Regarding the official announcement during Google I/O 2023 this yearArtificial intelligence service "Bard"Today (8/16), Google DeepMind scientist Dr. Huaixin Ji, who participated in the development of "Bard", explained in Taiwan how Google created "Bard".
While Google has long invested in AI development, including the familiar Google Assistant digital assistant, services like Google Search, Google Maps, YouTube, Gmail, and Google Translate have long been powered by AI. Even the cameras on Pixel phones use AI to achieve incredible nighttime photography.
However, the actual operating model behind "Bard" is somewhat different, even significantly different, from the AI already used in various Google services. This is because the AI used in many Google services often uses deep learning from massive amounts of data to make optimal decisions based on different situations. For example, it can determine whether a user frequently opens an app or uses a feature, or provide the most appropriate answer based on a keyword. However, the interaction that "Bard" aims to achieve is more like human interaction, not only providing the best answer at the moment, but also adjusting it based on subsequent learning.
The actual difference is more similar to Google Assistant's predecessor, Google Now, which primarily used voice recognition to convert spoken content into text for search, then displayed the answer or search results on a webpage. However, the system itself may not actually "understand" the user's search terms, simply comparing data and selecting content that is closer to the keywords. However, subsequent updates of Google Assistant began to integrate large-scale natural language models and, through learning mechanisms, began to "understand" the user's actual needs, and compared the results to more relevant answers.
For users, it can be difficult to discern the differences in the actual operating modes. Dr. Ji Huaixin explained that Bard's operating principle isn't simply training the English version and then converting the user's language through a mechanism similar to Google Translate. Rather, it's about enabling Bard to truly understand specific languages and, in turn, understand the user's actual needs. Therefore, adapting to different languages isn't simply a matter of connecting translation mechanisms; it requires extensive language training.
Although "Bard" initially only supports English interactions, by connecting to the Transformer natural language model developed by Google in the past and converting through an intermediate language, it can quickly learn languages including Chinese through deep learning and understand the meaning expressed in these languages.
Bard initially integrated the large-scale natural language processing (NLP) framework, LaMDA, which Google announced in 2021 and boasts 1370 billion parameters. Accelerated training using its own TPUs allows Bard to understand reasonableness, specificity, fun, safety, authenticity, and integrated, multi-step Q&A, even allowing for fine-tuning. At Google I/O 2023, Google announced the integration of the new PaLM 2 large-scale natural language model into Bard, enabling more natural communication and interaction, and providing more appropriate responses.
Compared to the early Google Assistant, which could only provide single, fixed answers and even seemed stiff and unnatural during the communication process, Google hopes to achieve more natural interaction and a true understanding of user needs in "Bard". Even if users interact in different languages, it can still provide reasonable and correct answers, and can even adjust and correct the answers based on subsequent learning.
However, Dr. Ji Huaixin also emphasized that in real life, the "answers" to many things are actually multifaceted and may even vary depending on the way users ask questions. In addition, the operation behind artificial intelligence still relies on a large amount of learning data and even involves more complex issues such as privacy and ethics. Therefore, it has become a difficult problem that current artificial intelligence cannot avoid and must face.
Therefore, Google is currently still providing the "Bard" service in beta form. In addition to providing the best answer, it will also provide different candidate answers. The purpose is to fine-tune it through user feedback so that artificial intelligence can make more accurate decisions. In addition, the source of the quotation will be added to the answer content to avoid bias in the answer.
Dr. Ji Huaixin stated that Google will invest in the research and development of artificial intelligence technology with a more responsible attitude, and "Bard" is only one of the current development projects. The future goal is to allow artificial intelligence to penetrate more life scenarios, but it will also be cautious to avoid causing discrimination, injustice, or affecting the privacy and security of users, and to enable artificial intelligence to accelerate the growth of more industries.
