Large language models (LLMs) like ChatGPT have transformed the way we interact with AI, there is still a lack of understanding around how they work and their limitations. In a commentary published in Wen Wei Po on 18 June 2025, Professor Siu-Ming YIU, Associate Director (TPg) of the School of Computing and Data Science (CDS), provides insights into the training of these models and the associated challenges and limitations.
Understanding the Capabilities and Limitations of Large Language Models
Upon the release of ChatGPT by OpenAI, the world was astonished by its remarkable capabilities. However, the emergence of ChatGPT was not an overnight breakthrough. This is the result of years of accumulated research, and the technology behind it has been under development for many years.
Training Large Language Models
There are three key elements to consider when training AI systems: data, annotated data, and the underlying AI models and algorithms. From a broader perspective, the training of a LLM follows a similar process, with one notable difference: it requires an enormous amount of data. Most LLMs, including ChatGPT and DeepSeek, depend on open data sources, such as billions of web pages and documents available online, while also incorporating proprietary datasets. It is not feasible to manually annotate such vast amounts of data.
This is where advanced techniques can be employed. One simplified way to understand this process is through an analogy: imagine teaching a child to construct conditional sentences like “If…, then…”. An educator will provide a few examples, invite the child to create new sentences following the same structure, and then provide feedback on their work. Similarly, in training LLMs, we only need to annotate a small subset of data. The system then learns and reinforces whether its answers are correct, greatly reducing the need to label billions of web pages and documents. We also depend on the underlying AI models and algorithms to make this process work. Based on the training data and the foundational model, the system builds a LLM capable of answering users’ questions.
It is also important to acknowledge that LLMs have their limitations. Just as the materials provided to children shape
Challenges and Limitations
Despite their impressive performance in language generation, most current AI systems are not yet capable of doing everything. For example, if we ask a model: “Given two buckets, one holding 5 litres and the other 3 litres, but neither is marked, how can you obtain exactly 2 litres of water?”—this is a simple logical problem, but most LLMs cannot provide the optimal or even an accurate answer.
It is likely that most systems will undergo continuous enhancement, facilitated by enhanced training in logical problem-solving methodologies. However, it should be noted that the answers provided may vary depending on the training data (especially proprietary datasets) and the differences in the underlying AI models and algorithms. This emphasises the necessity to understand both the capabilities and limitations of LLMs.