Two sandboxes now are available for government agencies and businesses in Singapore to develop and test generative artificial intelligence (AI) applications.
Both sandboxes will run on Google Cloud’s generative AI toolsets, including its Vertex AI platform, low-code developer tools, and graphical processing units (GPUs). Pre-retained generative AI models also will be made available, encompassing Google’s own large language model Palm, its partners’ AI models, as well as open source options.
The initiative is part of a partnership agreement inked earlier this year between the US cloud vendor and the Singapore government to establish an AI Government Cloud Cluster. The cloud platform runs within a dedicated environment on Google Cloud and aims to drive AI adoption in the public sector.
One of the sandboxes will be used exclusively by government agencies, while the other is available to local organizations. Collectively, the two controlled cloud-based environments will be provided at no cost for three months, for up to 100 use cases or organizations.
These will be selected through a series of workshops to be held over 100 days, during which participants will receive training by Google Cloud engineers to identify use cases that can be supported with generative AI and built using toolsets from the sandbox.
The Smart Nation and Digital Government Office (SNDGO) will administer the government sandbox. while Digital Industry Singapore (DISG) will manage the sandbox set aside for local businesses.
Singapore, which introduced its national AI strategy in 2019, has more than 4,000 researchers publishing on AI today, according to Josephine Teo, minister for communications and information. While the research output was “more than respectable”, she noted that Singapore still had “some distance to go” in translating this research into meaningful use cases across different verticals.
“Such efforts are important because they push us to iron out kinks,” Teo said during the launch of the sandboxes. “These kinks could be data issues, but they could also be issues to do with security. Responsible implementation of AI needs us to ensure you do not leave these questions unanswered, but to develop satisfactory answers. This will allow us to unlock the fullest potential of AI for Singapore.”
Google Cloud’s Asia-Pacific vice president Karan Bajwa noted that rolling out generative AI within an organization requires a different approach from general-purpose, public-facing generative AI applications. The former should be deployed with governance and robustness, requiring enterprise-grade generative AI applications to deliver responses based on curated and company-approved data sources.
Data security also is critical for generative AI models used to train enterprise applications, to ensure sensitive information will not be leaked. AI models, too, will need to be calibrated and finetuned for the industry within which an organization operates, such as healthcare or financial services, Bajwa said.
To optimize their cost, companies also will want to choose the right model for the business challenge they are looking to resolve, he said, adding that the larger the model, the more costly its usage.
The sandboxes provide a “clear pathway” for companies to “quickly, easily, and responsibly” build their own generative AI applications with the necessary data security and governance, he said.
Organizations already signed up to participate in the sandbox initiatives include the Ministry of Manpower, Government Technology Agency (GovTech), American Express, PropertyGuru Group, and Tokopedia.
The public sector’s CIO office, GovTech, taps AI and natural language processing engines to power its virtual intelligence chat assistant (Vica) platform. It is looking to leverage generative AI and plans to move all 88 chatbots, currently used within the sector, to generative AI models by year-end.
GovTech already is employing generative AI to operate seven chatbots that are used for two intranets, the Housing Development Board, and Singapore Polytechnic, among others.
According to GovTech, generative AI has cut the number of hours needed to train the model by 10-fold and generated more natural responses. To reduce the risk of hallucinations, responses are generated from data drawn directly from the relevant government agency. Queries about the weather, for instance, are answered using data extracted from the National Environment Agency’s API.
Generative AI models then are used mainly to train the chatbots to better understand questions. Chatbots trained by generative AI have demonstrated an ability to answer 85% to 90% of questions, compared to about 75% for non-generative AI chatbots.
Organizations should make the effort to train AI models to address risks such as hallucinations, said Jimmy Ng, DBS Bank’s CIO and head of group technology and operations.
Large language models learn from any publicly available data, which may not necessarily be quality data. While taking a plug-and-play approach may work for some, businesses need AI applications that have the necessary guardrails and security, Ng said during a panel discussion at the launch.
Organizations then should sharpen large language models with their own knowledge database to improve the quality of the training data, he added.