Navigating the AI Frontier 2024

The development of AI agents began in the 1950s,6 and since then they have evolved from simple rule-based systems to sophisticated autonomous entities capable of complex decision-making. Early AI was characterized by deterministic behaviour, relying on fixed rules and logic that made these systems predictable but unable to learn or adapt from new experiences. Advances in AI research introduced systems that could handle larger datasets and manage uncertainty, leading to probabilistic outcomes and non-deterministic behaviour. This shift enabled more flexible and dynamic decision-making, moving beyond rigid frameworks. The 1990s marked a significant turning point, as machine learning applications became more widespread. AI systems began to learn from data, adapt over time and improve performance. The introduction of neural networks during this period laid the foundation for deep learning, which has since become essential to modern AI. Since 2017, the rise of LLMs has transformed AI’s capabilities in natural language understanding and generation. These models use vast amounts of data to produce human-like text and engage in complex language-based tasks. Today’s AI agents use various learning techniques, including reinforcement learning, or transfer learning, allowing them to continuously refine their abilities, adapt to new environments and make more informed decisions. Key technological trends 2.1 Over the past 25 years, the increase in computing capacity, the availability of large quantities of data on the internet and novel algorithmic breakthroughs have enabled significant developments in the base technologies behind recent advances in the capabilities of AI agents. These are briefly described below. Large models Large language models (LLM) and large multimodal models (LMM) have revolutionized the capabilities of AI agents, particularly in natural language processing and the generation of text, image, audio and video. The emergence of large models has been driven by several technological advances and by the transformer architecture, which has paved the way for a deeper understanding of context and word relationships, considerably improving the efficiency and performance of natural language processing tasks.7 In summary, advanced AI models have enabled better understanding, generation and engagement with natural language. Machine learning and deep learning techniques A range of techniques have greatly improved AI models through increased efficiency and greater specialization. Some examples of machine- and deep-learning techniques include:1. Supervised learning: facilitates learning from labelled datasets, so the model can accurately predict or classify new, previously unseen data.8 2. Reinforcement learning: enables agents to learn optimal behaviours through trial and error in dynamic environments. Agents can continuously update their knowledge base without needing periodic retraining.9 3. Reinforcement learning with human feedback: enables agents to adapt and improve through human feedback, specifically focusing on aligning AI behaviour with human values and preferences.10 4. Transfer learning: involves taking a pretrained model, typically trained on a large dataset (e.g. to recognize cars) and adapting it to a new but related problem (e.g. to recognize trucks).11 5. Fine-tuning: involves taking a pretrained model and further training it on a smaller, task-specific dataset. This process allows the model to retain its foundational knowledge while improving its performance on specialized tasks.12 These and other learning paradigms are often used in combination and have dramatically expanded the problem-solving capabilities of AI agents in various areas of application. The evolution of AI agents is detailed in Figure 2, while the agent types are further expanded in the following section. Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 9