The following is the transcript of the 6th episode of the Taiwanology podcast. It was produced by CommonWealth Magazine, hosted by Kwangyin Liu, and was first aired Jul. 11, 2023.  

Listen to the episode:【Taiwanology Ep.6】Big tech vs. big gov：How Taiwan uses AI for good

Delving deeper, a Commonwealth Magazine feature unveiled Taiwan's significant role in the AI domain, with a spotlight on the Mediatek Research Lab. This Taiwanese team has embarked on pioneering the world's first large-language model for traditional Chinese, akin to ChatGPT. The question arises: What implications does AI hold for Taiwan's future? As AI's presence in daily life grows, how do we safeguard against the pitfalls of deepfakes and AI-driven distortions? A tantalizing prospect also looms – will all vocalists transform into virtual entities?

To shed light on these matters, we are joined by Ethan Du, founder of Taiwan AI Labs. His distinguished history, including spearheading Microsoft's AI assistant Cortana and contributing to ChatGPT's creation, underscores his insights. Noteworthy is Commonwealth Magazine's collaboration with AI Labs on a deep learning-based podcast voice project, fostering a perfect setting to explore AI's future in Taiwan. With anticipation, we invite Ethan to share his perspectives, revealing a path forward through the intricate terrain of AI.

CW: What the Taiwan AI Labs does and what are your flagship projects? 

So after I left Microsoft in 2017, Taiwan AI Labs is the organization I  founded in Taiwan. That was the very first open AI research institute in Asia. We started as a non-profit, but of course, if we have very good projects like ChatGPT, we can offer a solution for industries too. 

CW: Why did you think Taiwan needs something like AI Labs? 

During my tenure as a principal data manager at Microsoft, overseeing Cortana, I witnessed the impending AI revolution poised to reshape the global landscape. Recognizing distinct patterns and trends, I pondered whether there might be a more crucial avenue of exploration beyond the prevalent approaches in the United States and nations like China and Russia.

This contemplation spurred the inception of Taiwan AI Labs. Drawing parallels to our earlier endeavor with PTT, a pioneering non-profit social network, we prioritized equitable engagement over profit. While American platforms often commercialize through ads, our emphasis in Taiwan remained on fostering connections while championing the equal and unique voice of every individual.

Inevitably, our journey into artificial intelligence acknowledges its potential to emerge as a future superpower.

CW: What did you see when you were heading the team creating Cortana? What could Cortana do that really scared you? 

Both Microsoft and Google stand as responsible entities, with a shared commitment to ethical technology. Yet, their optimization models inevitably gravitate toward profit, given their presence on Wall Street. This compels us to tread cautiously in the realm of artificial intelligence research, prioritizing human-centered values.

Taiwan AI Labs was born from this imperative. Pioneering an open-air research paradigm, we lay emphasis on privacy, algorithmic integrity, and transparency - a novel approach placing human concerns at the forefront. This makes us a trailblazing human-centered research institute, advocating a distinct ethos.

CW: I really remember that you were  one of the first tech people in Taiwan to talk about responsible and ethical AI. 

How do people  have access to Taiwan AI Labs products? What are some of the projects that you have worked on?

Our scope encompasses a diverse array of projects spanning various domains. Within human-computer interaction, our endeavors span traditional language and speech analysis, speaker comprehension, and machine translation. Our focus extends beyond mere conversation to encompass news comprehension, endowing us with a global awareness.

Beyond news analysis, our capabilities extend to identifying news manipulation - an advanced field of study. Further, we draw inspiration from Taiwan's artistic landscape, honing music understanding and even crafting AI singers capable of rendering songs with professional finesse. This enriches the realm of human-computer interaction.

On a distinct trajectory, our medical initiatives encompass medical imagery analysis, such as chest X-rays for COVID diagnosis. Our journey also meanders into genomic research, a path I traversed prior to my Microsoft tenure in 2006 while working at the Human Genome Research Institute under NIH in the United States.

Taiwan AI Lab amalgamates elite researchers from diverse fields. Our collaboration spans premier medical institutions, wherein top researchers collaborate closely. Similarly, the synergy between our top-tier computer science and machinery engineering experts and artists, akin to movie producers, fuels our projects in the realm of human-computer interaction.

CW: AI researchers are voicing concerns about its dangers for humanity. How are these discussions going, and what insights can you provide?

