AI Concepts & Fundamentals

Large Language Model (LLM)

An AI trained on massive text that predicts and generates language.

Quick Answer

A large language model (LLM) is a deep learning model trained on massive amounts of text to predict the next word in a sequence. That simple objective, at scale, lets it generate fluent writing, answer questions, summarize, and follow instructions. LLMs power tools like ChatGPT and Claude, but they can produce confident, wrong answers, so their output needs checking.

The Full Picture

LLMs exist because language is how most human knowledge is stored — in reports, specs, emails, and contracts. A model that can read and produce language fluently can act on all of that unstructured text, which older software could only search literally, not understand.

Mechanically, an LLM is trained to predict the next token (roughly, the next word-piece) given everything before it. Trained on enormous text corpora, it internalizes grammar, facts, and patterns of reasoning as statistical relationships. At use time you give it a prompt and it generates a continuation, one token at a time. The dominant architecture is the transformer, which lets the model weigh how much each earlier word matters to the next (A Survey of Large Language Models, 2023).

The headline capability is generality: one model handles summarization, translation, drafting, classification, and question-answering without being retrained for each. The headline risk is hallucination — because the model generates plausible text rather than retrieving verified facts, it can state something false with complete confidence. It also only knows what was in its training data up to a cutoff date.

In construction, LLMs are what let software read a 400-page specification and answer questions about it, draft a scope of work, or summarize differences between drawing revisions. Because of hallucination, serious tools ground the model in the actual project documents (a technique called RAG) and keep an expert reviewing the output rather than trusting raw generation.

Real Examples

→Spec Q&A: An LLM reads a specification and answers 'what are the concrete testing requirements?' with the relevant clauses pulled together in plain language.
→Scope drafting: Given a set of construction documents, an LLM drafts a first-pass scope of work for the concrete trade package for an estimator to refine.
→Without AI vs with AI: Without an LLM, comparing two spec revisions is a line-by-line manual read; with an LLM, the model surfaces what changed and the reviewer confirms the material ones.

Common Misconceptions

People assume: An LLM looks up facts in a database.

Actually: An LLM generates text by predicting likely words from patterns it learned in training — it isn't querying a fact store. That's why it can invent citations or numbers that look right but don't exist. Grounding it in real documents (RAG) is how tools make it reliable.

People assume: A bigger or newer LLM is always more accurate.

Actually: Scale improves fluency and breadth, but every LLM still hallucinates and has a knowledge cutoff. For a specific task like reading your project's specs, how the model is grounded and verified matters more than which model is the largest.

Does MeltPlan Solve This?

Not directly

A large language model is a general-purpose engine, not a construction product. MeltPlan applies LLMs to a narrow, high-value job — reading construction documents to draft bid-package scope, answer questions about specs, and support design review — grounded in your actual project files and checked by estimators, rather than offering an open-ended chatbot you have to fact-check yourself.

Frequently Asked Questions

How does a large language model work?

It's trained on huge amounts of text to predict the next word in a sequence. Repeated across billions of examples, this teaches it grammar, facts, and reasoning patterns as statistical relationships. When you prompt it, it generates a response one token at a time based on those patterns.

What is an LLM hallucination?

When a model produces text that is fluent and plausible but factually wrong — an invented citation, a made-up number, a nonexistent code section. It happens because the model generates likely-sounding language rather than retrieving verified facts. It's why LLM output must be checked.

What are LLMs good and bad at?

Good at summarizing, drafting, classifying, and answering questions over text. Bad at guaranteed factual accuracy, precise math, and anything past their training cutoff. Grounding them in real source documents and keeping a human reviewer addresses most of the weaknesses.

What is the difference between an LLM and generative AI?

Generative AI is the broad category of models that create new content — text, images, audio, or code. An LLM is the text-focused kind. All LLMs are generative AI, but generative AI also includes image and audio models that aren't LLMs.

Related Terms

Artificial Intelligence (AI)Software that performs tasks normally requiring human intelligence.Deep LearningMachine learning with many-layered neural networks that read raw images and text.RAG (Retrieval-Augmented Generation)Grounding an LLM in real documents by retrieving relevant passages before it answers.Context WindowThe amount of text an AI model can consider at once, measured in tokens.AI AgentAn AI that plans, uses tools, and takes multi-step actions toward a goal.Embedding (AI)Turning text or data into a vector of numbers that captures its meaning.

More AI Concepts & Fundamentals Terms

Sources

  1. Stanford HAI — AI Index Report
  2. IBM — What Are Large Language Models (LLMs)?
  3. Zhao et al. — A Survey of Large Language Models (arXiv:2303.18223)
MELTPLAN