What Is a Large Language Model (LLM)?

A large language model (LLM) is a type of artificial intelligence built on a transformer-based neural network, trained on hundreds of billions to trillions of tokens of text to learn statistical patterns in language. OpenAI's GPT-3 was trained on 300 billion tokens in 2020; Meta pretrained Llama 3 on more than 15 trillion in 2024. The result is a model that can generate coherent text, answer questions, summarize documents, write code, and hold multi-turn conversations - all without being explicitly programmed for any individual task.

LLMs are the foundation of model families such as OpenAI's GPT, Anthropic's Claude, Google's Gemini and Meta's Llama. Within a few years they have moved from research curiosity to enterprise infrastructure, powering everything from customer service bots to internal knowledge assistants to automated data pipeline documentation. Understanding what they are - and what they are not - is now a prerequisite for anyone building or governing enterprise AI.

How LLMs Work

LLMs are built on the transformer architecture, introduced in 2017 by a team of mostly Google researchers in the paper Attention Is All You Need. At their core, they work by predicting what comes next. Given a sequence of tokens (words, subwords, or characters), the model learns to predict the most probable continuation. Trained across enough text with enough compute, this simple objective produces a model capable of remarkable generalization.

Tokens, Attention, and Context Windows

Text is broken into tokens. For English text, OpenAI's rule of thumb is about 0.75 words per token (100 tokens for roughly 75 words); the exact ratio depends on the tokenizer and the language. The model processes tokens through layers of self-attention, which lets each token relate to every other token in the current input. This attention mechanism is what allows LLMs to track long-range dependencies (the subject of a sentence 400 tokens ago, a reference made three paragraphs back) that earlier models like LSTMs struggled with.

The total number of tokens the model can process at once is the context window. GPT-3 (2020) had 2,048 tokens. As of 2026, several Anthropic and Google models offer windows of 1 million tokens, and Meta advertised 10 million for Llama 4 Scout. Bigger is not automatically better: Anthropic's documentation notes that accuracy and recall degrade as the token count grows. A larger context window means the model can hold more of a conversation, document, or codebase in working memory - which directly affects the quality of enterprise RAG applications.

Pre-training and Fine-tuning

LLM development has two main phases. Pre-training is expensive and done by foundation model labs (OpenAI, Anthropic, Google, Meta): train on internet-scale text, produce a general-purpose model. Fine-tuning is cheaper and often done by enterprises: take a pre-trained model and adapt it to a specific domain, task, or behavior using a smaller curated dataset. Fine-tuning can teach the model a company's terminology, tone, or specialized knowledge without retraining from scratch.

Training Data and Scale

The "large" in large language model refers to two dimensions: the number of parameters (weights) in the neural network, and the volume of training data. GPT-3 had 175 billion parameters. Most frontier labs no longer publish parameter counts; in April 2025 Meta described Llama 4 Behemoth, then still in training, as having nearly two trillion total parameters. Researchers have described emergent abilities - capabilities absent in smaller models but present in larger ones (Wei et al., 2022) - although later work argued that some of these jumps are artifacts of how performance is measured (Schaeffer et al., 2023).

Training data quality matters at least as much as quantity. Models trained on poorly curated data learn to replicate its flaws - biases, factual errors, inconsistencies. The "garbage in, garbage out" principle applies with multiplicative force when the model has memorized its training distribution and will confidently reproduce it.

The Data Provenance Problem

For enterprise AI teams, training data provenance is increasingly a legal and governance concern. Questions that now land on data teams: What data did this model train on? Are we allowed to use it for commercial output? Do we understand what biases it may have absorbed? These are data governance questions - the same discipline that governs enterprise data assets now needs to extend to the AI systems that consume and generate data. Regulation points the same way: the EU AI Act requires providers of general-purpose AI models to publish a sufficiently detailed summary of the content used for training (Article 53(1)(d)), an obligation that applies from 2 August 2025.

Capabilities and Limits

LLMs are genuinely good at a cluster of tasks: synthesizing information across many sources, generating coherent long-form text, translating, summarizing, answering questions with context, writing and explaining code, and reformatting structured data. These capabilities transfer across domains - a model trained on general text can discuss medical literature, legal contracts, and financial disclosures without separate training on each.

Hard Limits That Won't Be Fixed by Scale

Three limits are structural - they don't disappear just by making models bigger:

• Knowledge cutoff. LLMs are trained on a static snapshot of text. They don't know what happened yesterday, don't update continuously, and will confidently answer questions about events after their cutoff as if they had seen the data. This is why enterprise deployments commonly pair LLMs with retrieval systems.
• Hallucination. LLMs generate plausible-sounding text by predicting likely tokens. When asked about something outside their training distribution, they often don't say "I don't know" - they generate a confident-sounding answer that may be entirely fabricated. See the separate entry on AI hallucination for the full treatment.
• No private knowledge by default. A foundation LLM knows nothing about your company's data, your internal systems, your customers, or your processes - unless you give that context in the prompt or through a retrieval mechanism. This is why RAG and fine-tuning exist, and why the quality of the context you provide is often the limiting factor in enterprise AI quality.

