Large Language Models: Architecture and Limitations

Examine the architecture of large language models in depth, understanding their capabilities and fundamental limitations.

12 min 3 quiz questions Grade 10

Understanding LLM Architecture Explains Their Behavior

Many of the surprising behaviors of large language models — their fluency, their hallucinations, their biases — follow directly from their architecture. Understanding how they work explains why they behave as they do.

Transformer Architecture

Modern LLMs are based on the transformer architecture, introduced in 2017. Transformers use a mechanism called attention to weigh the relevance of different parts of the input when generating each output token. This allows them to capture long-range dependencies in text.

Attention Mechanism:A component of transformer models that allows the model to weigh the relevance of different parts of the input when generating each output, enabling capture of long-range dependencies.

Token:The basic unit of text that LLMs process — roughly a word or word fragment. LLMs predict the next token based on all previous tokens.

Why LLMs Hallucinate

LLMs are trained to predict the most likely next token given the context. They do not have a separate "fact database" to consult — they generate text based on statistical patterns. When asked about something outside their training data or at the edge of their knowledge, they generate plausible-sounding text that may be false.

The Stochastic Parrot Problem

Critics have described LLMs as "stochastic parrots" — systems that generate statistically likely sequences of words without understanding their meaning. This captures something important: LLMs can produce fluent, coherent text about topics they have no genuine understanding of.

Stochastic Parrot:A critique of LLMs arguing they generate statistically likely text without genuine understanding of meaning — producing fluent but potentially meaningless or false output.

Use LLMs for What They Are Good At

LLMs excel at tasks that benefit from pattern matching and fluent text generation: summarizing, explaining, brainstorming, drafting. They are unreliable for tasks requiring factual accuracy, current information, or genuine reasoning about novel situations.

Ready to test your knowledge?

3 questions · Grade 10 level

Attribution, Authorship, and the Ethics of Credit Responsible AI Use: Principles and Practice