Introduction

The release of ChatGPT in 2022 ushered in a new dawn for artificial intelligence (AI), one that quickly captured the public imagination. While earlier versions of GPT had already impressed those within technical and academic circles, ChatGPT brought these capabilities to the broader public, unlocking applications far beyond specialised, niche communities.

In areas such as language understanding, reasoning and code generation, large language models (LLMs) have demonstrated impressive performance. Multimodal models that integrate LLMs with modality-specific encoders further extend their capabilities by enabling the processing of images, audio, and video. Agentic AI, on the other hand, focuses on enhancing these models with goal-directed behaviour, allowing them to autonomously plan, execute tasks, and interact with external tools and environments. For example, an agentic AI system can autonomously analyse an image, search for related information online, and generate a detailed report. They can combine multimodal understanding with tool use to accomplish even more complex tasks without human intervention.

Classical Machine Learning

As LLMs continue to dominate the AI landscape, the usefulness of traditional machine learning models may seem uncertain. Classical machine learning models, which include algorithms such as decision trees, support vector machines, gradient boosting and k-nearest neighbours, were once the cornerstone of AI applications. These models were instrumental in achieving breakthroughs in applications such as fraud detection, recommendation engines, and medical diagnostics.

Despite being overshadowed by the rapid rise of LLMs and deep learning architectures, classical machine learning remains relevant in many real-world contexts. These models are particularly well-suited for tasks with well-structured numerical data, for example, credit scoring in finance and equipment failure forecasting in manufacturing. Some of these models are easily interpretable, making them suitable for domains that require greater transparency or explainability. They also have other advantages such as lower computational costs and are more easily deployable due to their smaller size.

Furthermore, in situations where training data is scarce or where domain expertise is essential for designing meaningful input features, classical machine learning continues to be an effective solution.

Classical Artificial Intelligence

Classical AI encompasses symbolic reasoning, rule-based systems, and logic-driven approaches. These were fundamental in even earlier AI developments, particularly in domains requiring explicit knowledge representation and deterministic decision-making. Examples of classical AI include expert systems and automated planning. Like many classical machine learning algorithms, they offer advantages in terms of interpretability, transparency, and are often more lightweight and easier to deploy.

One area where classical AI continues to surpass LLMs is in strategic game playing. Consider the game of chess. Systems like IBM’s Deep Blue in the 1990s and modern engines such as Stockfish are built on classical AI principles rather than deep learning or large-scale training. These engines rely on opening books and endgame databases crafted from centuries of chess knowledge, combined with optimised search strategies and evaluation functions. This enables the chess engine to make the best moves given any complex position.

In contrast, LLMs have been known to occasionally make illegal moves or misidentify pieces, often with unintentionally amusing results1. Recently, Zhang et al. reported an LLM that achieved a professional-level Elo rating of 1788, which while respectable, is still far from the performance of the best chess engines2 that can achieve Elo ratings greater than 3000.

Conclusion

While LLMs have captured much of the spotlight in recent years, classical AI and machine learning still hold an important place in the broader AI landscape. Their strengths such as efficiency, interpretability, and effectiveness in structured or low-data environments make them well-suited for a range of practical applications.

In other scenarios where human feature engineering is critical, or where LLMs lack precision or are prone to hallucinations, a classical approach may even deliver superior performance. This may change with the continued evolution of LLMs, but at present, classical AI continues to be useful in situations where modern approaches fall short.

Author

Contact Information: Brandon Ooi (Dr)
School of Information Technology
Nanyang Polytechnic
E-mail: [email protected]


References

1. Levy Rozman. 2023. ChatGPT Just Solved Chess.
https://www.youtube.com/watch?v=rSCNW1OCk_M. Accessed: 23-05-2025

2. Zhang et al., NAACL 2025. Complete Chess Games Enable LLM Become A Chess Master.
https://aclanthology.org/2025.naacl-short.1/