Get Lost in the Magic of Autumn: 5 Unforgettable Novels to Snuggle Up with This Season
Autumn is the season for reading. I asked the generated AI chatbot, “I would like you to give me a list of novels that are set in autumn.” They include Han River’s “Vegetarian”, which won the 2024 Nobel Prize in Literature, “Norwegian Wood” (Haruki Murakami), “Uncharlie” (Lucy Maud Montgomery), and “The Legend of Autumn” (Jim). Harrison), “Night Train to Lisbon” (Pascal Mercia), “The Color of Memories” (Richard Yates), “Heartless” (Lee Kwang-soo), “Whispers in the Trees” (Shin Kyung-sook), “Stranger Room” ” (Choi In-ho) and others were introduced.
He then asked the question again, “How was autumn portrayed in Han River’s novel “The Vegetarian”? The generated AI chatbot replies, “The novel “The Vegan” richly expresses the life and emotions of the main character through beautiful autumn scenery,” and kindly informs the source (sports.chosun.com) (Microsoft ·Copilot). However, if you actually search for the source, you won’t find any explanation that the novel’s background is autumn. It’s a “wow” caused by the phenomenon of hallucination. One AI chatbot replied, “’Vegetarian’ is a novel that delicately depicts the lonely emotions of autumn and the inner changes of humans.” When asked again about the rationale, the answer was, “I’m sorry. “Autumn was incorrect information. I apologize for categorizing it incorrectly,” (Claude of Enthropic) admits his mistake.
On the other hand, Chat GPT classifies Choi In-ho’s “Stranger’s Room” instead of “Vegetarian” as a novel with an autumn background, but when asked if they would like to include the basis for autumn, they answered, “The main character is in a city.” The story revolves around the feeling of alienation and loneliness felt in the middle of nowhere, and the lonely atmosphere of autumn is imbued with the emotion of the work.”
사진 확대 Hallucination (parrot shit) generated by the author from MS Copilot AI One of the banes of generative AI, the so-called “gobbling” or “hallucination”, remains. The generated AI chatbot’s replies are just a “probabilistic parrot” that provides the most probabilistically similar corpus, and it doesn’t even know what it’s saying. As a result, there are many cases in which users make decisions based on websites guided by chatbots, only to suffer damages resulting in liability issues. So how can we reduce hallucination errors in AI models?
First, there is the “Retrieval-Augmented Generation” (RAG) technique. The principle of RAG is similar to using Wikipedia or Open Book when taking an exam. You can quickly find the information you need and create answers by referring to your notebook containing expected exam questions and answers, as well as related textbooks. RAG searches in real time for the latest documents and web pages related to your question, increasing the reliability and accuracy of your answers. Rather than the conventional method of finding answers from vast amounts of learned information (a dumping ground for all kinds of information), this method generates answers by connecting related external databases and documents in real time. For example, if you use the RAG method to ask about vegetarian seasonal information, the accuracy will be high because the latest information will be searched and provided from related books or news.
Another solution is the Small Language Model (SLM). Because large language models (LLMs) provide answers based on vast amounts of information from a variety of sources, false answers can often occur. On the other hand, if a small language model (SLM) is trained specifically for a specific domain, it can reduce unnecessary information and provide tailored responses. For example, rather than searching for information in a huge library that contains all the world’s knowledge, using specialized books or dictionaries that only cover a specific topic will increase the reliability and accuracy of the information. SLM has smaller parameter sizes, works more efficiently in a given domain, and can be optimized for your work. It has fast data processing speed, is capable of real-time response, can operate smoothly in mobile environments, and can significantly reduce development and maintenance costs.
사진 확대 Chatbot model generated by the author using MS Copilot AI In addition to this, Reinforcement Learning from Human Feedback (RLHF) is useful for reducing gossip. Reinforcement learning based on human feedback can improve performance by rewarding accurate and relevant responses in place of factually incorrect or irrelevant information. Additionally, fine-tuning a model based on specific, trusted datasets within a domain increases consistency and accuracy. Applying domain-specific tweaks and RLHF together can reduce whitewashing and provide more reliable information.
사진 확대 “Reading season” autumn image created by the author from MS CoPilot AI Generative AI nonsense appearing in Nobel Prize-winning works and literary works is laughable, like the case of King Sejong the Great throwing his Macbook. However, if such errors occur in fields such as medicine, finance, national defense, and law, there is a risk of fatal consequences for human life and property. This is an area where even the slightest mistake cannot be tolerated. As we saw earlier, there are various solutions, but “human feedback” with specialized knowledge and insight is always valuable. Human beings have developed exceptional insight and expertise through constant thinking and learning. He is the primate of all things who creatively solves problems even when given only a small amount of information. It is humans, not AI, that come up with new solutions like RAGs (Retrieval Augmented Generation) and SLMs (Small Language Models). Human thought is the source power that reduces “AI hallucination.”
For a long time, humans have relied on searching rather than thinking, and now even searching has become a chore, and we have entered an era where commands are given through artificial intelligence prompts. However, in autumn, which is the season for reading, I think deep thinking is more appropriate than searching or commanding. No matter how much artificial intelligence develops, the scent and value of thinking humans will always be great.
enlarge photo [ヨ·ヒョンドクKAIST G-School院長/技術経営大学院教授]