LLMs in the Enterprise

Enterprise LLM deployment typically takes one of three architectures:

1. API access to foundation models - connect to OpenAI, Anthropic, Google via API. Fast to deploy, no training overhead, but all prompts go to third-party infrastructure. Data residency and privacy concerns apply.
2. Self-hosted open models - run open-weight models such as Meta's Llama or Mistral's models on your own infrastructure. Data stays in-house, but requires ML engineering capacity for deployment, inference optimization, and updates.
3. Fine-tuned domain models - adapt a base model to enterprise-specific vocabulary, tone, and domain knowledge. Often the strongest results on narrow, specialized tasks, with the highest upfront cost.

Most enterprise use cases on all three architectures share one further need: retrieval-augmented generation to ground the model in current, private, enterprise-specific knowledge. The LLM without an information retrieval layer is a general-purpose text generator. With a well-governed retrieval layer, it becomes an enterprise intelligence system.

Data Governance for LLMs

LLMs introduce new requirements for data governance that most existing frameworks don't address. Three areas are most critical:

Input Governance: What Context You Feed the Model

Every prompt carries data. If employees use an LLM to answer questions about customers, patient records, or financial projections, the sensitive data in those prompts may leave the corporate network. Input governance - classifying what data can go to which models, under what conditions - is a new operational requirement for data teams.

Output Governance: Verifying Generated Content

LLM outputs can't be trusted as-is. They require verification: Does this answer match our metadata? Is this summary accurate? Did the model cite real sources? Organizations that have invested in a data catalog and knowledge graph can point the LLM at governed, authoritative sources - making verification tractable. Without that infrastructure, each output is a manual fact-check.

Model Governance: Version, Drift, and Accountability

Like software, models have versions. Like data, they can drift. A model that performed well in March may perform differently in June - the same input producing different outputs after a provider update. Researchers documented exactly this when they compared the March and June 2023 versions of GPT-4 and GPT-3.5 and found substantial changes in behavior on the same tasks (Chen, Zaharia and Zou, 2023). Model governance tracks which version is deployed, monitors for performance drift, and maintains an audit trail of what the model was used for. This is the domain of LLMOps.

How Dawiso Supports LLM Workflows

Dawiso does not train or host language models. Its role in the enterprise LLM stack is to be the governed context layer that LLM-based assistants and agents can draw on. When a user asks an AI assistant "what does churn rate mean in our context?" or "which datasets feed the monthly revenue report?", the answer should come from documentation the organization maintains and governs - not from an LLM's general training data.

The Business Glossary holds the agreed definitions of business terms, and Dawiso links glossary terms automatically where they are mentioned in object content. The Data Catalog provides the inventory of data assets, with data owners and data stewards assigned to them. The Dawiso MCP Server exposes this documentation to any MCP-compatible agent via the Model Context Protocol: agents can search objects, read object details, trace lineage and preview data samples, and write tools are offered only to user roles that permit them. Dawiso MCP is currently available to cloud customers.

The result: LLM answers about business terms and data assets that can be grounded in, and checked against, the catalog objects behind them.

Conclusion

Large language models are a genuinely transformative technology - but their enterprise value depends heavily on the infrastructure around them. An LLM with no reliable context is a confident text generator prone to hallucination. An LLM connected to a governed knowledge graph, a maintained business glossary, and well-cataloged data assets is the foundation of enterprise intelligence. The model provides the language capability; data governance provides the trustworthiness.

Sources

• Vaswani et al., Attention Is All You Need, arXiv, June 2017.
• Brown et al., Language Models are Few-Shot Learners, arXiv, May 2020 (GPT-3: 175 billion parameters, 300 billion training tokens, 2,048-token context window).
• Meta, Introducing Meta Llama 3: The most capable openly available LLM to date, April 18, 2024 (pretrained on over 15 trillion tokens).
• Meta, The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation, April 5, 2025 (Llama 4 Scout 10M context window; Llama 4 Behemoth nearly two trillion total parameters).
• OpenAI Help Center, What are tokens and how to count them? (English rules of thumb).
• Anthropic, Context windows, Claude documentation (1M-token windows; context rot).
• Google, Long context, Gemini API documentation.
• Wei et al., Emergent Abilities of Large Language Models, arXiv, June 2022.
• Schaeffer, Miranda and Koyejo, Are Emergent Abilities of Large Language Models a Mirage?, arXiv, April 2023.
• Chen, Zaharia and Zou, How is ChatGPT's behavior changing over time?, arXiv, July 2023.
• Regulation (EU) 2024/1689 (AI Act), Article 53(1)(d) and Article 113, EUR-Lex.
• Dawiso Help, Dawiso Model Context Protocol (MCP).
• Dawiso Help, Business glossary.
• Dawiso Help, Role of ownership in data governance.