What is Agentic AI? Stanford Explains LLM Planning & Problem Solving (2026 Guide)

Large Language Models (LMs) are essentially sophisticated machine learning models designed to predict the next word in a sequence based on a massive corpus of training data. As Insop Song from GitHub Next explains, the journey of a language model begins with a pre-training phase, where the model consumes vast amounts of internet text and books to understand linguistic patterns and global knowledge. However, a pre-trained model alone is often difficult to control or use for specific tasks.

To bridge this gap, the industry employs post-training techniques, specifically instruction tuning and Reinforcement Learning from Human Feedback (RLHF). This stage aligns the model with human preferences, teaching it to follow specific commands rather than just completing sentences. By training on datasets formatted with instructions and expected outputs, the model becomes a more versatile tool for applications like coding assistants and conversational interfaces like ChatGPT.

Despite these advancements, developers must understand that an LM is still a probabilistic engine. It generates the most likely next token, which can lead to issues if the prompt is ambiguous. Clear communication is the bedrock of effective AI interaction. As we move toward more complex systems, the quality of these foundational models determines the potential of the higher-level agentic structures built upon them.

To extract the maximum value from modern LMs, developers must employ strategic prompting techniques. Writing clear, descriptive instructions is non-negotiable; as Insop Song notes, the model cannot read your mind. Detail is your friend. Furthermore, providing 'few-shot' examples—showing the model the exact format and style you expect—significantly boosts the consistency of the output. This is particularly vital in production environments where structured data is required.

Another critical technique is 'Chain of Thought' (CoT) prompting. Instead of asking for a final answer immediately, you instruct the model to think step-by-step. This 'time to think' allows the model to allocate more attention to its own reasoning process, often correcting errors that would occur in a 'one-shot' response. For complex tasks, breaking down the prompt into a chain of simpler sub-tasks ensures higher accuracy at each stage.