Since the latter part of the previous year, OpenAI's release of ChatGPT, in collaboration with Microsoft where I had a role overseeing Microsoft Cortana's funding, has sparked extensive conversations. The potency of platforms like PTP has been acknowledged for its potential to revolutionize various industries.

Take the news and translation sectors, for instance. ChatGPT's adeptness in generating news pieces and translations has garnered attention, alleviating the need for human intervention. However, a pertinent discourse has arisen concerning potential unconscious biases embedded within AI. As ChatGPT becomes an indispensable tool, there are concerns about cognitive implications. Notably, the model's training data is predominantly sourced from the United States, reflecting an American linguistic flavor that could clash with interests in other regions – illustrated by China's swift decision to develop their own AI language model after banning ChatGPT. This underscores the dual nature of artificial intelligence: it offers unprecedented capabilities while inciting apprehension regarding its potential impact and control.

CW: I think we're likely to see people having their jobs replaced by AI. which led to the next question, what are you going to do with your life when you don't need to work anymore? 

“I would say that we humans will always find a new position.”

When the discourse turns to the possibility of artificial intelligence supplanting humans, my perspective emerges: while AI undoubtedly learns from human expertise, its role lies in augmenting our capabilities rather than usurping them. Humans possess an innate adaptability, continually forging new paths and maintaining a role as overseers of AI endeavors.

We can have a lot of different artificial intelligence to support you.  Then you can just handle what you already know and try to explore what you don't know.  

CW: Regarding Taiwan's perspective, how do models like OpenAI's Chinese variant with mainland Chinese worldviews affect us?

An extensively discussed concern in AI model training revolves around the apprehension of bias. This concern is particularly pronounced in the context of Chinese language data, where a substantial portion originates from simplified Chinese, rather than the traditional form used in Taiwan and Hong Kong. This disparity can inadvertently lead to AI models adopting viewpoints that align more closely with mainland China. Consequently, many Chinese documents and labeled data stem from simplified Chinese sources, shaping the AI model's output in turn.

This issue extends to Taiwan, where even ChatGPT sometimes employs simplified Chinese, thereby reflecting a more mainland-oriented perspective that contrasts with the distinct manner of expression in Taiwan.

CW: Mediatek research lab that is using  traditional Chinese text to train their large language model. So what do you see the potential in that? 

Artificial intelligence learns from human input, including language usage. When data originates in Taiwan, it better suits Taiwanese people; likewise, globally prepared data for Chinese language aligns more with China. Consequently, AI algorithms can favor majority opinions, raising concerns about marginal society impact and reinforcing the need for trustworthy AI algorithms.

Addressing the issue of AI bias and unconscious bias, evaluating their effects is essential. Taiwan has established an infrastructure to safeguard privacy and AI integrity through a proactive approach. This includes a reliable AI program utilizing federated learning and federated validation techniques to ensure privacy and mitigate bias.

CW: What is federated learning?

Data governance is crucial in ensuring the reach and variety of artificial intelligence algorithms. Respecting data ownership is fundamental, and we adopt a respected data governance approach. Unlike centralized models of big tech or government approaches, our strategy embraces distributed ownership. Institutes retain their data, while individuals keep data on their devices, with a shared common format. An open architecture allows AI algorithms to operate on individuals' data without extracting sensitive information.

To maintain privacy, we reduce personal identifying information (PII), keeping data within institutes or devices. AI algorithms learn from this data without extracting it, aggregating experiences into an AI model using federated learning. This approach preserves privacy while achieving AI model enhancement through collective experience, ensuring data anonymity and safeguarding personal information.

CW: I'm thinking of an example, a recent example during the pandemic. I think Taiwan was very  famous for creating the contact tracing system. Is it a good example of using the anonymized data and federated? 

Tu: Yes, that is a very good example. So during the pandemic, we have a  decentralized approach using a federated analytics approach. We know the transmission of the virus without a centralized database. So that will preserve our privacy. So there's no big government  database to trace everyone. You only keep anonymous contact list in your cell phone.  

In addition to the contact tracing app, we harnessed federated learning to develop an innovative chest x-ray classifier, the first of its kind to receive TFDA approval. This classifier discerns COVID presence from chest x-rays, offering a non-invasive alternative to traditional testing methods. The utility of this advanced technology emerged prominently during the early phases of the pandemic, enabling swift and accurate COVID detection without the need for invasive procedures like nasal swabs.

