Since November of 2022, the world feels to have been thrown several sci-fi films into the reality of our future. The wide-release of ChatGPT initiated many questions around AI technology. How do we make sure it is safe? How can we use it to better our world? How do we control it? And many governments are asking the same one question: How can we make our own?
In February of 2023, one week after China announced work on Baidu’s Ernie Bot, Taiwan announced its goals in the development of Taide, the country’s own efforts to capture some of the benefits and counter the consequences of the technology. Taide was released last week on April 29.
The integration of large language models (LLMs) offer massive gains in economic and political power, instantly making their place in geopolitics and the international tech race critical. Many governments, including Taiwan’s, are motivated by economic growth and security to develop something that gives them a little more control of the industry within their own markets. Taide, Taiwan’s own domestic LLM, will allow the country to capture some of the economic benefits of the industry while protecting its population against foreign influence.
LLMs in Geopolitics
Large Language Models mimic the neural pathways found in the human brain. Developers feed the model relevant information–or data parameters–to train the system to respond to queries and questions as appropriately as possible. Within the constraints of the training data, LLMs have the power to do anything from answering questions on current events to translating documents and performing copywriting and code generation tasks. The power of these softwares makes the industry extremely lucrative. Private individuals may regularly turn to ChatGPT or ErnieBot to answer a question on current events or assist in writing an email. In addition, many organizations and private firms contract the technology to build into their business models, using the software to counter supply chain obstacles and simplify customer service operations.
Because of its innovative capacity, the LLM industry is estimated to add up to $4.4T annually to the global economy, making it a key driver in international economic and geopolitical competition as countries race to develop their own models before barriers to market entry become too high. However, the technology has been flagged as an area of international concern because of its ability to covertly shape perceptions and limit access to certain types of information, including information censored by governments. Depending on the data an LLM is trained on, it will reflect whatever restrictions of information or nuanced biases in its output present within its training parameters. Because of the natural biases within all of us, including the developers of LLMs, it is extremely difficult to create a model that reflects perfect objectiveness in every single output. However, an even more onerous task for a country’s population would be to counter LLMs that are purposely trained to discreetly shape attitudes and perspectives that are not in that country’s interests.
Taiwan & LLMs
After only 18 months into the generative AI revolution, it is still very difficult to pin down the long-term effects of LLM technology on geopolitics. Chatbots that cheerily field productivity-draining customer inquiries on retail websites are built on the same systems that have the power to enable mass disinformation campaigns, and the ability to generate code to maximize supply chain operations could even be used to test military invasion and defensive strategies (as has been the case with ErnieBot and potentially even ChatGPT in the near-future). Recognizing the benefits of joining this new phase of the tech race–and the risks of being left behind–Taiwan announced its intentions to build a domestic LLM only three months after ChatGPT became bookmarked in our browsers.
On April 29th, Taiwan’s National Science and Technology Council released the Trustworthy AI Dialogue Engine, or Taide. This new model will function similar to AI chatbot front-runners, but on a smaller, more compact scale. Meant to serve the Taiwanese domestic population almost exclusively, Taide is trained on domestic data, with the majority of parameters consisting of news publications and publicly available government reports that are written in traditional Chinese characters (as opposed to simplified characters used in China) and are infused with local grammar and colloquialisms. Because of these parameters, Taide’s outputs will better align with Taiwanese language patterns, a complaint of ErnieBot and ChatGPT’s English-to-Mandarin translation outputs. Just like with ChatGPT and ErnieBot, private individuals would be able to turn to Taide for help with menial tasks while corporations and other organizations could add it into their digital operations to improve efficiency. $7.4 million USD has been budgeted for development, and the technology has been rolled out in April of 2024.
Taiwanese technology specialists are optimistic about the benefits offered by Taide. Taiwan’s export-driven economy is made-up by critical firms in technology, IT, manufacturing, and logistics industries–sectors where LLMs are in high-demand for its software generating, debugging, inventory forecasting, and demand prediction capacity. By offering a domestic model, LLM consumption could stay in-country, capturing more of the economic value of production and consumption and boosting the country’s GDP.
Taide is also predicted to offer many security-related benefits. Over the last few years, Taiwanese officials have expressed concern over the proliferation of TikTok–also known by its Chinese-language name, Douyin (抖音)–on Taiwanese phones. The app has been accused of gathering user data and distorting information in ways that align with PRC interests. At the end of last year, Taiwan’s Information Environment Research Center (IORG) found Taiwanese tiktok users were 10% more likely to agree with arguments skeptical of the US while leaning toward China’s political positions.
Similar to social networking platforms, AI chatbots have a pronounced ability to infiltrate culture and social norms through a restriction of content and nuanced messaging. Because the model is trained on censored and biased information, ErnieBot redirects questions on taboo topics like Tiananmen Square, and proclaims Taiwan to be part of China. Even ChatGPT is questionable, as it pulls some of its Chinese-language outputs from PRC-published outlets and sources. When asked on the topic, Yuh-Jye Lee, principal investigator with the Taiwan Information Security Center, warned that “if Taiwan’s youth are going to use ERNIE as much as they are using Douyin, we are going to have a serious problem.” A domestic model would also prevent the risk of the gathering of Taiwanese user data by foreign companies with links to nefarious actors.
Conclusion
Taiwan is already the world’s leading AI chip maker, with demand for this specialized hardware now being the biggest growth-driver for the country’s highly-geopoliticized semiconductor industry. Given Taiwan’s skilled workforce and high-tech research & development infrastructure, government officials, as well as foreign governments, are optimistic in the country’s ability to develop an LLM that grows its economy and protects its users against foreign influence.
However, some Taiwanese specialists are concerned whether the benefits of its development will outweigh the costs. Taide’s seven billion data parameters are a tiny fraction of ChatGPT’s 175 billion, making its operating capacity significantly less powerful. In addition, the computing power currently available to the project is limited, a factor that could significantly hinder the adoption and scaling up of Taide. Overall, the project needs more funding. One veteran AI expert argued that Taide would need closer to NT$3 billion (US$100 million) to create a more comprehensive model that would be widely used.
Despite these facts, many working on Taide are quick to state that the software should not be considered as a “Taiwanese version of ChatGPT”, and it does not necessarily need to match its capabilities for Taide to be valuable. Lee of IORG, who is also one of many contributors working on the project, argues that “decent is good enough.” Investments in Taide have already yielded massive geopolitical benefits in the form of a Science and Technology Cooperation Convention between Taiwan and France, signed in November of 2023, that features AI security as one of its six areas of focus.
In the short-term, it is unlikely Taide will match the strength of its competitors. However, the software moves Taiwan forward on its innovative path. Along with potential economic and security benefits, Taide signals Taiwan’s commitment to strengthening its digital democracy. In the long-term, we can expect AI development to play a critical role in the country’s geopolitical positioning as Taiwan continues to show the world how innovation and democratic values can work together in furthering a free and open world order.
(Photo by Maxim Hopman on Unsplash)
