发展本土AI基础模型能力

14 Mr Gerald Giam Yean Song asked the Prime Minister (a) whether there exists a whole-of-Government plan to develop indigenous capabilities in creating and deploying artificial intelligence (AI) foundation models, including establishing a National Foundation Model Research Institute, so as to help local researchers build core AI foundation model expertise; and (b) if such an AI foundation model plan already exists, what are its core and strategic aims.

The Senior Minister of State for Communications and Information (Dr Janil Puthucheary) (for the Prime Minister) : Sir, Singapore's research capabilities in artificial intelligence (AI) are well-regarded.

We generally rank among the top 10 countries based on publications at leading AI conferences across areas, such as natural language processing and computer vision.

We also have an active AI industry research scene, with industry research and development (R&D) centres by companies, such as Salesforce, SAP and Alibaba.

As such, Singapore has a strong base of AI researchers who are already working on various aspects of foundation models (FMs), such as helping FMs better understand visual and audio inputs, evaluating the performance of FMs and developing FMs that are more suited for regional use cases.

It is not usually necessary or beneficial to rely solely on indigenous capabilities to advance our economic interests. For example, Singapore has built up thriving aerospace manufacturing and aviation services without developing or owning aircraft technologies.

That said, we recognise the value of investing in research capabilities to support talent development and enterprise innovation. We have a vibrant research ecosystem for AI and will continue to support useful endeavours, including various aspects of FMs.

Mr Speaker : Mr Gerald Giam.

Mr Gerald Giam Yean Song (Aljunied) : I thank the Senior Minister of State for his reply. Could I just clarify then that there is no whole-of-Government plan to develop indigenous AI FMs in Singapore? I am asking this because many other countries, including the United Kingdom, Japan, Korea, even Saudi Arabia and the UAE, are committing to ambitious programmes to develop their own indigenous FMs.

On the other hand, Deputy Prime Minister Lawrence Wong said last month, during his trip to the United States that Singapore is not at the cutting edge of ITS. The latest ideas and innovations will still come from American universities and companies, but that we can be a fast adopter. So, can I clarify that the approach for Singapore is that we are going to be a fast adopter and not going to lead in the building of FMs?

Secondly, does the Government see any long-term risks in outsourcing the development of FM research and development?

And lastly, when using the large language models (LLMs) developed by foreign companies, how does it ensure that the confidential data is not disclosed?

Dr Janil Puthucheary : Sir, I am mindful of time. On the last question that Mr Gerald Giam has asked, it is large area about how we protect the privacy of our systems and the management of data around AI. I think it is a whole area. I would encourage them to file a question on that so that we could have a discussion.

Let me try and address his first two questions.

I would say though that I would push back a little and disagree with his characterisations that we are either going to outsource or we are going to develop capabilities locally. And, in truth, you have to do both and even large continent-sized countries are doing both. We do recognise the potential of FMs to significantly enhance productivity to enable new services. And most use cases in the world leverage on existing FMs, either off-the-shelf, customise via methods, such as retrieval augmentation of finetuning.

We do have groups that are experimenting and building FMs on such issues, such as to better translate regional languages, improving our understanding on how to govern FMs.

But we do so in a prudent and phased manner. We are starting with smaller size models. We have received industry interest to collaborate. These efforts are relatively nascent and we will be examining our plans on how to scale up further with details to be announced later this year.

We are supporting an initiative to build our own family of FMs through AI Singapore, named SEA-LION, or Southeast Asian languages In One Network, and training to better address regional use cases, which is a gap in existing FMs.

So, we are building these models in a phased manner. We have developed two small ones so far. They have been published. They are open for public and research use and we have received industry interest.

The Government also has been proactively looking at how to develop FMs to improve our own productivity.

My key point then, Sir, is that I would characterise the issue not as either indigenous capabilities or outsourcing but, in truth, the appropriate combination of both and we have already begun on this journey.

12.02 pm

Mr Speaker : Order. End of Question Time. Introduction of Government Bills. Minister for Home Affairs.

[ Pursuant to Standing Order No 22(3), provided that Members had not asked for questions standing in their names to be postponed to a later Sitting day or withdrawn, written answers to questions not reached by the end of Question Time are reproduced in the Appendix. ]