Interviewee / Allan Yiin, Chief Technology Officer
By Cheng-Yuan Wu

    Artificial intelligence (AI) is creating “a more expansive world.” CTO Allan Yiin, a graduate of both Department of Chemistry and Graduate Institute of Journalism at National Taiwan University, excels in science, engineering, programming, and logical reasoning. As a long-time Microsoft consultant, Yiin “used to believe that once you reach a certain age, the things you can do become limited to the scope of your innate talent or professional ability. Now, with only a few prompts, even someone like me, with no music background, can release songs on Spotify!” Yiin’s passion for exploration is just as drastic and intense as his devotion to learning. 

 

Brainstorming with AI on the Road to Creativity

    Shortly after giving a presentation on using AI to write a murder mystery novel at the 2023 Generative AI Conference, Yiin uses this presentation again as an example to share with us the principles by which writers interact with AI to co-write a script.

    In text generation, humans remain the creators of a story’s overall structure–an empty shell waiting to be filled with content that brings the story to life. In this context, generative AI comes in rather handy – both as a generator of text and collector of information. Generative AI is important for “tearing down an echo chamber,” Yiin said, “It seems that great writers in history often led difficult lives, but it is precisely because of their struggles that they were able to produce such profound works. Now that AI is able to role-play on behalf of them, they no longer need to go through all the challenges in life themselves.”
 

1. No ideas yet? Ask AI to “prove by exhaustion”

    For novice learners just starting to learn about AI interaction, it is unclear to them to what extent generative AI can perform specific tasks. You should avoid overthinking and simply begin with the scientific method of exhaustion, also known as proof by cases. When deciding which styles to adopt to write a script, ask AI to prove, by exhaustion, a variety of writing styles, then adopt each style to help you think during the process.
 

2. Informative contents – from “famous person role-play” to the “language of the lens”

    Add more “people” elements. In addition to thinking from a first-person perspective, don’t forget the importance of role-playing. Select different authors’ names to guide your writing, then note which ones prove helpful and which do not–this makes gathering ideas and information much more efficient. Yiin “sometimes asks AI to tell me what authors there are, though some of the authors may be a ‘bluff.’” He advises random checks, as every attempt made is a revelation of a new piece of information about the thing of interest.

    You might start to explore other possibilities for defining storytelling techniques, such as experimenting with reverse chronology using ChatGPT or incorporating film editing concepts like the language of the lens. CTO Yiin uses the language of the lens as an example to demonstrate prompt engineering. “‘Point-of-view shots, low-angle shots, drone overhead shots, or shots at other camera angles’ can be taken to produce detailed images; instructions such as ‘fast dynamic movement’ or ‘time-lapse photography’ can be used to produce dynamic pictures. You can also ask AI for terminologies/jargons.”
 

3. AI as a Professional Consultant and a Punctilious Craftsman

    The most important part of using AI to design scripts is conversing non-stop with AI. Last year, Yiin authored a murder mystery novel about a series of murders involving a person in charge of an AI startup company set against the backdrop of AI being the mastermind behind the murders and the victims being developers of service delivery recommendation apps. “In order to design an AI that kills a person in a way that is commensurate with the real world, I engaged in a back-and-forth discussion with ChatGPT. Eventually, I decided to kill the developers by triggering a fatal allergic reaction caused by a peanut allergy.” To ensure a logical and reasonable murder process, Yiin prompted ChatGPT to provide information on the sequence of physiological responses following peanut ingestion. The process began with hyperemia in the smooth muscle, followed by a cascade of other reactions outlined by ChatGPT in precise order. Yiin finds this process of creation amusing, “I can quickly connect the vast medical knowledge to my work without having to consult another expert.”
    “Writing scripts with AI depends largely upon the use of ‘pseudocode.’”

