Google’s Bard generative AI engine can now read emails, summarise documents, and even fact-check itself, the company has announced as a slew of tech giants outline next-generation products that will shape the way consumers and businesses embrace AI.
The release of Bard Extensions marks a significant step forward as Google fights back against the disruption of ChatGPT by extending Bard – which is built on Google’s PaLM 2 AI large language model (LLM) – to integrate data from Google services including Gmail, Docs, Drive, Google Maps, YouTube, and Google Flights and Hotels.
Described as “Bard’s most capable model yet”, the new generative AI engine leverages “state-of-the-art reinforcement learning techniques to train the model to be more intuitive and imaginative,” Bard director of product management Yuri Pinsky explained in showing how the tools can integrate a range of disparate information.
Ask it to check your emails to find out when travel companions are available, for example, and then create an itinerary that includes flight and accommodation details for those dates, and it will search across multiple Google properties and present the answer as a coherent and ongoing conversation.
Significantly, a new ‘Google It’ feature in Bard will force the engine to double-check its outputs – exposing problematic generative AI ‘hallucinations’ by cross-checking its output against Google Search results and highlighting statements that are verifiably true, or potentially inaccurate.
The update – which also strengthens Google’s response to the AI-powered Microsoft 365 Copilot productivity tool that company outlined earlier this year – sees Bard “getting even better at customising its responses so you can easily bring your ideas to life,” Pinsky said.
Yet while PaLM 2 is good at processing domain specific knowledge, Google’s long game is built around Gemini, its upcoming competitor to the GPT-3.5 and GPT-4 LLMs that power OpenAI’s ChatGPT.
Gemini was teased at Google I/O earlier this year with a promise from CEO Sunar Pichai that it would be “fine-tuned and rigorously tested for safety” before being released – a benchmark that had tongues wagging after reports suggested that Google recently gave a few select businesses access to an early version of the system.
Will 2024 be the year of Gen AI everywhere?
For all the work being done to improve generative AI sevices running in the cloud, the efforts of Google and OpenAI are just part of a fast-growing ecosystem that is also working to push the technology out to consumer devices and business desktops.
Amazon, for one, this week released a rash of new and updated consumer devices that bring generative AI capabilities to its popular Alexa assistant – enabling it to converse in a more free-form, human-like way compared to its existing robotic voice.
Even as data centre operators invest billions to add extra capacity to support data-intensive LLMs like ChatGPT and Bard, many vendors envision a future where computationally-intensive AI features run locally on user computers and other devices.
Intel, for one, this week teased ‘AI PCs’ built around upcoming Core Ultra chips that will enable its AI Everywhere strategy by incorporating a Neural Processing Unit (NPU) that accelerates AI data processing to run applications like Microsoft Copilot – which is also due in a Microsoft Dynamics 365 Copilot version for business customer relationship management and ERP applications – on standalone PCs.
Use of the OpenVINO deep-learning model will enable LLMs to run on laptops even when they’re disconnected from the Internet – keeping data private and empowering personalised Gen AI systems that support workers’ everyday activities.
“We see the AI PC as a sea change moment in tech innovation,” Intel CEO Pat Gelsinger said.
Analyst firm Gartner has seen near universal executive support for standalone, customised generative AI capabilities, and the support of major tech vendors will normalise the technology in 2024 and beyond – yet Gartner distinguished VP analyst Arun Chandrasekaran warns business leaders to weigh possible use cases carefully, and to build AI “tiger teams” that actively work with line-of-business executives to identify business processes most ripe for innovation.
Existing generative AI use has followed five major approaches, Chandrasekaran told the firm’s recent IT Symposium/Xpo conference – noting that adding it to existing workflows, like Microsoft Copilot or Google Bard do, is “the least disruptive way to deploy generative AI, but it’s going to be restricted to that app [and] your ability to tune the models.”
Other approaches include ‘prompt engineering’ to steer AI engines towards the desired outcome; “complex and expensive” fine-tuning of LLMs or building them from scratch; and ‘retrieval augmented generation’ (RAG), which combines enterprise data with existing LLMs to “ground these models with a sense of reality”.