i卡會員

The launch of ChatGPT at the end of last year has allowed people of different nationalities to easily use generative AI in their life and at work. The number of ChatGPT users reached the one million mark within five days of the registration opening. The growth of its user base was 15 times faster than that of Instagram and 30 times quicker than Spotify. The total number of users exceeded 100 million within two months. The number of daily active users was approximately 13 million. The new upgraded version, ChatGPT Plus, available through a paid subscription plan even stopped selling temporarily due to excessive demand.

CB Insights, the technology research company focusing on startups, released its “11 Tech Trends To Watch” this February, in which it mentioned that generative AI would be an important topic this year. In fact, global venture capital investment in generative AI has been rising since 2017, in both case numbers and amounts, and reached a peak last year. Visual media, generative interfaces, texts, voice and programming are the five hot fields.

August of 2022 was an important turning point. Stable Diffusion launched by the startup Stability AI attracted millions of users and became a hot topic online putting pressure on Google, Meta and OpenAI. OpenAI launched its Dall-E 2 in September, and ChatGPT was opened to the public in November. Google introduced its chatbot Bard in February this year. Since then, many startups have developed generative AI tools including C3.ai, BigBear.ai and SoundHound AI, and their share prices also have been raised high. The Chinese search engine giant, Baidu, also launched its ERNIE Bot chatbot in March.

In Taiwan, the programming, copywriting, idea generation, art teaching and animation industries have all been benefiting from using ChatGPT for human-machine cooperation at work. The outstanding performance of generative AI in dialogue and answering questions has resulted in its introduction by companies in the fields of finance, telecommunication and customer service. For example, E.SUN Financial Holding uses ChatGPT as a replacement for its original chatbot for customer service and for KYC (Know Your Customer) due diligence which can significantly shorten the lead-time for customer account opening. Meanwhile, telecommunication firms are working with Taiwan Web Service Corporation (TWSC), an ASUS company, to deploy their own customer service chatbots using in-house data and open-source models. Some companies hope to directly master the field of optimized large language models to avoid monopoly domination by foreign players. MediaTek, Academia Sinica and the National Academy for Educational Research (NAER) are collaborating in the retraining and optimization of the open-source model BLOOM by using traditional Chinese corpuses. The model is now available for download and can be used for Q&A, text editing, advertisement copywriting and customer service applications.   

Due to a lack of data volume, computing power, and financial resources required for the development of AI foundation models, it is difficult for Taiwan to compete with the international heavyweights. However, domestic companies at least have suitable options for foundation models and customized models.

Taiwanese companies can directly access the APIs of Microsoft or Google to introduce generative AI that is immediately useful and low cost. However, this approach does not allow for easy customization and can only be utilized on the cloud. Data stored on the cloud is exposed to the risk of potential service disruptions in the future. Another approach for Taiwanese companies is to use cloud-based training tools offered by the global players and then adjust the large language models by uploading their own data. Whilst this meets the needs of specific domains, there is still the risk of service disruption. Alternatively, Taiwanese companies could use open-source models combined with their own in-house data to develop models to suit their own applications. This facilitates the safest and most accurate services. Whilst it requires multiple rounds of training and significant computing power, it ensures the ownership of the core technology. The model can also be deployed solely within companies to avoid the leakage of confidential information and the problems associated with information security.

Given its characteristics, generative AI can be utilized in many other use cases in Taiwan in addition to finance, telecommunication and customer care. These include medical service, healthcare, mental counseling, legal services, tutoring and educational institutions, manufacturing, IC design and software, e-commerce, 3D design, animation and music creation, new drug and new material development, etc. 

It will not be easy for latecomers in Taiwan to compete with the global players in the field of general-purpose generative AI if the number of model parameters is large and large data volumes, computing power and financial resources are required. However, it is not without possibilities. Based on research literature, practical experience, computing power and annotated data volumes, a generative AI model with 1B to 20B parameters would be quite competitive for a specific domain. This is a feasible technical range for domestic companies to consider.

One example is that of BioMedLM which was jointly developed by the Center for Research on Foundation Models (CRFM) of Stanford University and MosaicML specifically for medical Q&A tasks. It can summarize a patient's needs by generating succinct and relevant questions. It is a relatively small model based on GPT-2 and trained on PubMed abstracts and full texts from the National Center for Biotechnology Information, National Library of Medicine website operated by the National Institutes of Health (NIH) of the U.S. The model achieved a score of 50.3% when tested on MedQA biomedical question answering tasks (US Medical Licensing Exam-style questions) and outperformed many larger models. However, BioMedLM only has 2.7B parameters, which is small in comparison to most general-purpose generative AI models.

This indicates that small models that are retrained with data from a specific domain can perform as well as or better than large models. However, the required data volumes, computing power and cost would be significantly less. Other well-performing generative AI models such as CodeGen for code generation and DALL-E and Diffusion Model for image generation all fall into this category. Certain industries in Taiwan e.g., high-tech design or manufacturing will not want to retain confidential data on the cloud in order to protect business secrets and information security, so they will not consider cloud-based solutions offered by international heavyweights. This local demand for generative AI in specific domains may be an opportunity for solutions offered by domestic companies.