Pseudocode and GPTs

    By using pseudocode, Yiin can prompt AI to produce a novel, a picture book, animation, and other content. He illustrated the concept of pseudocode with this scenario: A language barrier exists between a Mandarin-speaking person and a Taiwanese-speaking person. To overcome this barrier, a third person who speaks “Taiwanese Mandarin” – analogous to pseudocode in AI – is needed to facilitate mutual understanding. In the past, humans had to learn machine languages (programming language) to instruct a machines. Now, machines have to learn human languages to understand our instructions. This communication gap, arising from the lack of a shared language, is why pseudocode emerged as a vital tool for balancing human and machine interaction.

    Yiin explains that processes are now designed using pseudocode, such as loops, if, and if-else statements. However, details of the execution are augmented with natural language descriptions. In the past, Chinese word-segmented writing typically involves writing complex program contents. Now, a simple statement like “Use vertical bar to segment text by semantics” would suffice. In other words, AI outputs are more effective when execution details are written in natural language and control flow in programming language. Compared to one-by-one prompt inputs, pseudocode “can standardize repetitive tasks,” said Yiin. For instance, in creating a picture book, a writer must maintain consistency in the style of illustrations and characters while ensuring the story remains, even as the plot evolves. This repetitive process necessitates back-and-forth conversations with AI, resulting in significantly high communication cost.

    When compiling a picture book, Yiin usually starts off with “a theme” which he then embeds in every page of his book to ensure coherency and consistency in the pictures’ visual style. Each page of the book will feature a composition, plot, and image description–a standardizable model provided that definitions are set using pseudocode in the first place. Thereafter, any picture book can be produced simply by specifying the “axis of the story” and “main visual style.” Yiin recounted his experience of using pseudocode to produce seven bedtime stories for children, one for every day of the week. The theme of his stories was to teach children the value of being virtuous. As long as each task is properly defined, seven stories about this one theme can be created. 

    GPTs, released by OpenAI, have gained widespread attention. GPTs are a new way for paid ChatGPT users to create a tailored version of ChatGPT to complete specific tasks or work and share that creation with non-paying users. Each GPT can only perform a small task. While it cannot execute complex process control, it can switch from a single-function to a multi-function composite model or composite process. Yiin suggests “setting a default function first. Say, a GPT has five functions, then first determine which function a user needs and proceed from thereon.”

 

Using ChatGPT with Pseudocode for Data Organization 

    For more information, see Yiin’s YouTube video on thesis/dissertation topic selection, which helps researchers choose their research topic with a pseudocode design process. In the video, Yiin adheres scrupulously to “the spirit of scientific research.” He performed the following actions on ChatGPT to select a medical research topic:
 

1. Perform a sweep search: I asked GPT to find 30 research articles related to a topic from a medical journal. 
2. Read the articles: I issued a specific instruction, asking GPT to read every article to avoid fabricated content or hallucinations.
3. Group content based on similar topics: I asked GPT to list the exact number of articles and specify which group each article belongs to.
4. Read each group of articles and identify research gaps: I asked GPT to read the content of each group, provide the median for a specific group, then tell me the research gaps in this group of articles.
5. Rate the research gaps: The listed gaps are all potential topics. Thus, I asked GPT to rate or score the gap by its research value, execution difficulty, or possible risk of failure. I can then make a final decision accordingly.

    At the mention of the problem of verifying LLM-generated contents, Yiin admitted the difficulty of tackling hallucinations in LLMs. According to Haziqa Sajid, LLMs can hallucinate due to overfitting errors in encoding and decoding and training bias, thus generating outputs that are syntactically and semantically correct yet disconnected from reality based on false assumptions, such as fabricated, factually incorrect content. Nevertheless, here are some small details you can watch out for:
    If ChatGPT is extracting data or reading a PDF article, it will display a prompt stating that it is extracting data or reading the article. If you realize ChatGPT gives you an answer without displaying any prompt, then the answer given is a lie. When hallucination occurs, the given DOI of an article may be 1234567, which is probably also a lie. You must set up a mechanism to check and inspect the generated content. You can also prepare another prompt telling GPT to “tell me the reason and indicate the data source,” which makes it difficult for GPT to deceive you because the cost of deception will be high. The absence of any fact-check or inspection mechanism will significantly increase the possibility of hallucination.

