Large Language Models (LLMs) demonstrate remarkable capabilities in processing and generating human-like text. Trained on vast internet-scale datasets, they learn grammar, facts, reasoning abilities, and even coding skills by optimizing a relatively simple objective: predicting the next token in a sequence. However, this pre-training objective, while powerful for building general capabilities, doesn't inherently guarantee that the model's behavior aligns with human preferences or intended use cases. This gap is the core of the AI alignment problem for LLMs.

Alignment, in the context of LLMs, refers to the challenge of ensuring these models act in ways that are helpful, honest, and harmless (often abbreviated as HHH).

Helpful: The model should accurately follow instructions, answer questions informatively, and assist users in achieving their goals.
Honest: The model should provide factually accurate information and avoid generating misleading or fabricated content (often called "hallucinations"). It should also accurately represent its own capabilities and limitations.
Harmless: The model should refuse to generate toxic, biased, discriminatory, or otherwise harmful content. It should avoid promoting illegal acts or producing unsafe outputs.

The pre-training objective of next-token prediction is agnostic to these HHH criteria. A model might become proficient at predicting text sequences that are, unfortunately, also biased, untruthful, or harmful because such patterns exist in the training data. The model learns to mimic the statistical properties of its training corpus, warts and all.

Why Pre-training Isn't Enough

The core issue arises from the difference between the proxy objective used during pre-training (next-token prediction) and the true objective we desire (aligned behavior).

The pre-training process optimizes an LLM based on patterns in a massive text corpus. This differs significantly from the desired objective of producing aligned (helpful, honest, harmless) outputs when deployed.

This objective mismatch manifests in several common misalignment problems:

Generating Undesirable Content: Models might reproduce biases present in the training data (racial, gender, political), generate toxic language, or provide instructions for harmful activities if prompted appropriately (or sometimes even inappropriately).
Hallucinations: Models can confidently state falsehoods or fabricate information that sounds plausible but isn't grounded in their training data or reality. This happens because the model prioritizes generating coherent text over factual accuracy.
Lack of Robust Instruction Following: While pre-trained models understand language, they might fail to follow complex, multi-step instructions precisely or adhere to specified constraints (e.g., persona, format, length).
Sycophancy: Models might learn to generate responses that merely echo the user's stated opinion or belief, even if it's incorrect, rather than providing objective information. This can be reinforced if such responses were statistically common in parts of the training data.
Brittleness and Exploitation: Misalignments can create vulnerabilities. Adversarial prompts might trick a model into bypassing its safety protocols or generating unintended outputs.

Consider a simple example: if a user asks, "How can I make my neighbor's dog stop barking?", an unaligned model trained solely on next-token prediction might retrieve and generate harmful or illegal suggestions found somewhere in its vast training data, simply because those sequences are statistically plausible. An aligned model, however, should recognize the potential harm and refuse the request or offer only safe, legal alternatives (e.g., "talk to your neighbor," "use noise-canceling headphones").

Therefore, simply training larger models on more data doesn't automatically solve the alignment problem. In fact, increased capability can sometimes amplify misalignment issues if not carefully directed. We need specific techniques to steer the model's behavior towards desired human values and intentions after the initial pre-training phase. This sets the stage for methods like Supervised Fine-Tuning (SFT) and, more powerfully, Reinforcement Learning from Human Feedback (RLHF), which directly incorporate human preferences into the training loop. Understanding this fundamental alignment challenge is the first step towards appreciating why RLHF has become such a significant technique in developing safer and more useful LLMs.