In a bid to strengthen its position in the increasingly competitive landscape of AI, Google has announced significant updates to its Bard AI model, introducing new features and capabilities. These updates come as Google faces off against Microsoft, which has its own partnership with OpenAI, a key player in the AI domain.
Bard extensions: Empowering users
One of the standout features unveiled by Google is Bard Extensions, designed exclusively for paid users of Google products like Gmail, Docs, and Sheets. These Extensions offer users enhanced integrations with Bard, a formidable competitor to OpenAI’s ChatGPT. The Extensions serve as a powerful toolset, allowing users to streamline their tasks and interactions with Bard.
Within Bard Extensions, a prominent addition is the “Double-Check” feature. This innovation aims to address the longstanding issue of inaccurate information generated by large language models. Double-Check evaluates each sentence provided by Bard and cross-references it with internet citations. Users will see sentences color-coded as green or orange. Green indicates corroborated information supported by credible sources, while orange signals information without corroboration, prompting users to exercise caution.
Sissie Hsiao, Vice President at Google and General Manager for Google Assistant and Bard, shared a practical example of how Bard Extensions can revolutionize daily tasks. She described how a parent dealing with numerous school-related emails could use Bard to summarize the information contained within these emails quickly and efficiently. This demonstrates the potential time-saving and organizational benefits that Bard Extensions offer.
Bard and Google search: A symbiotic relationship
Despite Google’s reputation as a search giant, Hsiao emphasized that Bard and traditional search are complementary rather than contradictory. While traditional Google Search helps users find information across the web, Bard’s Extensions enable users to search their private data corpus effectively. This ability to navigate and retrieve information from personal documents and emails provides a unique and valuable dimension to AI-powered tools.
Regarding Bard’s popularity and adoption, Hsiao didn’t reveal specific numbers but highlighted the substantial momentum Bard has gained. Users are actively exploring and embracing the new features, uncovering novel use cases that continue to expand Bard’s utility.
Bard’s evolution: A constant experiment
Hsiao described Bard as an ongoing experiment in generative AI. Google remains committed to pushing the boundaries of what generative AI can achieve. The introduction of Bard Extensions represents a significant leap forward, enabling Bard to interact seamlessly with other tools—a concept referred to as “agentive capabilities.”
As Bard evolves, Google envisions it transforming into a personalized AI agent capable of handling complex tasks. Hsiao described a future where Bard can not only summarize emails but also place orders on behalf of users or create schedules by sifting through vast amounts of information. This vision underscores the idea of AI as a valuable assistant that empowers individuals to accomplish higher-order tasks through natural language interactions.
When asked about the possibility of Bard booking summer camps or similar tasks, Hsiao affirmed that such capabilities are on the horizon. While the transactional aspect might be relatively straightforward, the real value lies in Bard’s ability to process and organize information efficiently. This capability can significantly enhance productivity in handling various information sources, including emails and texts.
Bard as a unifying AI agent
Hsiao emphasized that Bard’s potential extends beyond Google’s ecosystem. Bard aims to be a unifying agent that bridges multiple tools and services from various companies. Users interact with a multitude of platforms, and Bard’s ambition is to seamlessly integrate and streamline these interactions for a personalized AI experience.
Google also announced upgrades to the underlying AI model powering Bard, known as PaLM2. While specific details about model sizes were not disclosed, the company emphasized that the latest launch significantly improves the quality of PaLM2.
Google employs fine-tuning and reinforcement learning with human feedback (RLHF) to enhance the model’s performance. Hsiao likened this process to a culinary recipe, where adjustments are made to achieve the desired taste. Fine-tuning aims to minimize instances of model “hallucinations,” ensuring that it generates accurate and contextually relevant responses.
Regarding the costs associated with running larger and more sophisticated models like PaLM2, Hsiao indicated that Google constantly balances costs and latency. The focus is on optimizing both the cost-effectiveness and speed of these models, with ongoing refinements and optimizations.
The awaited Gemini model
Intriguingly, Hsiao hinted at a forthcoming model called Gemini, which will be integrated into Bard. While specific details and release dates were not provided, Gemini represents a substantial advancement. It’s described as a new model architecture with innovative training methods and a considerable size. Google aims to leverage this technology to introduce new capabilities and skills within Bard, further enhancing its versatility.
