that says "The sky is blue" could be transformed into the question "What color is the sky?"
Researchers created their own question and answer dataset using Wikipedia and other web pages. They called the datasets WikiDialog and WebDialog.
These new datasets are a thousand times larger than existing datasets, which gives conversational language models an opportunity to learn more.
Researchers reported that this new dataset helped improve conversational question and answer systems by more than 40%. It's hard to imagine a scenario where Google wouldn't train a conversational AI on such a large data set.
However, as Google rarely comments on its underlying technologies in detail, we cannot say for sure whether Bard was trained with the WikiDialog and WebDialog datasets.
Regardless, what we know is that Bard is a tool designed to improve the Google search experience, allowing users to find information more easily and in more useful formats.
New Google Searches: Great Language Models With Attributed Sources
According to the research article, large language models (LLMs) have shown remarkable results with little or no direct supervision, and there is growing evidence that they can be useful in information search scenarios. The authors believe that an LLM's ability to attribute the text it generates is crucial to making it trustworthy in this scenario.
The study authors propose Attributed Question Answering as an important first step in developing LLMs with attributed sources. The idea is to create a language model capable of providing an answer with supporting documentation which, theoretically, guarantees that the answer is based on something.
This technology is specific to question and answer tasks. The goal is to create better answers - something Google certainly wants for Bard.
Attribution allows users and developers to evaluate the "reliability and nuance" of responses. Allows developers to quickly review the quality of responses as sources are provided.
An interesting note is a new technology called AutoAIS which correlates strongly with human raters. In other words, this technology can automate the work of human raters and scale the process of rating answers given by a large language model (like Bard).
According to the researchers, although human assessment is considered the gold standard for evaluating the system, AutoAIS has a strong correlation with human assessment at the system level. This offers promise as a development metric in situations where human assessment is infeasible, as well as being able to be used as a noisy training signal. While this is an experimental technology, it shows that Google is exploring ways to produce high-quality, reliable answers.
Technology that edits responses to ensure factuality in language models
This is where new research from Cornell University comes in, exploring a different way to attribute sources for what a language model generates and even edit a response to correct it.
The proposed system, called RARR (Retrofit Attribution using Research and Revision), is capable of automatically finding the attribution for the output of any text generation model and then post-editing the output to correct unsupported content while preserving maximum the original output.
The research summary states that RARR improves attribution and preserves original input more significantly than previous editing models. Furthermore, its implementation requires only a few training examples, a large standard web search and language model.
This research is important because it helps solve a common problem with language models - the lack of mechanisms for attribution to external evidence. As natural language technology continues to advance, developing new solutions to this problem is critical to improving the accuracy and reliability of language models.
How to get access to Google Bard?
Although it is an experimental tool, the good news is that Google is accepting new users to test Bard. Just access the official page of Bard and fill out the application form.
However, it is important to remember that Bard is not a new version of Google search, as the company stated. He is a functionality that uses AI to provide more accurate and relevant information about a certain subject.
For those working in web publishing or SEO, it is important to understand the limitations and possibilities offered by Bard. Although it is a promising tool, there is still much to be explored in terms of its potential to improve internet search.
In short, if you're curious to try Bard and see how it can help you with your internet searches, now is the time to request access. And who knows, maybe the Bard could become a essential tool for your work in the future.