 

The Randomness of AI and a Different Workflow

    Natural language descriptions will be an integral part of programming language. The ability of AI to comprehend human language hinges on not only technology but also the language used in human–AI interactions. Yiin affirms the cognitive gap existing in the interaction between humans and LLMs. When certain variables are present in text-to-text models, the gap consequent to text-to-image conversion increases. Take Midjourney as an example: to generate images, users typically type in a prompt composed of “a bunch of commas, and a stack of characters.” Although Midjourney generates captivating images, it actually does not understand sentences at all. “It has transformed into a new language or instruction (rather than a purely natural language), and if it reverts to ‘precise control’, it is not the direction I anticipated,” said Yiin.

    What makes AI fascinating is its “randomness,” said Yiin. The concept of creating an AI environment is quite different to past concepts. Take video production as example: film directors often have a clear vision for visual design and camera techniques, with videos shot by individuals who exercise a high degree of control over filming and lighting equipment. In contrast, a key characteristic of generative AI is its randomness, which makes it uncontrollable: the more you try to control AI, the more uncontrollable it becomes. The harder you attempt to fine-tune it, the more you may just find yourself struggling to achieve the desired results. “It is this power of randomness that makes people realize that the solution we have in mind is not necessarily the best solution,” said Yiin, “Our past aesthetic experience will confine us, limiting our exploration to a specific sphere of influence. However, beauty knows no bounds. It is because of its randomness that AI is able to generate some very interesting and amazing outcomes.”

 

The Future of Industries

    The multitasking nature of AI makes small and medium enterprises (SMEs) more resilient. An AI-powered workplace makes individuals more flexible and productive, offering SMEs greater advantages compared to large corporations–a belief held by Yiin. In an industrial environment, “AI won’t replace humans, but a person with AI will replace a group of people.” On the other hand, Yiin said, “There may be some people, who, like me, suddenly feel that life actually has a lot to offer, so, feeling reluctant to continue working for a large corporation, they move on to small businesses.” This thus changes the stereotype about the small businesses’ inferior market status.

    The emergence of LLMs enables the use of API to make websites accessible in multiple languages. Yiin believes that “LLMs should be able to overcome language barriers in all form of communication, except for face-to-face interactions. Additionally, the cost of adopting LLMs will be lower for transnational companies.” In Taiwan, SMEs doing business with large corporations are often haggled into providing customization services at no extra cost. Therefore, even though business deals with large companies are very profitable, most of the money is spent on delivering these services. Yiin foresees that “if SMEs become mainstream in the future, they won’t have much bargaining power and will concede to working with the functions you provide. In future needs assessment and function designs, you should begin by focusing on needs shared across people/countries/cultures. I think this is imperative because then you can do business in other countries.”

    Therefore, we need to develop a stronger “abstraction” capability. In the past, we were likely inclined to divide jobs and tasks strictly according to professional expertise, as long as individuals fulfill the responsibilities assigned to them. However, the emergence of generative AI gave Yiin a glimpse of a new trend: Gone are the days of such division of labor. Now, we must find the commonalities within differences (e.g., common models and abstract features shared across several different books), much like how ChatGPT transforms all languages into a single form to predict the next word–this represents a powerful form of abstraction. When abstraction is complete, problems of this type (text interaction) can be resolved all at once. Yiin further elaborates, “In truth, generative AI is not about breaking a problem into smaller problems, but finding commonalities and solving the problem all at once. This way of thinking is quite different.”

    On one of the major social concerns about AI: Will it replace humans? Yiin thinks AI is not mature enough yet in its control of the “sense” for certain professions. In creating storyboards, AI remains inferior to professionally trained directors and producers. According to Yiin, existing storyboarding tools tend to produce results that are less than ideal, with some even telling the “wrong” story. Such expertise—professional knowledge or aesthetic literacy that is acquired only through professional training—is probably not something that can be compensated for by generative AI. Nevertheless, AI in this context helps lower the “startup cost,” particularly during the idea generation stage. The efficiency with which AI generates results enables effective communication with film directors and sponsors, facilitating the visualization of ideas and, “more importantly, reducing the pain of being misunderstood during the communication process,” said Yiin.