This passage of text is written as summary to Ayu Purwarianti’s presentation with the same title, Large Language Model: Potensi dan Limitasi at Maranatha University, December 7th 2024. It may contains over-simplified explanations and inaccuracies as the writer was too ashamed to open good ol pen and paper in a conference full of tech students and professionals. Upon finding difficulties in trying to recall the lecture, the writer is determined to squash down the icky feeling of appearing too studious later on if such opportunity arises again. (Wish me luck!)
Ayu Purwarianti, Associate Professor from Bandung Institute of Technology, giving introduction to the audience;
At the center of Artificial Intelligence technology, there lies Large Language Model at work. It is a subset of Machine Learning that generates data based on learned pattern of a huge pool of input. It is a new solution that created the recent boom of Artificial Intelligence use in public. Large Language Models are enabled by the advent of Graphic Processor Unit (GPU) as the workhorse to train AI model with.
There is a key difference between traditional Machine Learning and the current Large Language Model (LLM). Traditional Machine Learning requires personnel to manually input labels for different groups of data. The labeled group of data then can be used to create a pattern or model for each group it is trained with. This is a laborious task, but can delivers accurate result per the quality of data it is fed with. Large Language Model does not require the labor of manual labeling and grouping of data as what required in traditional Machine Learning. LLM gathers its own patterns from large pool of data based on statistic of what comes next after a word to word. It’s really convenient to train Artificial Intelligence using LLM method. Anyone can do it as long as they have a large, quality pool of data.
In recent years, we are bombarded by the practicality of AI use. We can generate text for speech, mail, presentation, general copywriting, or even fiction. We can transcribe and translate meeting sessions in real time to notes. We can transform our ideas to images or even videos by prompting. Another common use case is info searching and the whole process of hiring from the point of view of both recruiter and the recruitee. AI is surely a convenient addition to our current modern world life style, isn’t it?
We cannot deny the big impact that AI brought to us in real life. Developers can improve their productivity by letting copilot handle the begrudging aspect of writing boilerplate codes. Students have much easier time to write essay, teachers need to think twice before giving assignment in form of paragraphs. Human Resource offices reported 80% of their applicants use AI to generate documents without changing the content into more human-proofread output. There are more to come in future especially at the fore front of risk profiling and media generation.
The convenience that AI had brought to us, of course, is not without some drawbacks. Despite how smart Artificial Intelligence had become, it is only codes running through large chunk of data and spitting out the most likely answer to a question. It doesn’t necessarily understand the data it had been trained with. The term for such description in Machine Learning is Stochastic Parrots. The machine essentially parroting back an answer without knowing whether it is true or not. We must be vigilant to check the validity of AI’s answer to our question. However, many AIs are equipped with such complex model, the answers would sound convincing at first glance even though it is actually invalid. AI made connections by statistic and often arrived to the wrong conclusions. This is what people usually called Hallucination in AI. This phenomenon can mislead its users and have fatal impact to our life.
Other than the validity of answers, we also need to be wary of few things when it comes to AI. First of, using AI can circumvent the long process of how we learn things. In the long run, it can make our brain atrophic to critical thinking. This is especially dangerous when the notion is combined with AI inaccuracies and rampant scamming. Next up, AI can be biased based on the data it had been trained to. If you ask AI how many houses there are in a picture, the answer can vary model to model depending on the data it had been fed with. In case of profiling, a model would call a doctor he, and a nurse she. This is an indication to bigger problem at hand. AI can make snap judgement for law and medical system that is partial to one ethnicity, gender, or even financial background. Other than that, we need to be wary of targeted scam, hoax news, and impersonation of a person or even public figure. AI can generate convincing text, sound, and pictures, which depending on the target audience, people may forget checking the validity of it in favor to reacting to the inflammatory content.
Now, the countermeasure for AI problems are already on the go. For documents, people can run through AI checker to see whether or not the writing is AI generated. For video, we can check for inconsistencies and weird renderings in the video because AI still hasn’t perfected the art of making flawless video just yet. Actually, the technology to detect AI is still less to be desired as the current ones declare AI generated output as human’s. That’s kind of worrying, isn’t it? Well, the development for watermarking technologies had already begun and we can hope it will be a salvation in the future. For the meantime, we just need to be careful of the information we’re receiving and itching to spread.
UNESCO guide to ChatGPT use, published in April 2023;
When it comes to fact checking, there is also a stark difference between using the usual web page search and LLM-powered search. The older Information Retrieval (IR) system works by indexing the most relevant search results, then presenting it in form of list of links, complete with info snippets for each links. We can get our intended information by clicking the links and judge by ourselves, whether the contents are relevant or sufficient enough. Using LLM for IR makes the process of getting result much faster for us since the model generates summary in human language. Still, the LLM search has risk of hallucination with no way to confirm the validity without digging in somewhere else, and the info presented might even be outdated. We would have better accuracy searching info with run of the mill web page search, but cannot be denied that AI-search is way faster.
Retrieval-Augmented Generation (RAG) system has recently emerged as a bridge to the best of both worlds. RAG system uses not only its own pool of data, but also actively search up-to-date information and references it. When given a query, it would give us summary with numbered links as references we can check ourselves. The summary may still contains false information, but at the very least we are given the means to investigate the context of why it is still there, in the list. In the modern age where everything is fast, we can, and need to verify the information we are given so readily.
The difference between LLM and RAG search result; Currently RAG is still in need of exploration and improvement.
Top comments (0)