Large Language Model
A transformer trained on vast text to predict the next token, gaining broad language and reasoning ability.
Definition
A large language model (LLM) is a transformer with billions of parameters trained on large text corpora to predict the next token. From this simple objective, it acquires grammar, facts, and useful reasoning and generation abilities.
Context length, the amount of text a model can consider at once, is a defining constraint: it bounds how much a model can attend to in a single request. Retrieval-augmented generation sidesteps this by fetching relevant passages into the context rather than relying on parameters alone.
Because a language model predicts plausible text rather than retrieving verified facts, it can state falsehoods with the same fluency as truths, a failure mode called hallucination. Mitigations include grounding responses in retrieved documents, training the model to express uncertainty, and keeping a human in the loop for consequential decisions. Treating model output as a draft to be checked, rather than an authority, is the sound default for any serious application.
Lifecycle
- Pretraining: next-token prediction on broad data.
- Fine-tuning: adapt to instructions or a domain.
- Alignment: shape behavior toward helpful, honest output.
- Inference: generate text one token at a time.
Strengths and limits
LLMs are fluent and broadly capable but can produce confident errors (hallucinations), reflect biases in their training data, and lack guaranteed factual accuracy. They are tools for drafting and synthesis, not sources of ground truth.
Why it matters
LLMs have reshaped how people interact with software, from writing to coding to search.
Fusion connection
Kronos uses language models to help organize and cross-check its technical documentation, while every physics claim remains anchored to simulation and the program's frozen canon, never to a model's unverified output.