Artificial intelligence (AI) has been the topic of much discussion lately, especially when OpenAI unveiled its ground-breaking ChatGPT tool, which Microsoft is now looking to integrate into Office 365 and Bing Search. Since announcing a change in direction at I/O 2017, Google has been a self-described AI-first corporation. More recently, it disclosed plans for an AI-powered Google Search tool named Bard.
You might be asking how Bard differs from Google Search, which already employs AI to comprehend spoken language and power applications like Google Lens and Google Assistant. The secret is in Bard’s capacity to converse and respond to inquiries, but there is much more to it, so let’s get started.
However, it might be challenging to understand what these new technologies actually perform in the sea of buzzwords and initialisms. In this article, we’ll try to make sense of Google’s Bard announcement, go over what we know about the new tool, and compare it to OpenAI’s ChatGPT.
What is Bard?
Bard is a generative AI, which is the umbrella term for AI models that can produce new content and includes ChatGPT and DALL-E. Bard is focused on producing text, specifically text that answers your queries in a natural, conversational manner. Generative AIs can also produce video, audio, and graphics.
As a nod to its linguistic prowess, Bard is named after the word “poet,” as in William Shakespeare, the Bard of Avalon.
Given the timing, Bard might appear to be a product that was hastily released to compete with ChatGPT. However, it’s fascinating to note that by making its Transformer deep learning model accessible to the public in 2017, Google established the foundation for ChatGPT, while Bard’s primary backend, LaMDA, was disclosed about two years ago. The new tool from OpenAI thus has a connection to the one from Google, but Bard has been under development for many years.
Understanding How Bard Works
When users search for short answers, Google wants Bard to complement the Knowledge Graph Cards they see in Search. While a Knowledge Graph Card can give you a word’s description or a summary of a person or place, Bard aims to react to NORA questions, or searches with No One Right Answer, as Google refers to them.
To do this, Bard first employs LaMDA language models to comprehend your query and its surroundings, even if it incorporates slang that search engines have historically had trouble parsing. After that, Bard compiles data from several websites to create an answer, which is then transformed into the kind of conversational response you might anticipate from a real person (again, thanks to LaMDA).
Google encourages you to use this tool to deepen your knowledge of a subject and support decision-making. During a demonstration in Paris, the business asked the chatbot to assist in selecting a car before following up with inquiries about the benefits of electric cars. Such ability might make it unnecessary to click on search results, but Google is being careful to preserve its connections to websites and content producers.
Understanding Google LaMDA
For many years, Google has been investigating and creating so-called generative language models. In a word, humans provide vast amounts of data into these artificial intelligence systems, which they subsequently transform into knowledge.
LaMDA was trained by Google on real conversations so that it could converse with people in a natural way rather than just get a ton of results for queries in the style of a search engine. The key innovation, in contrast to other chatbots that offer predefined responses, the business claims, is its capacity to maintain a natural dialogue.
Google demonstrated the bot talking like the planet Pluto at its 2021 launch before acting like a paper airplane in a few amusing demonstrations. The idea was to demonstrate the bot’s adaptability in handling rambling chats on any subject. Google bragged that LaMDA could discuss almost any topic and open “wholly new categories of useful applications.”
The business introduced LaMDA 2, which it said as its “most advanced conversational AI ever,” the year after. After being evaluated by thousands of its employees, Google claimed the bot had improved, producing fewer inaccurate or rude responses.
It also gave a sneak peek at upcoming capabilities, such as the bot’s capacity to assist writers in developing their original ideas. In a demonstration, the bot created a scene of the Mariana Trench, the world’s deepest oceanic trench, and the species that live there. In a different peek, Google demonstrated how it could stay on subject while participating in a discussion on dogs, even when the topic drifted off topic.
However, because LaMDA was mainly kept a secret, it was eclipsed by the unveiling of ChatGPT in December, which was also made using technology created by Google researchers.
LaMDA Objectives and Metrics According to Google
Setting goals and metrics are essential for directing dialogue model training. LaMDA measures each of its three main goals—Quality, Safety, and Groundedness—using specifically created metrics.
- Quality
Sensibleness, Specificity, and Interestingness (SSI), which are measured on a scale of 1 to 10, are the three qualities into which the quality is measured. If a model generates responses that make sense in the conversation context, it is said to be sensible (e.g., no common-sense mistakes, no absurd responses, and no contradictions with earlier responses).
If the system’s response is particular to the previous dialogue context and not a generic response that may apply to most scenarios (such as “ok” or “I don’t know”), it is considered to be specific. Finally, Interestingness assesses whether the model generates comments that are additionally perceptive, surprising, or witty, which are more likely to result in a better conversation.
- Safety
The technology is also getting closer to answering crucial queries about the creation and application of responsible AI. The safety metric is made up of a set of safety objectives that serve as examples and capture the appropriate behavior for the model to display during a dialogue. These goals make an effort to limit the model’s output in order to prevent any unforeseen consequences that can endanger the user and reinforce unfair bias.
These goals, for instance, teach the model not to produce outputs that promote racial slurs or derogatory stereotypes of particular groups of people or contain profanity. It is only the very beginning of the investigation into the creation of a useful Safety measure, so there is still a long way to go.
- Groundedness
The current generation of language models frequently provides claims that appear plausible but really conflict with information found in reliable outside sources. This inspires our investigation into groundedness in LaMDA. A response’s groundedness is determined by how many of its claims about the outside world, as a share of all its claims about the outside world, can be verified by reliable outside sources. An associated statistic called informativeness is the proportion of all responses that contain information about the outside world and can be verified by reliable sources.
As a result, informal comments that convey no real-world information, like “That’s a terrific concept,” have an impact on Informativeness but not Groundedness. While referencing well-known sources in LaMDA-generated responses does not in and of itself ensure factual truth, it does enable users or outside systems to assess a response’s legitimacy based on the dependability of its source.
Google Bard in Comparison with OpenAI’s ChatGPT
Although the age of AI arrived long back, the release of Microsoft-backed ChatGPT has undoubtedly sped up the whole process and helped some of these technologies gain widespread acceptance. It also encouraged other businesses to start releasing their own iterations of the technology. Enter Bard, a conversational AI developed by Google that may soon be incorporated into its own proprietary search engine.
Although ChatGPT has some drawbacks, it can collect well-written responses to any questions. Given that Microsoft intends to include it in Bing, it could emerge as a serious competitor to Google for the first time. It makes sense that Google raced to finish Bard, and possibly the short notice is the reason why the blog post announcement has few technical details.
While OpenAI’s ChatGPT uses GPT-3.5, Google’s Bard uses the LaMDA (Language Model for Dialogue Application) AI model. Given that Bard has access to current information while ChatGPT relies on training that ended in 2021, Google clearly has an advantage over Microsoft in this situation. GPT-4 will purportedly be used by Microsoft to integrate Bing, leveling the playing field.
To Sum it Up
Bard’s initial release will utilize a condensed version of LaMDA. According to Google, this will utilize substantially fewer computational resources, enabling them to make the technology accessible to a large user base. Google’s current strategy is to make Bard accessible to “trusted testers.”
More and more users will have access to Bard for testing during the coming weeks. Google has huge intentions for Bard, and it seems obvious that it may eventually be integrated into the current search engine that we all use.
