What is LLM?
LLM (Large Language Model) is a neural language model trained to predict, score, or generate token sequences from learned parameters and supplied context. Its task behavior depends on the model revision, tokenizer, prompt, decoding policy, tools, retrieval, and evaluation protocol.
Quick Facts
| Full Name | Large Language Model |
|---|---|
| Created | 2018 (GPT-1), scaled significantly from 2020 (GPT-3) |
| Specification | Official Specification |
How It Works
Many LLMs use decoder-style Transformer architectures and are pretrained with a token prediction objective, then adapted with supervised training, preference optimization, tool use, retrieval, or product-specific controls. There are also encoder, encoder-decoder, multimodal, and small language model variants. Parameter count, benchmark score, or fluent output alone does not establish reliability for a task. An LLM can produce useful drafts, transformations, classifications, and tool plans, but it can also be wrong, omit constraints, follow untrusted instructions, expose sensitive context, or produce inconsistent outputs. Evaluate a dated model revision with representative inputs, clear success criteria, adversarial cases, cost and latency budgets, human review where needed, and a rollback path. Retrieval and lower-temperature decoding can change behavior but do not guarantee factuality or safety.
Key Characteristics
- Predicts or generates tokens conditional on learned parameters and supplied context
- Uses an architecture, tokenizer, context limit, and decoding policy with version-specific behavior
- Can adapt behavior through prompts, retrieval, tools, or parameter updates
- May generalize across tasks but requires representative evaluation for each deployment
- Has measurable quality, latency, context, memory, cost, privacy, and safety trade-offs
- Can generate plausible unsupported statements and requires verification controls for high-impact use
Common Use Cases
- Conversational AI assistants and chatbots for customer support
- Content generation including articles, marketing copy, and creative writing
- Code generation, completion, and debugging for software development
- Language translation and cross-lingual communication
- Text summarization and information extraction from documents
Example
Loading code...Frequently Asked Questions
What is the difference between LLM and traditional NLP models?
Many traditional NLP systems were trained or engineered for a narrow task, while LLMs are often pretrained broadly and adapted through prompts, retrieval, fine-tuning, or tools. The boundary is not absolute: smaller or task-specific models can be more accurate, cheaper, private, or easier to validate. Compare the exact models on representative inputs and failure cases.
How much data and compute is needed to train an LLM?
Requirements vary substantially by architecture, parameter count, token budget, sequence length, precision, hardware, parallelism, and training objective. Public reporting is incomplete for many models. Estimate a concrete run from the chosen configuration, data governance requirements, hardware availability, energy, evaluation, and operational costs rather than applying a universal token, GPU, or price figure.
What are hallucinations in LLMs and how can they be reduced?
Hallucinations are outputs that are unsupported, incorrect, or fabricated relative to the task evidence. Mitigations include scoped retrieval with source checks, constrained outputs, verification tools, calibrated abstention, human review, and task-specific evaluation. RAG, fine-tuning, or lower-temperature decoding can change error patterns but cannot guarantee factuality.
Can LLMs be run locally without cloud APIs?
Yes, some model revisions can run locally or on-device. Feasibility depends on model architecture and size, quantization, context length, KV cache, batch size, backend, hardware memory, throughput target, and license. Measure the actual revision under the expected workload and account for security updates, data retention, access control, and monitoring.
What is the context window and why does it matter?
The context window bounds the tokens a particular model configuration can process across input, generated output, and sometimes system or tool content. A larger nominal window does not guarantee effective use of every position. It affects memory, latency, cost, truncation, retrieval design, and evaluation; test relevant long-context cases instead of relying on a headline limit.