CW: We were just talking about music. I would  like to play a few seconds of the song for our listeners. Take a listen.

CW: Okay, so this is the voice of Taiwanese singer Sandy Chen, 陳珊妮, or was it?  So Ethan, you said you worked together in creating Taiwan's first AI-assisted music single. So what we were hearing was not really Sandy's voice, or was it her voice? And who came  up with the idea? 

Tu: Actually, Sandy came to us and we discussed can artificial intelligence learn  how she sings? So this was a discussion. Then within a couple weeks, we trained our AI algorithm  to learn how Sandy sings. 

CW: How many songs did you use to train the model?  

It's between 40 to 60 songs. We learned how Sandy sings first. And after that,  it was very interesting that Sandy had an idea. She asked can I publish this song?  Therefore, Sandy used the vocal from the AI singer and worked with their team to publish this song  without telling the team member, this song is actually not sung by her.  

CW: Okay, so in her team, how many people knew that was the AI vocal?  

Tu: Only the lyricist-writer and Sandy. After they published this song on KKBOX and Spotify for a week, nobody noticed that.  Everyone just felt, oh, Sandy has a really good song.

CW: The people in her company, were they all assuming that was her voice?  What was their reaction when it was revealed?

Soon after the release, Sandy and I revealed that the song wasn't actually sung by Sandy, but by our AI singer. This revelation surprised many, as they struggled to discern the difference. The project's reception on the international stage has been notable. While traveling in the United States, I've encountered friends who are avid listeners of Chinese music, expressing admiration for the AI singer and acknowledging its future potential. Yet, this excitement is tempered by concerns, including from professional singers who contemplate their role in a landscape where AI can replicate such performances.

This dynamic echoes the past, akin to the MP3 era when professional singers grappled with how to navigate the evolving music industry. Just as they once pondered the fate of their CDs, today's artists confront questions about their place in a realm where AI singers are making waves.

CW: CommonWealth and Taiwan AI Labs cooperated and created an AI  news reading assistant, called “Sky” using Luo Mei, our colleague’s voice.  

And I should say that every time I listen to this,  it gives me goosebumps  because it really does sound like her.  I feel like she was talking to me.

CW: Could you talk about the Sky project  that you have been working on with us?  And how was the voice created? 

Another fascinating endeavor is the "Sky" project, a collaborative effort with CommonWealth Magazine. In this initiative, we engaged with Luo Mei, the vocal presence of Sky. Drawing from her past podcast recordings, we harnessed this wealth of content to train an artificial intelligence agent capable of emulating her speech. 

Consequently, Luo Mei no longer requires a physical presence in the studio for voice recording; podcasts can be produced using AI algorithms, utilizing her existing files. This innovation eliminates the need for new recordings as the AI model can be trained using the available material.

CW: So for people who understand Mandarin Chinese,  go to 闖天下 and find AI 讀新聞.  You will be able to hear Sky.

Frequently, vocal owners initially react defensively to hearing their replicated voices, finding it unsettling that their voice is echoed without their control, a sentiment shared by many.

In contemplating the future, a critical juncture beckons. Will our interaction with AI be illuminated by a promising glow or overshadowed by uncertainty? This pivotal moment necessitates thoughtful consideration, encompassing not just regulations but the methodologies that ensure AI aligns with, rather than contradicts, our human essence.

As evident in projects like Sky and the AI vocal venture with Sandy Chen, we have established mechanisms to uphold ethical AI usage in harmony with the vocal owner's intentions. A profit-sharing model further solidifies this ethical commitment. Now is the opportune time to deliberate on regulations, business models, and the vast landscape of creative and intriguing AI applications. The path ahead beckons us to illuminate AI's potential while upholding human values.

CW: Indeed, the discourse on profit sharing is gaining traction, especially as we witness instances where graphic artists, for instance, protest against their creations being employed to train AI models, subsequently generating new works like Midjourney. Musicians and singers likely grapple with similar concerns. 

In this context, it's reassuring to have entities like Taiwan AI Labs at the forefront, championing a responsible and ethical AI landscape. As we navigate this evolving realm, the potential of AI is immense, yet its impact could be daunting if not harnessed adeptly.

---

Have you read?

• AI medical records assistant Copilot 
• Apps, musicians, language tutors: Taiwan’s AI visionaries 
• The story behind Taiwan’s ChatGPT 

Uploaded by Ian Huang